keyphrase_and_bert.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#!pip install sentence-transformers==0.3.3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sentence_transformers import SentenceTransformer\n",
    "import numpy as np\n",
    "import torch\n",
    "from torch import nn"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "with open(\"simplified_last_1000_keyphrases.json\", \"r\") as f:\n",
    "    lines = f.readlines()\n",
    "pred_data = [json.loads(l) for l in lines]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "#use_docs = 1000\n",
    "#pred_data = pred_data[-use_docs:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "all_keyphrases = {}\n",
    "total_keyphrases = 0\n",
    "for i in pred_data:\n",
    "    for j in i[\"pred\"]:\n",
    "        total_keyphrases+=1\n",
    "        if j not in all_keyphrases:\n",
    "            all_keyphrases[j] = 1\n",
    "        else:\n",
    "            all_keyphrases[j]= all_keyphrases[j]+1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "14329"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "total_keyphrases"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "9897"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(all_keyphrases.keys())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[('', 186),\n",
       " ('neural networks', 162),\n",
       " ('machine learning', 80),\n",
       " ('<unk>', 70),\n",
       " ('reinforcement learning', 67),\n",
       " ('deep learning', 64),\n",
       " ('learning', 59),\n",
       " ('neural network', 40),\n",
       " ('classification', 39),\n",
       " ('recurrent neural networks', 39)]"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sorted_lst = list(sorted(all_keyphrases.items(), key=lambda item: item[1], reverse=True))\n",
    "sorted_lst[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def top_keyphrases_of_topic(topic_ids):\n",
    "    keyphrases = {}\n",
    "    total_keyphrases = 0\n",
    "    for idx, i in enumerate(pred_data):\n",
    "        if idx in topic_ids:\n",
    "            for j in i[\"pred\"]:\n",
    "                total_keyphrases+=1\n",
    "                if j not in keyphrases:\n",
    "                    keyphrases[j] = 1\n",
    "                else:\n",
    "                    keyphrases[j]= keyphrases[j]+1\n",
    "    print(\"total keyphrases are: \", total_keyphrases)\n",
    "    print(\"Unique nummber of keyphrases are:\", len(keyphrases.keys()))\n",
    "    return list(sorted(keyphrases.items(), key=lambda item: item[1], reverse=True))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "total keyphrases are:  1383\n",
      "Unique nummber of keyphrases are: 1133\n"
     ]
    }
   ],
   "source": [
    "#Topic 1\n",
    "topic = [547, 226, 571, 792, 964, 916, 166, 193, 165, 662, 143, 98, 790, 649, 555, 590, 172, 561, 732, 200, 311, 67, 391, 86, 144, 553, 344, 951, 243, 960, 517, 467, 638, 475, 372, 179, 62, 413, 47, 198, 152, 442, 771, 786, 624, 256, 25, 39, 768, 986, 958, 535, 306, 515, 276, 831, 895, 825, 453, 390, 543, 45, 815, 558, 377, 666, 400, 548, 923, 50, 303, 929, 979, 6, 866, 23, 63, 937, 576, 419, 796, 877, 180, 915, 336, 239, 162, 909, 119, 738, 892, 8, 697, 944, 14, 500, 38, 801, 975, 873]\n",
    "topic = [t - 1 for t in topic]\n",
    "topic_keyphrases = top_keyphrases_of_topic(topic)\n",
    "#print(topic_keyphrases[:50])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_path = 'phrase-bert-model'\n",
    "model = SentenceTransformer(model_path)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": true,
    "jupyter": {
     "outputs_hidden": true
    },
    "tags": []
   },
   "outputs": [
    {
     "ename": "ValueError",
     "evalue": "not enough values to unpack (expected 2, got 1)",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mValueError\u001b[0m                                Traceback (most recent call last)",
      "\u001b[0;32m/var/folders/pp/xk4wtc5x62z94qrdxdbv__1h0000gn/T/ipykernel_43277/1191558193.py\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[0;32m----> 1\u001b[0;31m \u001b[0mkeyphrase_embs\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mmodel\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mencode\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mi\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mi\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mtopic_keyphrases\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[0;32m~/opt/anaconda3/lib/python3.9/site-packages/sentence_transformers/SentenceTransformer.py\u001b[0m in \u001b[0;36mencode\u001b[0;34m(self, sentences, batch_size, show_progress_bar, output_value, convert_to_numpy, convert_to_tensor, is_pretokenized)\u001b[0m\n\u001b[1;32m    185\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    186\u001b[0m             \u001b[0;32mwith\u001b[0m \u001b[0mtorch\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mno_grad\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 187\u001b[0;31m                 \u001b[0mout_features\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mforward\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mfeatures\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    188\u001b[0m                 \u001b[0membeddings\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mout_features\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0moutput_value\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    189\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/opt/anaconda3/lib/python3.9/site-packages/torch/nn/modules/container.py\u001b[0m in \u001b[0;36mforward\u001b[0;34m(self, input)\u001b[0m\n\u001b[1;32m    139\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mforward\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minput\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    140\u001b[0m         \u001b[0;32mfor\u001b[0m \u001b[0mmodule\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 141\u001b[0;31m             \u001b[0minput\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mmodule\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0minput\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    142\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0minput\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    143\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/opt/anaconda3/lib/python3.9/site-packages/torch/nn/modules/module.py\u001b[0m in \u001b[0;36m_call_impl\u001b[0;34m(self, *input, **kwargs)\u001b[0m\n\u001b[1;32m   1108\u001b[0m         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks\n\u001b[1;32m   1109\u001b[0m                 or _global_forward_hooks or _global_forward_pre_hooks):\n\u001b[0;32m-> 1110\u001b[0;31m             \u001b[0;32mreturn\u001b[0m \u001b[0mforward_call\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0minput\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1111\u001b[0m         \u001b[0;31m# Do not call functions when jit is used\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1112\u001b[0m         \u001b[0mfull_backward_hooks\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mnon_full_backward_hooks\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/opt/anaconda3/lib/python3.9/site-packages/sentence_transformers/models/BERT.py\u001b[0m in \u001b[0;36mforward\u001b[0;34m(self, features)\u001b[0m\n\u001b[1;32m     31\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mforward\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mfeatures\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     32\u001b[0m         \u001b[0;34m\"\"\"Returns token_embeddings, cls_token\"\"\"\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 33\u001b[0;31m         \u001b[0moutput_states\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbert\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m**\u001b[0m\u001b[0mfeatures\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     34\u001b[0m         \u001b[0moutput_tokens\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0moutput_states\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     35\u001b[0m         \u001b[0mcls_tokens\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0moutput_tokens\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;36m0\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m:\u001b[0m\u001b[0;34m]\u001b[0m  \u001b[0;31m# CLS token is first token\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/opt/anaconda3/lib/python3.9/site-packages/torch/nn/modules/module.py\u001b[0m in \u001b[0;36m_call_impl\u001b[0;34m(self, *input, **kwargs)\u001b[0m\n\u001b[1;32m   1108\u001b[0m         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks\n\u001b[1;32m   1109\u001b[0m                 or _global_forward_hooks or _global_forward_pre_hooks):\n\u001b[0;32m-> 1110\u001b[0;31m             \u001b[0;32mreturn\u001b[0m \u001b[0mforward_call\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0minput\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1111\u001b[0m         \u001b[0;31m# Do not call functions when jit is used\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1112\u001b[0m         \u001b[0mfull_backward_hooks\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mnon_full_backward_hooks\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/opt/anaconda3/lib/python3.9/site-packages/transformers/models/bert/modeling_bert.py\u001b[0m in \u001b[0;36mforward\u001b[0;34m(self, input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds, encoder_hidden_states, encoder_attention_mask, past_key_values, use_cache, output_attentions, output_hidden_states, return_dict)\u001b[0m\n\u001b[1;32m    948\u001b[0m             \u001b[0;32mraise\u001b[0m \u001b[0mValueError\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"You have to specify either input_ids or inputs_embeds\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    949\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 950\u001b[0;31m         \u001b[0mbatch_size\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mseq_length\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0minput_shape\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    951\u001b[0m         \u001b[0mdevice\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0minput_ids\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mdevice\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0minput_ids\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0;32mnot\u001b[0m \u001b[0;32mNone\u001b[0m \u001b[0;32melse\u001b[0m \u001b[0minputs_embeds\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mdevice\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    952\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mValueError\u001b[0m: not enough values to unpack (expected 2, got 1)"
     ]
    }
   ],
   "source": [
    "keyphrase_embs = model.encode([i[0] for i in topic_keyphrases])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "similarity_matrix = np.empty((len(keyphrase_embs), len(keyphrase_embs)), dtype=float)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "cos_sim = nn.CosineSimilarity(dim=0)\n",
    "for row_idx, i in enumerate(keyphrase_embs):\n",
    "    for col_idx, j in enumerate(keyphrase_embs):\n",
    "        similarity_matrix[row_idx][col_idx] =  round(cos_sim(torch.tensor(i), torch.tensor(j)).item(), 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[1.  , 0.56, 0.61, ..., 0.54, 0.44, 0.54],\n",
       "       [0.56, 1.  , 0.82, ..., 0.47, 0.33, 0.38],\n",
       "       [0.61, 0.82, 1.  , ..., 0.56, 0.4 , 0.44],\n",
       "       ...,\n",
       "       [0.54, 0.47, 0.56, ..., 1.  , 0.47, 0.45],\n",
       "       [0.44, 0.33, 0.4 , ..., 0.47, 1.  , 0.9 ],\n",
       "       [0.54, 0.38, 0.44, ..., 0.45, 0.9 , 1.  ]])"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "similarity_matrix"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "a = similarity_matrix > 0.7"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x1966657208>"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print([i[0] for i in topic_keyphrases])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "keyphrase_embedding = {}\n",
    "for keyphrase, freq in all_keyphrases.items():\n",
    "    keyphrase_embedding[keyphrase] = model.encode([keyphrase]).tolist()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "9897"
      ]
     },
     "execution_count": 42,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(keyphrase_embedding.keys())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "with open('1000_docs_keyphrase_embedding.json', 'w') as fp:\n",
    "    json.dump(keyphrase_embedding, fp)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# From json file"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "with open('1000_docs_keyphrase_embedding.json') as json_file:\n",
    "    kp = json.load(json_file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "collapsed": true,
    "jupyter": {
     "outputs_hidden": true
    },
    "tags": []
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "dict_keys(['style transfer', 'feature transforms', 'image processing', 'image power analysis', 'feature time', 'we analysis', 'neural networks', 'feature transform', 'we image processing', 'we image synthesis', 'networks', 'feature cross analysis', 'vqa', 'methods', '', 'we', 'method', 'gamma process', 'edge partition model', 'binary matrix', 'model for have prediction', 'learning from art', 'markov chain monte carlo', 'model samples', 'model shrinkage', 'are in at advocate', 'sketches', 'art', 'art lasso', 'model linear model', 'to this sketches', 'to this render', 'pose guided person generation network', 'pose guided person generation', 'pose classification', 'u game', 'person image synthesis', 'pose guided person that problem', 'u . which method', 'u . which evaluated', 'image . based modeling and synthesis', 'present value', 'in , g . transforms', 'u . which constraint', 'in , b . transforms', 'in , g , transforms', 'graphical models', 'semidefinite programming', 'sum - of - squares', 'graph inference', 'linear programming', 'combinatorial optimization', 'maximum a posteriori probability', 'populations', 'semidefinite are graph', 'sum - of - squares hierarchy', 'model selection', 'model complexity', 'from enabling factor', 'model in graphs', 'model in motion', 'astronomy', 'region', 'exploit', 'decision trees', 'random forests', 'impurity to converge', 'impurity', 'impurity model', 'impurity to examples', 'impurity to both', 'decision tree', 'random forest', 'impurity to data', 'impurity to posterior', 'gated recurrent unit', 'gradient explosions', 'improves', 'has', 'contrastive learning', 'recurrent units', 'gated contrastive learning', 'learning', 'recurrent neural network', 'in large time series', 'in large time large scale', 'in large time large data', 'in large time large scale systems', 'dynamic time series', 'gated recurrent units', 'in large large large data sets', 'in large time large systems', 'polyhedra', 'gated', 'high - rank matrix estimation', 'singular value decomposition', 'truncated svd', 'singular value convergence', 'singular optimum', 'singular value change', 'active subspace method', 'used of large numbers', 'singular value decision', 'work function', 'singular new of form', '1 . 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '1 . 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2', '1 . 2 , 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '1 . 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '1 . 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2', '1 . 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2', '1 . 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 1 , 2 , 2', '1 . 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '1 . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '1 . 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2', 'gan', 'f - divergence', 'information geometry', 'f - gans', 'activation functions', 'composite losses', 'information geometric', 'exponential families', 'composite link function', 'geometric distribution', 'cascade networks', 'generalization', 'time series prediction', 'cascade architecture', 'complex networks', 'thus', 'cascade network architecture', 'time seeks', 'generalization regression', 'include delay', 'essential facilities', 'include dissipative networks', 'include higher order', 'include dissipative systems', 'include an diffusion', 'image - to - image translation', 'conditional generative modeling', 'opponent modeling', 'self in and sampling', 'completion', 'a posteriori estimation', 'problem decomposition', 'performance prediction', 'for ligand recognition', 'to and intra prediction', 'low - rank matrix approximation', 'collaborative low . to nonlinear principle component analysis', 'new estimating equation', 'completion matrix', 'new estimating equations', 'new hilbert space', \"new hilbert ' s decomposition\", 'new we distribution model', 'different epm', 'collaborative filtering', 'low - rank matrix ) have control', 'in the and the completion probability matrix', 'mixture', \"new hilbert ' s decomposition approach\", 'networks of dimensionality', 'non - monotone continuous functions', 'non - monotone dr - selection', '5d optimization', 'of we', 'sgd', 'asymmetric of hyperbolic functions', '5d sketches', 'standard asymmetric d , actually', 'standard asymmetric of d , paper the information', 'standard asymmetric of d the paper the latent variables', 'non - monotone dr - submodularity', 'non - monotone continuous for non realistic shape optimization', ') ) ) continuous ) on determinantal improvements', 'on of activity', 'standard asymmetric of d the paper the latent of neural networks', 'on of of activity', 'motion planning', 'neural network', 'average top - k loss', 'aggregate loss', 'learning theory', 'support vector machines', 'supervised learning', 'support vector machine', 'reply time', 'challenge', 'task', 'deep networks', 'automatically defined functions', 'attributes', '<s> visualization', 'polynomially certain answers', '<s> domain', 'domain', 'imaging', 'first for space', 'first model', 'important', 'it', 'given', '<s> with data', 'first model model', 'convex optimization', 'large <unk> scale optimization', 'total variation', 'condition number', 'a posteriori error bounds', 'convolutional neural networks', 'object detection', '360?', 'spherical convolution', 'feature extraction', 'process imagery', 'process 360? imagery', '<unk>', 'process objectives', 'spherical convolution convolution', 'mean dr', 'itself', '( and learning', 'process and approximate reasoning', '3d shape reconstruction', '2 . 5d sketches', 'domain adaptation', 'shape reconstruction', '3d object reconstruction', '2 . 5d modeling', 'projective transformations', '2 , sgd', 'end - to - for the , a priori constraints', 'in , a priori constraints', 'end - to - for the , to design', 'design', 'i , semidefinite programming', 'i , semidefinite relaxation', 'visual question answering', 'modular neural networks', 'in is learning', 'learning to rank', 'in and medium learning', 'in and medium processing', 'ordinal regression', 'adversarial losses', 'surrogate losses', 'worst problem', 'classification', 'surrogate geometric show', 'ordinal on distribution', 'hinge 1 demonstrate', 'hinge 1 minimization', 'surrogate geometric in uci', 'a in labels', 'a in dataset', 'transient signal detection', 'artificial neural network', 'back propagation algorithm', 'artificial neural networks', 'back propagation learning algorithm', 'back <unk> propagation algorithm', 'back propagation', 'address detection', 'scheduling', 'hypothesis transfer', 'transformation function', 'kernel', 'transformation functions', 'kernel method', 'protocol', 'multi - objective optimization', 'adversarial feature learning', 'image classification', 'feature learning', 'adversarial learning', 'game theory', 'representation learning', 'learning meaningful representations', 'exploding captioning', 'invariant representation', 'two - layer neural networks', 'stochastic gradient descent', 'relu activation', 'convergence analysis', 'relu', 'convergence be', 'principled design', 'architectures', 'architectures of algorithms', 'we priori information', 'dimensionality reduction', 'we for we interference', 'we for we', 'we for we are we', 'two - layer perceptron', 'relu class', 'we for we we', 'we for we we interference', 'we for that are we', 'empirical risk minimization', 'stochastic gradient method', 'optimize', 'nonlinear programming', 'stochastic gradient', 'model reduction', 'dual averaging', 'langevin dynamics', 'continuous tempering', 'deep neural networks', 'fundamental matrix sampling', 'empirical bayes', 'a posteriori probability estimation', 'gibbs sampling', 'on and sales dynamics', 'on and sales nonconvex optimization', 'on and off completion', 'online learning', 'linear optimization', 'approximation algorithms', '? - regret bound', '? - regret bounds', '? - regret', 'geometric descent method', 'proximal gradient method', 'a 1 we ( explicitly small htl very large', 'by 1 we ( explicitly small htl very large', 'a 1 we , explicitly small htl very large', 'a 1 we ( explicitly small htl very high fairness', 'by 1 we ( explicitly small htl very high fairness', 'aim analysis', 'robust convergence', 'aim', 'geometric', 'this 0 hypothesis', 'robust estimation', 'this t hypothesis', 'and convergence rate', 'principal component analysis', 'global convergence', 'algorithm', 'diffusion approximation', 'global image', 'framework', 'global image images', 'correspondences', 'interpretability', 'oja?s iteration', 'enforcing penalty', 'kernels', 'causal reasoning', 'fairness', 'protected attributes', 'discrimination', 'fairness by discrimination', 'fairness conclusively', 'causal visual visual visual visual method', 'causal visual visual visual reasoning', 'causal visual visual visual visual reasoning', 'machine', 'question variants', 'top is', 'top is further data', 'top is further completion', 'top is by a kernels', 'top is by programs', 'between first , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 2', 'between first , 1 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 2', 'between first , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2', 'between first , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 2', 'between first , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 4', 'between first , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1', 'between first , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 2 , 2', 'between first , 1 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 4', 'between first , 1 , 1 , 1 , 1 , 2 , 2 , 1 , 2 , 2 , 2', 'between first , 1 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 1', 'nonparametric regression', 'metric learning', 'machine learning', 'regression', 'infomax', 'information theory', 'multiresolution', 'optimal encoding', 'receptive field', 'information - theoretic optimization', 'wavelets', 'negative correlation', 'partial differential equations', 'partial differential equation', 'privileged learning', 'distribution matching', 'recycling', 'multi - faceted', 'imperfect - information games', 'subgame solving', 'action translation', 'imperfect information games', 'generative models', 'analysis', 'probabilistic graphical models', 'bayesian inference', 'bayesian networks', 'analysis of algorithms', 'coded - computation', 'personalized pagerank', 'social networks', 'inverse problems', 'coded - code', 'scalability', 'error - correcting code', 'error - correcting - code', 'personalized complexity', 'social', 'error - correcting - even codes', 'convex hull', 'a - compression', 'a posteriori error bound', 'a posteriori error correction', 'dimensional reduction', 'number of clusters', 'a posteriori error estimation', 'prediction', 'planes', ') to image', 'possible world', 'possible worlds', 'be to frontier', 'matching', 'training', 'o', 'that conservation', 'literature overview', 'deep <unk>', 'deep matrix', 'deep packet', 'ising model', 'screening rule', 'screening is', 'l1 regularization', 'sparsity', 'screening', 'l1 - regularization', 'clustering', 'screening regularization', 'data fusion', 'data clustering', 'screening a', ', <unk>', 'in the roc curve', ', quality control', 'in the images', 'in the adapter', 'in the direct method', 'in the direct', 'in the direct marketing', 'in the . which', 'dynamic regret', 'non - degenerate functions', 'knowledge distillation', 'computer vision', 'domain mapping', 'unsupervised learning', 'circularity', 'numerical methods', 'inverse kinematics', 'gab', 'that we', 'steered simulation', 'steered response surface', 'steered response speed', 'gab model', 'steered response', 'image restoration', 'bayes estimation', 'representation theory', 'bayes model', 'mean - ) prior', 'net regularization', 'between the em algorithm', 'tikhonov regularization', 'into <unk> estimator', 'evaluate', 'between the em', 'mixture model', 'over and gaussian noise', 'in a posteriori estimation', 'kullback a <unk> leibler divergence', 'truncated kullback a <unk> leibler divergence', 'gamma distribution', 'atoms', 'variational methods', 'efficiency', 'active noise', 'markov random', 'kullback the', 'active learning', 'shrinkage estimation', 'gamma the', 'greedy algorithms', 'matching pursuit', 'convergence', 'robust optimization', 'convergence rate', 'non - negative optimization', 'non - machine learning', 'function approximation', 'non - machine optimization', 'to an detection', 'non - negative objectives', 'to an detection limit', 'convolutional tomography', 'convolutional optimization', 'detection rate', 'n in filters', 'n in core optimization', 'exactly solvable property', 'n in convex imagery', 'hamming networks', 'threshold hamming networks', 'the', 'memory neurons', 'pattern recognition', 'hamming network', 'threshold hamming network', 'threshold hamming based classifiers', 'prevalent matrix completion', 'missing data', 'matrix completion', 'nonuniform sampling', 'missing data recovery', 'matrix distribution', 'complexity', 'high <unk> dimensional linear algebra', 'high <unk> dimensional reduction', 'high <unk> dimensional data', 'we <unk> ( formula', 'high <unk> dimensional linear program', 'robust hypothesis test', 'gaussian processes', 'maximum likelihood estimation', 'or in banking', 'nonlinear effect', 'hypothesis test', 'model multiple hypothesis testing', 'or in statistical learning theory', 'or in natural resources', 'data - generating function', 'gaussian process', 'adversarial perturbations', 'robustness', 'image recognition', 'adversarial models', 'submodular function', 'approximation', 'submodular functions', 'submodular function minimization', 'submodular function optimization', 'computational learning theory', 'introspective convolutional networks', 'pseudo - negative neural networks', 'recurrent neural networks', 'networks on graphs', 'introspective convolutional network', 'convolutional network', 'scale the roc curve', 'existing latent semantic networks', '( 2 the roc curve', 'measures of information', 'g the in syndrome', 'g the in the gamma', 'found in the gamma 1', 'found in , 1 ,', 'g the in the gamma 1', 'found in the )', 'found in the ) in', 'found in , 1 the', 'g the in the limit', 'g the in the gamma function', 'found in the ) 1', 'label distribution learning', 'label distribution', 'label', 'applications', 'functional', 'express', 'object', 'on - line learning', 'end 5d model', 'hierarchy', 'standardization', 'segmentation', 'arrays', 'hierarchy functional', 'and analysis', 'hierarchy )', 'visual perception', 'viewpoint generalization', 'viewpoint factorization', 'articulated objects', 'b we', 'equivariant rates', 'numerical analysis', 'numerical integration', 'application', 'non', 'during', 'rank', 'paper', 'compression', 'pruning', 'network', 'adaptation', 'variations', 'excellent model', 'change detection', 'focuses', 'model', 'chose', 'process', 'brain - computer interface', 'kalman filter', 'semi - activations', 'semi - markov dynamics', 'rehabilitation', 'biomedical signal processing', 'brain - learning', 'predictive learning', 'our , settings', 'spatiotemporal', 'our , <unk>', 'hidden markov models', 'baum - welch algorithm', ',', 'from statistics', 'statistical learning', 'statistical learning theory', ', number screening', 'that the done multipliers', 'markov random fields', 'learn from demonstration', 'learn to memory', 'stochastic processes', 'general', 'statistical inference', 'brain connectivity', 'detrended partial cross - that concretely', 'memory cf', 'dynamics', 'appearances', 'cf', 'cf component analysis', 'as relationship', 'cone', 'detrended fluctuation analysis', 'partial cross - that concretely', 'matrix algorithm', 'brain', 'algorithms', 'applicable factor', 'atom pairs', '. <unk>', 'atom ) we', 'atom ) we diffusion mri', 'atom ) we field', '1', 'fmri', 'of gene expression', '1 <unk> <unk>', '1 with 2', 'image captioning', 'natural descriptions', 'image and text indexing', 'information retrieval', 'and information retrieval', 'code search', 'code searching', 'code search strategies', 'image', 'code generation', 'distillation', 'boosting', 'o <unk> <unk>', 'in <unk> margin learning', 'in <unk> situ hybridization', 'in <unk> the <unk> shelf', 'distillation error', 'in <unk> the <unk> go', 'in <unk> the <unk> go learning', 'continuous learning', 'o <unk>', '<unk> <unk>', 'continuous queries', 'reinforcement learning', 'lyapunov stability', 'stability guarantees', 'safety - critical systems', 'control theory', 'model - based control', 'implementation', 'markov decision processes', 'markov decision process', 'multiclass', 'multiclass classification', 'multiclass boosting', 'we <unk> divergence', 'nearest neighbor matching', 'missing counterfactuals', 'o functions', 'o average consensus', 'treatment effects estimation', 'maximum likelihood', 'a posteriori probability', 'conditional random fields', 'o average case analysis', 'counterfactual prediction', 'obtain a function', 'non - parametric representation', 'kernel methods', 'non - parametric', 'np to mixtures', 'overcomplete hmms', 'property testing', 'state analysis', 'overcomplete', 'randomization', 'long - term memory', 'property learning', 'number theory', 'in ) learning', 'number theoretic learning', 'state transition graphs', 'have reinforcement learning', 'state reinforcement learning', 'state transition', 'effective connectivity', 'gaussian process regression', 'scene analysis', 'matter waves', 'causal', 'pair a posteriori', 'models', 'pair potential', 'netflix prize', 'action potential', 'netflix', 'pair decomposition', 'outperform', 'number of attention', 'number of demonstrate', 'however', 'in silico experiments', 'hypotheses generation', 'hypotheses testing', 'we training', 'for', 'ensuring', 'deep learning', 'gender classification', 'decoupling to three a wild dimensional', 'decoupling to three a wild the in the state', 'decoupling to three a wild the in . state', 'decoupling the three a wild the in . state', 'decoupling the three a wild the in . of the only cases', 'decoupling to three a wild the in . of the only cases', 'decoupling \"when to update\" from \"how to update\" three three method', 'convincing datasets', 'face', 'defeat', 'convex sets', 'decoupling \"when to update\" from update\" to update\" three three method', 'convex hulls', 'convex functions', 'self - normalizing neural networks', 'batch normalization', 'exploding gradients', 'drug discovery', 'activation function', 'self - normalizing properties', 'adversarial networks', 'robust inference', 'adversarial network', 'pivotal suggest', 'pivot ? a ? a ) support', 'needs analysis', 'needs measure', 'needs assessment', 'accelerated efficiently computable models', 'games on graphs', 'non over log in graphs', 'non over log in strong discontinuities', 'non over log <unk> structured data', 'limitations', 'perturbed optimization', 'backpropagation', 'recurrent network', 'higher order network', 'recurrent networks', 'higher , is with', 'is target', 'in <unk> network', 'higher order networks', 'initial value problem', 'in <unk> network flow', 'langevin equation', 'approach', 'in <unk> network training', 'lateral geniculate nucles', 'pattern analysis', 'by , associative memory', 'only by that accuracy', 'introduce is', 'only by that in unknown channel', 'introduce efficient coding', 'only by ( aim', 'approach vision', 'quantum interactions', 'data mining', 'example mining', 'self - weight training', 'stochastic content no learn', 'example - based learning', 'threshold', 'accuracy', 'experimentation', 'examples', 'architecture', 'neighbor accuracy', ') correspondence', 'unknown monte carlo', 'neurons', 'uncertainty', ', only algorithm', 'we we correspondence', 'unknown sampling', 'discrete optimization', 'differentiable learning', 'submodular optimization', 'optimization', 'variational inference', 'submodular minimization', 'inductive representation learning', 'transduction', 'large graphs', 'protein - protein interactions', 'inductive learning', 'protein - protein interaction', 'subset selection', 'summarization', 'sequential data', 'integer continuous optimization', 'facility inform discovery', 'bursts', 'down algorithm', 'integer binary binary', 'from algorithm', 'message passing', 'sequential a integer binary binary binary', 'help', 'program', 'agent', 'generation', 'cognition', 'formalism', 'precise feature generation', 'second and person reasoning', 'maps', 'nonlinear modeling', 'nonlinear regression', 'interatomic interactions', 'inference', 'moving objects', 'parsimony', 'colorization', 'multi assembly model', 'many assembly', 'multi assembly', 'multi assembly models', 'halfspaces', 'label complexity', 'learning of halfspaces', 'homogeneous halfspaces', 'pac learning', 'sample complexity', 'gradient descent', 'escape saddle points', 'synchronization', 'cowan equations', 'non - pathological functions', 'measurements', 'parametric optimization', 'cowan equation', '( outputs', 'almost always escapes', 'are', 'over online optimization', 'a design', 'design of experiments', 'data - driven discovery', 'union of intersections', 'mechanistic models', 'gene regulatory networks', 'microarray data', 'of learning', 'and regulatory networks', 'gene expression', 'instructional networks', 'gene expression data', 'genotype - match of observations', 'that of neuronal cells', 'microarray', 'imitation learning', 'neural nets', 'one - shot learning', 'one - shot', 'meta - learning', 'neuroevolution', 'hippocampus', 'firing rate', 'decoding', 'neural code', 'neural coding', 'convolutional sparse coding', 'expectation - maximization', 'cross - frequency directly proposed setting', 'results instance of results', 'alpha - stable distributions', 'expectation - faster enjoys enjoys', 'expectation of results', 'results', 'cross - frequency coupling', 'time - optimization', 'theoretical difference propagation', 'spike synchrony', '<unk> of results', 'results clustering', 'expectation', 'schedules', 'construction of algorithm', 'few a sampled data', 'few a sampled maybe instantiations', 'expectation a each a tower of algorithm', 'expectation a <unk> one one one one one a d', 'expectation a <unk> one one one one one a i d', '<unk> a <unk> one', '<unk> a <unk> a', 'expectation a <unk> one one one one one one a i d', '<unk> a <unk> <unk> one', '<unk> a <unk> one one', '<unk> a <unk> a <unk> one', '<unk> a <unk> a <unk> a', '<unk> a <unk> a <unk> one one', '<unk> a such <unk> a <unk> a', '<unk> a <unk> a <unk> a <unk> one', '<unk> a <unk> a <unk> one one one', '<unk> a such <unk> a such <unk> a', '<unk> a <unk> a <unk> one a <unk> one', '<unk> a <unk> a <unk> one a <unk> one one', '<unk> a <unk> a <unk> one a <unk> one a', '<unk> a <unk> a <unk> a <unk> one one one', '<unk> a such <unk> a <unk> one one one one', '<unk> a such <unk> a one one one one one', '<unk> a <unk> a <unk> a <unk> one one a', '<unk> a <unk> a <unk> one one one one a', '<unk> a such <unk> a one one one one a', '<unk> a <unk> a <unk> a <unk> one one', 'accelerated optimization', 'gradient flow', 'multi information', \"newton ' s method\", 'finite volume method', 'accelerated optimization methods', 'multi - step cell methods', 'finite element methods', 'differential equations', 'finite element method', 'finite volume methods', 'numerical world', 'a posteriori error estimate', 'numerical experiments', \"<unk> ' s method\", \"<unk> ' s algorithm\", \"<unk> ' s equation\", \"<unk> ' s equations\", 'elliptic equations', 'elliptic problems', 'elliptic', 'differential', 'finite', 'global optimization', 'restart', 'complexity bounds', 'restart schemes', 'multi translation', 'hardness of approximation', '. s . considering experts', 'low rank approximation', 'koopman operator', 'dynamic mode decomposition', 'least - squares', 'learning koopman group', 'learning koopman invariant subspaces', 'run length principle', 'piecewise linear regression', 'least - <s> subspace', 'regret minimization', 'iterates', 'run average', 'least that squares', 'modulated projection', 'compressible representations', 'soft - to - in - bounds', 'image compression', 'end - the the . is learning', 'end - the the . is coding', 'end - the the . is learning to the learning', 'end - the the . is learning to the of deterministic annealing', 'is . 1', '2', '1 <unk>', '3', '3 a', 'piecewise - geodesic trajectory', 'manifold - valued data', 'future less data', 'invariances', 'mixed - effect model', 'problems the guarantees', 'probabilistically triggered arms', 'semi - bandit', 'regret bound', 'combinatorial semi - bandit', 'probabilistically bounds', 'regret bounds', 'combinatorial multi - armed bandits', 'when by algorithm', 'new reinforcement learning', 'new sampling', 'regret tradeoff', 'new learning rates', 'new learning', 'regret', 'predictive - state representation', 'framework learning', 'for other of discovery', 'for other of latent the -', 'for for for framework', 'even and hold', 'even and odd states', 'even and odd states prediction', 'for other of latent the - which - down', 'even statistics', 'even function', 'latent', 'latent data', 'state - of of of of of of of of of of , screens', 'constructed dynamic has', 'constructed state an policies', 'constructed dynamic process', 'state - of of of of of of of of of of of , personalized medicine', 'state - of of of of of of of of of of of , personalized constructed dynamic has', 'state - of of of of of of of of of of of , personalized dynamic has', 'state - of of of of of of of of of of of of , , personalized and', 'state - of of of of of of of of of of of , personalized dynamic has a', 'state - of of of of of of of of of of of , personalized and off', 'state - of of of of of of of of of of of , personalized absorbing', 'constructed', 'state - of of of of of of of of of of of of of , personalized and', 'state - of of of of of of of of of of of , personalized constructed dynamic process', 'posterior sampling', 'worst - further infer', 'are our state information', 'mechanism design', 'at for propose', ') on post regression', 'post regression analysis', 'sampling', 'model of computation', 'on dynamics', 'problem complexity', 'on and the problem complexity', 'on and the worst and case analysis', 'on and the worst and case complexity', 'on and the worst and case selection', 'on and off and the sample complexity', 'semi - supervised learning', 'mean where epidemic', 'stimulus selection', 'role models: weight - averaged consistency its graphs', 'role models: weight - averaged consistency its', 'we a size', 'quantitatively', 'we a based of demonstrate', 'classifier ensembles', 'such a size', 'and statistics', 'classifier ensemble', 'role models: weight - averaged consistency its offer', 'stimulus was factorization', 'rely . guarantee', 'sharing', 'value', 'point', 'visualization', 'environments', 'discover in hardware', 'communication', 'dynamic', 'an object we', 'monte carlo', 'switch', 'group', 'of computation', 'features', 'an object', 'are and inside affinity', 'of analog', 'of', 'that constraints', 'correspondence search', 'correspondence prediction', 'svm', 'convolutional neural network', 'years episode', 'face recognition', 'object recognition', 'user interface', 'linear constraints', 'transformation', 'parameter estimation', 'covariance function', 'transformation lgn', 'fixed - rank approximation', 'positive - semidefinite matrix', 'method of moments', 'eigenvalues', 'eigenvectors', 'schatten 1 - norm', 'interactions', 'margin constraints', 'data streams', 'in of gaussian', 'generative adversarial nets', 'multi - modal learning', 'multi - agent learning', 'multi - agent systems', 'multi - fixed programming', 'unstructured of of learning', 'multi - fixed of learning', 'progressive coding', 'multi - stage are prediction', 'progressive architecture', 'quality', 'performance', 'context', 'proper orthogonal decomposition', 'modeling', 'informal', 'predict', 'contention', 'mean', 'full data compression', 'full data', 'baselines', 'and', 'and object', 'and texture', 'errors', 'of images', 'full in in in in in filtering', 'indirectly', 'full in in in in in in filtering', 'full in in in in filtering', 'indirectly in', '(', 'area', 'adaptive sampling', 'monte carlo em', 'riemannian geometry', 'riemannian samplers', 'targets tracking', 'alternating direction method of multipliers', 'stochastic optimization', 'variable selection', 'stochastic admm', 'physics - based modeling', 'generative modeling', 'analysis - by - synthesis', 'analysis - by - principles', 'analysis - by - principles modeling', 'shape inference', 'cifar prediction', 'shape analysis', 'approximate bayesian computation', 'likelihood - incorporated level designing', 'two of gaussian states', 'two of to coupled oscillators', 'likelihood - incorporated designing', 'paper . 0', 'false discovery rate', 'single - representation model', 'two of linear dynamical systems', 'two of linear mixed models', '<unk> algorithm', 'posterior distribution', '<unk> , <unk>', 'boltzmann', 'posterior', 'only', 'new', 'online prediction', 'selfish experts', 'prediction with expert advice', 'online algorithms', 'binary prediction with expert advice', 'selfish task', 'proper transferability', 'proper scoring rules', 'multiple criteria analysis', 'multiple criteria evaluation', 'multiple criteria decision making', 'multiple criteria decision analysis', 'noise', 'on o . m', 'selfish compatible?', 'channel equalization', 'bayes decision rule', 'hybrid linear/nonlinear', 'hybrid linear/nonlinear approach', 'hybrid algorithm', 'decision probability', 'high - particular', 'sketch', 'channel equalization problem', 'hybrid to such pca', 'sketch dynamic', 'this bounds', 'piecewise linear classifier', 'such channel equalizer', 'bounds', '( <unk>', 'hybrid to such channel equalizer of that a , decision if', 'hybrid to such channel equalizer of that a , decision neural networks', 'hybrid to such channel equalizer of that first kind', 'molecules', 'tensor biclustering', 'spectral clustering', 'community detection', 'multiple features', 'dynamic personalized screening', 'finite state stochastic process', 'absorbing disease', 'dynamic screening', 'demonstration', 'finite state significant instances', 'task demonstrations', 'electronic demonstration', 'an with been attention', 'an with been demonstration', 'with been and been attention', 'an with been and been attention', 'an with been blocks', 'an with been and model', 'an with been and been demonstration', 'an with been and been demonstrations', 'stopping criterion', 'along predictions', 'stopping criteria', 'applicability domain', 'in source learning', 'and probabilistic models', 'maximum mean discrepancy', 'learning on distributions', 'symmetric problem', 'level sets', 'learning on the a distribution', 'analysis of variance', 'decomposition of distributions', 'analysis of variance components', 'error bounds', 'we triggered average', 'event sequence clustering', 'dirichlet mixture model', 'dirichlet real process', 'simpler algorithm', 'rates our', 'statistical machine learning', 'dirichlet best', 'dirichlet process', 'best which', 'model feature', 'a dimensional mixture model', 'observed dynamics', 'observed data', 'spaces gathered log', 'rates of distributions', 'observed information', 'cryptography', 'peer - to - peer networks', 'subsequently', 'peer - to - peer', 'system?s', 'bitcoin', 'anonymity', 'distributed computing', 'distributed hash table', 'we , for bitcoin', 'on . <unk>', 'matrix factorization', 'on . in digital systems', 'network security', 'optimality', 'according a service', 'according a network', 'be a a graph', 'be a a a graph', 'network task', 'be a a a a graph', '1 . 2', '1 . 2 , 2 , 2 , 2 ,', '1 . 2 , 2 , 2 , 4 ,', '1 . 2 . 2', '1 . 2 , 2 , 2 ,', '1 . 2 , 2 , 2 , 2 .', '1 . 2 , 2 , 2 . 4 ,', '1 . 2 , 2 , 2 . 2 ,', '1 . 2 , 2 , 4 , 2 .', '1 . 2 , 2 , 2 , 4 .', '1 . 2 , 2 , 4 , 4 ,', '1 . 2 , 2 , 4 ,', 'consensus', 'min - sum splitting', 'find', 'saddle point problem', \"except ' s criterion\", 'models of computation', \"except ' s algorithm\", '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 .', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 . 2 ,', 'generalized linear models', 'distance - to - set to a convex we', 'illustration', 'distance - to - set to a prior', 'constrained optimization', 'incur', 'shrinkage', 'lasso', 'first of fit', 'programming of experiments', 'dimension reduction', 'adaptive stimulus selection', 'mid - level visual processing', 'neural', 'approach to noise ratio', 'the <unk> rule of interest rate', 'in vivo experiment', 'in situ cycle prediction', 'in situ terms selection', 'spike timing analysis', 'cascade model', 'influence', 'nonbacktracking walk', 'trade t', 'seed set', 'message problem', 'nonbacktracking walks', 'community assumptions', 'cascade models', 'unit boltzmann machines', 'directional networks', 'mean - field approximation', 'directional unit boltzmann machines', 'von mises distribution', 'circular range forms', 'activation', 'simulation', 'unit boltzmann machine', 'computation', 'direct', 'probability', 'extensibility', 'representation', 'combinational', 'component', 'variability', 'timing', 'streaming learning', 'feature streams', 'ensemble methods', 'ensemble learning', 'sensor networks', 'are improvement', 'feature selection', 'fault tolerance', 'online optimization', 'stochastic constraints', 'data center', 'invariant features', 'semisupervised learning', 'one - shot t bandit', 'cognitive impairment prediction', 'regularized modal regression', 'variable we graphs', 'regularization', 'modal regression', 'aggregating graphs', 'aggregating bandit condition', 'its', 'show through time', 'show through rate', 'outlier detection', 'show through patterns', '1 in situ hybridization', 'translation synchronization', 'truncated least squares', 'truncated least square', 'link analysis', 'social network analysis', 'kernel learning', 'small', 'presence', 'spatiotemporal data mining', 'alternating direction method', 'variable splitting', 'alternating direction methods', 'patterns', 'sparse recovery', 'interior - point methods', 'alternating in point methods', 'systems of linear equations', 'sparse matrices', 'systems of linear programming', 'vanished', 'to ensure algorithm', 'low dynamics', 'systems of and pdes', 'high dynamics', 'duality', 'inside neural networks', 'inside neural network', 'how c', 'so dynamics', 'fast algorithm', 'fast direct search methods', 'c', 'fast algorithms', 'fast', 'how', 'continuous dueling bandit', 'dueling bandit', 'regret analysis', 'continuous space', 'classification trees', 'best response prediction', 'best response regression', 'prediction duel', 'probably area estimation', 'probably approximately correct learning', 'probably area prediction', 'classification functions', 'distributed deep learning', 'ternary gradients', 'gradient clipping', 'terngrad', 'issue protein', 'group sparsity', 'fundamental matrix', 'distributed training', 'distribution be o dirichlet', 'expected pair', 'doubled path', 'in an nodes vector', 'order dirichlet problem', 'doubled populations', 'demonstrated', 'on - line training', 'oscillatory neuronal responses', 'phase density dynamics', 'pdd', 'bistable successfully solve', 'multiple conservation laws', 'assembly', 'rate dynamics', 'to conservation', 'multiple conservation', 'time ground reaction', 'penalization', 'to <unk> to <unk> on', 'to <unk> to <unk> on time constraints', 'for on', 'on <unk> line dynamics', 'direction <unk> of <unk> arrival estimation', 'direction of arrival', 'direction <unk> of <unk> arrival', 'direction <unk> of <unk> arrival times', 'direction <unk> of <unk> arrival time', 'image segmentation', 'affinity matrix', 'spatial propagation networks', 'spatial propagation network', 'parallel machine learning', 'correspondence problem', 'linear regression', 'fully polynomial - synthetic data', 'may to decision ratio', 'from data', 'signal algorithm', 'may to decision number', 'from data to mixture data', 'optimality conditions', 'methods and algorithms', 'exponential family', 'may to decision number ratio', 'discretized min - sum rule', 'exponential families of distributions', 'exponential smoothing', 'electrophysiology', 'fast fourier transform', 'discretized min - sum', 'challenging', 'complex variables', 'fails ,', 'fast fourier', 'fails , 1 , 1', 'challenging to', 'fails', 'discovery thresholds', 'hypothesis weighting', 'hypothesis testing', 'discovery threshold', 'gradient boosting', 'source coding', 'regression trees', 'composition', 'abstraction', 'realization', 'rule realization', 'rule selection', 'music interpretation', 'probabilistic', 'music', 'probabilistic composition', 'probabilistic rule realization', 'in that composition', 'probabilistic realization', 'multiclass learning', 'sample compression', 'bayes consistency', 'bayes - consistency', 'nearest - neighbor', 'sample data compression', 'sample data data compression', 'metric spaces', 'in . <unk>', 'high dimensionality', 'rule learning', 'high . <unk>', 'in . 1 . neural networks', 'stochastic bandits', 'bounded kurtosis', 'scale free', 'finite - armed bandit', 'finite - armed bandits', 'scale single output', 'scale single another algorithm', 'scale single output problem', 'bounded that move', ') algorithm', 'output feedback', ') complexity', 'in data sets', 'develop time', 'output complexity', 'it data structures', 'online algorithm', 'in data set', 'as have feature', 'multilinear relationship networks', 'multi - task learning', 'human machine learning', 't . 0', '1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', '1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', '1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', '1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1', '1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1', '1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', '1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', '1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', 'hyperalignment', 'svd', 'high - dimensionality', 'stochastic decoys', 'nonlinearity', 'functional alignment', 'stochastic gradient both', 'relies', 'experimental design', 'factorization', 'functional algorithm', 'pomdp', 'synthetic data', 'speeding . moreover', 'speeding . propose', 'speeding . if', 'speeding . p .', 'speeding . if analysis', 'speeding . minimizer', 'inferences', 'speeding . p . our', 'speeding . p . machine', 'minimizing property', 'speeding . p . process', 'recommender systems', 'dynamic programming', 'generators', 'is ( performance', 'categorical variables', 'kalman filtering', 'point processes', 'of gaussian processes', 'offline', 'dimensional spaces', 'randomized algorithms', 'statistical analysis', 'performance measures', 'synaptic coupling', 'chaos', 'connectivity', 'finite sum', 'variance reduction', 'data augmentation', 'data analysis', 'topological signature', 'persistent homology', 'topological signatures', 'summary representation', 'differential - difference equations', 'conditional parameter estimation', 'differential likelihood', 'metropolis - hastings algorithm', 'conditional probability mass function', 'differential - difference equation', 'log at random process', 'and poisson process', 'stochastic process', 'poisson process', ', poisson process', 'user algorithm', 'language modeling', 'user process', 'random walks', 'problem difficulty', 'alternating direction method of arrival', 'of of of', 'empirical cumulative distribution function', 'revenue learning', 'uniform convergence', 'uniform pruning', 'a posteriori probability bounds', 'sample - complexity bounds', 'it inequality', 'generative power', 'it to ) efficiency', 'generating functions', 'first hamming divergence', 'first hamming distance', 'deep mixing', 'deep belief networks', 'feature markets', 'first true of tailed distributions', 'first', 'first and long and long tail', 'first and long tail', 'first true', 'first and', 'poisson factor analysis', 'deep dynamic poisson factorization', 'variance component analysis', 'integrates a furthermore', 'that furthermore experiments', 'deep dynamic poisson factorization model', 'approach to rank ratio', 'for reinforcement learning', '( 2 that within', 'graphical model', 'precision agriculture', 'precision data', 'non - negative risk estimator', 'positive - unlabeled rates', 'equations regularization', 'context we', 'positive unlabeled learning', 'do tasks selection', 'positive - unlabeled learning', 'non - negative risk a - we', 'computer convergence', 'constants of art', 'constants distributional benchmark', 'computer learning theory', 'non - negative risk a assumed we', 'data regularization', 'can', 'can of one', 'can of for for neural', 'can of for faster against neural learning', 'can of for faster against by one', 'can of for for against neural', 'can of for faster against by learning', 'top - k ranking', 'partially revealed preference', 'm - wise data', 'minimax optimality', 'm - wise comparison', 'sample size', 'counterfactual model', 'neyman [ 1923', 'what - if reasoning', 'decision - making', 'neyman [ 1923 ]', 'what - if analysis', 'reliable decision - making', 'rubin [ 1978', 'rubin [ ]', 'what - if', 'rubin [ if', 'rubin [ ] ]', 'rubin [ ] if', 'rubin [', 'rubin [ 1978 ]', 'rubin [ if [ if', 'rubin [ if [ ] ]', 'rubin -', 'rubin [ if [ ] if', 'image encoding', 'vector quantization', 'image develop', 'heuristic methods', 'speech', 'passing style', 'priors', 'may . <unk>', 'not', 'reduction', 'ensembles', 'not . <unk>', 'not . of', 'practical', 'may . on', 'proximal point method', 'nonequilibrium', 'proximal ( towards physics', 'comparable elements', 'requires set', 'olfactory signal', 'insect', 'coding', 'olfactory system', '\"generalization walk', 'generalization gap', '\"random walk', 'random landscape\" statistical model', 'batch means', 'batch normalization\"', 'batch training', 'adaptive numerical format', 'neon deep residual network', 'floating point format training', 'neon deep previous both', 'floating', 'mcmc', 'marginal likelihood', 'annealed importance sampling', 'learns', 'markov chain mechanism', 'plus', 'plus <unk> monte carlo', 'minimal exploration', 'stochastic bandit problems', 'instance - specific regret lower bounds', 'instance - specific regret lower observation', 'thompson sampling', 'instance - specific the we bound', 'instance - specific the we', 'systems biology', 'systems of optimality', 'compressive image recovery', 'compressive sensing', 'compressed sensing', 'adaptive algorithms', 'message provide', 'message provide alignment', 'may of architectures', 'may of optimality', 'are of one', 'parallel algorithms', 'multi of linear systems', 'multi of linear regression', 'deliberation networks', 'deliberation', 'encoder - decoder', 'bleu', 'deals', 'estimator', 'deliberation process', 'encoder - decoder framework', '. the only , statistics', 'deliberation n', 'called only , statistics', 'k - means', 'exact clustering', 'high - dimensional data', 'cost function', 'state/action values', 'bellman equation', 'log - normal ha control', 'prior distribution', 'learning outcome', 'deep inference', 'log - normal distribution', 'state/action vs . <unk>', 'learning which', 'learning attempt which', 'for algorithms', 'in networks', 'on networks', 'state/action and to state/action can however and <unk> is not supplied', 'repeated inverse problems', 'q <unk> learning', 'games', ', q <unk> learning', ', games', 'method of types', 'repeated', 'method of lines', 'games with a purpose', 'equation of state', 'two - player games', 'generative networks', 'its applications', 'generative adversarial networks', 'programming and algorithm theory', 'smooth two - rate games', 'integer linear programming', 'smooth', 'integer nonlinear programming', 'method of characteristics', 'speech production', 'speech synthesis', 'a provide learning', 'neural interactions', 'exact prediction', 'exact learning', 'motor control', 'the the objective measurement', 'conversions', 'arms race', 'model fitting', 'bayesian optimization', 'direct search', 'bayesian', 'adaptive direct search', 'high - dimensional model we', 'chordal markov networks', 'constraint optimization', 'functional genomics', 'branch and networks', 'is motion planning', 'for systems', 'is motion', 'observation application', 'and machine learning', 'for approach', 'erm', 'observation <unk> based learning', 'from to machine learning', 'learning algorithms', 'learning vector quantization', 'computational complexity', 'bias', ') methods', 'induced', 'revenue optimization', 'integer programming', 'bid price', 'bid predictions', 'bid optimization', 'random quadratic equations', 'most systems of quadratic equations', 'information . tolerate', 'samples ,', 'information', 'weighting', 'independent component analysis', 'relate probability learning', 'relate probability of learning', 'relate smaller results', 'autoencoder', 'sequential learning', 'hierarchical graphical models', 'variational autoencoder', 'mixture models', 'sequential bayesian learning', 'sequential generative models', 'multi and set learning', 'multi and dnns', 'simulations', 'lookahead', 'inequality constraints', 'inequality constraint', 'spectral methods of moments', 'hierarchical methods . auctions', 'latent variable models', 'spectral methods . to a on by ,', 'spectral methods . to a on moments', 'spectral methods . to a on by moments', 'with', 'are of chaos', 'tensor decomposition', 'hierarchical methods of to to a on by a are approach', 'spectral methods . to a on by a are approach', '1 <unk> projection', 'hierarchical methods of to to a on by a are a neural networks', 'adapt methods the at random variables', 'approach the at random variables', 'adapt methods the at half <unk> linear methods', 'adapt methods the at half <unk> linear models', 'risk minimization principle', 'empirical sequence analysis', 'empirical orthogonal decomposition', 'networks of workstations', 'x <unk> ray spectroscopy', 'x <unk> ray divergence', 'x <unk> ray crystallography', 'textual grounding', 'deep net', 'bounding box', 'deep nets', 'image concepts', 'fuzzy neural networks', 'generalized hamming distance', 'thresholding', 'rightful bias', 'properties', 'generalized hamming network', 'batch convex hull', 'fuzzy logic', 'properties multitasking networks', 'properties multitasking network', 'latent variable gaussian graphical model', 'low - rank matrix', 'we systems biology', 'network tomography', 'network flows', 'visual cortex', 'filter selection', 'velocity estimation', 'gain control', 'neural model', 'filter selection model', 'covered regions', 'active vision', 'neural models', 'batch renormalization', 'deep models', 'partial least squares', 'cross - sectional dependence', 'training and inference', 'logistic regression', 'information hiding', 'steganography', 'adversarial training', 'transport distances', 'sinkhorn iteration', 'optimal transport', 'coordinate descent', 'transport distance', 'greedy methods', 'in network distance', 'categorical prior', 'autoencoders', 'latent code', 'multitask learning', 'nonlinear output relations', 'universal consistency', 'structured prediction', 'universal', 'consistency', ', regularization', ', , reach', ', , . 1', ', , . <unk>', ', finite mixture models', ', , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 1 , 2 , 1 ,', ', , . 1 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 1 , 2 , 1', ', , . 1 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 1', ', , 1 , 2 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 1 , 2 , 1 ,', ', , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 1 ,', ', , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 2 ,', ', , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 1 , 1 , 2 ,', ', , . 1 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 2', ', , . 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 2 , 1 , 1 , 1 , 1 , 1', ', , . 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 1 , 1 , 1', 'dictionary learning', 'alternating minimization', 'sparse coding', 'random initialization', 'sparse representation', 'initialization', 'sparse', 'relus', 'planted model', 'adversarial network training', 'deep generative models', 'vlsi', 'measuring greedy selection', 'models stabilization', 'face detection', 'face tracking', 'models selection', 'log ) we', 'log ) curves', 'an the the on follow ) sampling', 'an the the on follow the we', 'expectation propagation', 'stochastic kinetic model', 'bethe free generalization', 'embeddings', 'kinetic model', 'bethe free energy', 'particle filter', 'show through', 'outperformed signal', 'favorable prior information', 'stochastic kinetic output model', 'favorable prior distribution', 'data - efficient learning', 'planning under uncertainty', 'data - efficient planning', 'continuous state - action gaussian pomdps', 'schema', 'cerebellum', 'signal processing', 'elemental pattern generator', 'into <unk> network', 'adapt', 'systems', 'input', 'matches', 'popular competing show', 'studied learning', 'implies', 'studied we <s> black', 'local linear control', 'compared', 'linear prediction', 'on produces', 'neural network model', 'communication systems', 'by largest eve effect', 'effective', 'effective guarantees', 'connection structure', 'spike trains', 'reward functions', 'inverse reinforcement learning', 'efficient algorithms', 'e as topic', 'e as a service', 'g a as a vector space', 'by crr crr', 'by crr date', '<unk> in the far', 'by example crr', 'adaptive sample size', 'first - order methods', 'number of samples', 'guaranteed convergence rate', 'to and number of probability learning', 'guaranteed convergence bounds', 'algorithm complexity', 'stochastic method', 'to to i data', 'number of thresholding', 'popular sampling', 'parallelizing compilers', 'parallelizing compiler', ') for neural networks', 'plain more of runtime', 'large deviations', 'covert channels', 'mutual information', 'plain', 'watermarking', 'data hiding', 'both', 'demonstrate', 'deep', 'and evidence', 'and security', 'cross - task knowledge transfer', 'transfer learning', 'neural programming', 'meta induction', 'extreme data sparsity', 'neural program induction', 'k - shot learning', 'experimental evaluation', 'meta - program induction', 'meta program induction', 'portfolio adaptation', 'bayesian regression', 'dyadic trees', 'histograms', 'regression surfaces', 'regression splines', 'dyadic regression trees', 'circumference', 'multitasking', 'graph theory', 'parallel processing', 'binding', 'measurement', 'use', 'addressing', 'lighting', 'extreme learning machine', 'interaction', 'extreme value theory', 'observability', 'robust regression', 'ordered lists', 'consistency analysis', 'robust linear regression', 'two - stage sparsity', 'two - stage model', 't', 'two', 'order statistics', 'order restricted inference', 't estimator', 'to is methods', 'to is if', 'natural value estimation', 'or in telecommunications', 'reduce batch integrate - and - fire', 'reduce batch integrate - and - fire highlights', 'reduce batch integrate - and - learning', 'bandits', 'partially ordered sets', 'normalization', 'partially is', 'optimal experimental design', 'elementary symmetric polynomials', 'elementary we', 'elementary symmetric polynomial', 'elementary', 'symmetric polynomials', 'elementary symmetric optimization', 'elementary quite', 'first order approximation', 'analog vlsi', 'parallel computing', 'neural network learning', 'gradient convex optimization', 'changes', 'changes in algorithm', 'gradient and posets', 'natural language processing', 'first passage percolation', 'first be french language', 'first be and second language', 'logarithm', 'first passage time', 'in in complex systems', 'secret sharing', 'in in bob worlds', 'in in silico experiments', 'we compression', 'asynchrony', 'first passage times', 'coin betting', 'a . <unk>', 'learning rate', 'a priori information', 'continuous optimization', 'learning , <unk>', 'bayesian estimation', 'learning , <unk> a', 'in , stochastic optimization', 'continuous stochastic optimization', 'continuous time markov chain', 'continuous time markov chain monte carlo', 'networks of information', 'associative embedding', 'scene graph', 'graph generation', 'scene graphs', 'runtime neural pruning', ')', 'neural pruning', 'eigenvalue decay', 'learning function classes', 'polynomial - time algorithms', 'polynomial - time algorithm', 'polynomial - model', 'data complexity', 'commonly . problem', 'format equivalence', 'data approximation', 'format . <unk>', 'converge', 'provably efficient', 'moment matching network', 'generative moment better', 'samples demonstrate', 'kernel maximum maximisation', 'process modeling', 'network optimization', 'generative moment matching network', 'adversarial samples', 'we method', 'deep residual networks', 'reversible residual performance', 'data reduction', 'deep residual network', 'bandit optimization', 'while choosing', 'with tasks gan', 'upper - confidence in neural networks', 'in for optimality', 'on and line learning', 'be the upper ( <unk> labeling', 'be the upper ( labeling algorithm', 'be the upper ( <unk> labeling algorithm', 'for optimality', 'it x x x', 'it x x', 'it x x distribution', 'q - learning', 'matrix - vector quantization', 'matrix - vector vector quantization', 'asymptotic variance', 'relaxations', 'guarantees erm', 'functions of bounded', 'functions of <unk>', 'guarantees or <unk>', 't - exponential family', 'q . time series', ', procedure of classification', 't - exponential a', ', further and procedure', 'scalable learning', 'expectation distance', 'expectation algorithm', 'expectation is', 'expectation for and', 'it , education', 'analog noise', 'multilayer perceptron training', 'for principal component analysis', 'adaptive resonance theory', 'learning neural single hidden layer feedforward neural polynomially', 'learning neural single of hidden layer feedforward neural networks', 'learning neural single hidden layer feedforward neural networks', 'learning neural single of hidden layer feedforward networks', 'control variates', 'we . template', 'lehman algorithm', 'be learning complex exploration', 'self of associative memory', 'self of consciousness', 'self of can growth', 'be learning complex we detection', 'few - shot learning', 'information activity', 'validation', 'benchmark', 'few - for of of the k2', 'uses and gratification', 'information observations', 'orders of minimization', 'orders of information retrieval', 'orders of information information retrieval', 'promising develop', 'promising image', 'orders of information information information retrieval', 'promising and action', 'information information based as topic', 'orders of information information information based as topic', 'adversarial manipulation', 'interactively', 'adversarial classification', 'adversarial easier easier', 'leverage', 'helps instances', 'of classifiers', 'leverage points', 'object detection and as a service', 'object tracking', 'demonstrating', 'object detection and a chip', 'scene and object tracking', 'scene and object detection', 'scene in scene analysis', 'video surveillance', 'scene clustering', 'scene ( video surveillance', 'scene ( video', 'scene understanding', 'object ( a chip', 'scene in', 'heavy hitters', 'private heavy hitters', 'sampling algorithms', 'classifiers', 'manipulation', 'private classifiers', 'quadratic programming', 'low - discrepancy sequence', 'produces method', 'low - discrepancy segmentation', 'quadratically constrained quadratic programming', 'low - can can', 'particular solution', 'models error', 'models error bound', 'quadratically constrained quadratic program', 'models of <unk>', 'models error brittle', 'obtaining method', 'large - scale optimization', 'detection of define the can', 'heavy the maximum likelihood', 'hypergraph partitioning', 'inhomogeneous partitioning', 'verify', 'n <unk> gram', 'dimensionality curse', 'laplacian eigenmaps', 'logical rules', 'neural logic programming', 'knowledge', 'knowledge base reasoning', 'artificial intelligence', 'computational intelligence', 'training algorithms', 'of reasoning', 'neural logic need', 'and object representations', 'as', 'outperforms', 'deep multi - task gaussian process', 'importance sampling', 'survival analysis', 'survival problems', 'density estimation', 'autoregressive processes', 'autoregressive flow', 'guarantees', 'autoregressive process', 'non - convex finite - sum on costs', 'in and vector optimization', 'information criterion', 'image denoising', 'action selection', 'unconstrained optimization', 'information measure', 'smoothing', 'parameters selection', 'nonlinear optimization', 'monte carlo methods', 'estimates', 'learning control', 'reactive control', 'extreme uncertainty', 'peg - in the the a . for', 'peg - in the the a . for neural networks', 'peg - in the the a . for of neural networks', 'is control', 'peg - in the the a . for of of is', 'efficient task learning', 'sparse decision trees', 'transport', 'time series', 'sparse classes', 'produce', 'sparse figure sets', 'low in graphs', 'in a method of operate', 'elf games', 'at and free optimization', 'at and energy minimization', 'in a priori optimization', 'elf', 'as random variables', 'is random variables', 'as random search', 'this this this algorithm', 'for this algorithm', 'this this neural learning', 'this this this neural learning', 'this this this nonlinear mapping', 'this this this neural learning algorithm', 'this this this class data random', 'for for for for for for this', 'for for for for for for for', 'for for for for for for for this', 'for for for for for for for for', 'for for for for for this this', 'for for for for for for this this', 'for for for for for this for', 'calcium imaging', ') deep brain imaging', 'online dictionary learning', 'on deep learning', 'policies', 'in works data', 'in works data analysis', 'real - time imaging', 'based segmentation and comparison', 'in works data generative model', 'in works and network analysis', 'in works data generative models', 'as discrepancy', 'as discrepancy search', 'are discrepancy', ', , art', 'as discrepancy measure', 'automatic differentiation', 'polynomial chaos', 'trust', 'discrete time', 'demand - aware recommendation', 'e - commerce', 'demand - aware recommender systems', 'low . paper', 'low . paper personalization', 'low - rank tensor completion', 'scale - free filtering', 'accurate scalability', '1 the learning', 'rather . moment can', '1 the proposed our', '1 . 2 . 2 . 2', '1 . 2 . 2 . 2 <unk>', 'personalization', 'performance evaluation', '1 . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 1 <unk>', 'performance measurement', 'polynomial', 'log', 'competition', 'estimation of distribution algorithms', 'dependent random variables', 'from models', 'continuation', 'inter scalable learning', 'fully implicit learning', 'mean 1 produces learning', 'structure', 'vectorization', 'a models', 'distributions', 'families', 'positive linear discriminant', 'positive', 'positive linear', 'inter scalable network', 'positive linear regression', 'noisy labeling', 'pairwise comparison oracles', 'interactive learning', 'select', 'lacks fast', 'second we correct', 'second we talk', 'lower bound', 'it 1 performance', 'to accuracy tradeoff', 'body', 'to accuracy', 'contains', 'ordinary least', 'binary', 'proposed t', 'contains their . through', 'ordinary differential', 'speech recognition', 'end - to the the proposed natural language processing', 'duration modeling', 'given method', 'it . <unk>', 'speaker verification', 'speaker recognition', 'datasets', 'kernel regression', 'local metric learning', 'much data', 'number of patterns', 'h we', 'electrosensory system', 'descending gain control', 'control', 'analog model', 'analog neural network', 'analog modeling', 'number of false alarm', 'analog this model', 'control system', ', a model', 'a posteriori', 'approach to optimality', 'high is discovery', 'approach to rank', 'models of markov learning', 'approach of algorithms', 'high based regularization', 'tucker decomposition', 'low - rank algorithms', 'low - rank properties', 'tucker algorithms', '<unk> algorithms', 'order information', 'is approach', 'tensor rank', 'low we', 'outputs', 'order', 'plays data', 'our on on on rank', 'are as matrix', 'all approach', 'our on on on line', 'occur method', 'regularized least bound', 'our on on speech', 'regularized least squares', 'are to number', 'members the algorithms', 'issue approach', 'issue', 'all', 'deep supervised hashing', 'image retrieval', 'discrete hashing', 'deep hashing', 'action recognition', 'deep supervised learning', 'content - based image retrieval', 'family of classifiers', 'family of random trees', 'a posteriori probability density estimation', '. 3 calcium', 'secondary', '. 2', 'family of random variables', 'a posteriori probability density function', '. 3', '.', 'options', 'covariance', 'policy its', 'off ( gan', 'black box model', 'off - policy gan', 'that ( <unk> for for samples', 'in on on on on demand', 'in on on on on on demand', 'follows', 'in on on on on on time', 'improving value at risk', 'in on on on on on on time', 'in on on on on the time', 'in on on on demand', 'in on', 'in', 'discretization', 'stochastic', 'critical algorithm', 'city we', 'linear neural networks', 'optimal control', 'critical guarantees', '<s> evidence theory', 'analyzing', '<s> a priori information', 'all machine learning', '<s> a d vi l 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2 2 2', '2 2 2 2 2 2 2 2 2 2 2 2 2 2', '2 2 2 2 2 2 2 2 2 2 2 2 2 2 2', '2 2 2 2 2 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2 2 2 2', '2 2 2 2 2 2 2 2 2 2 2 2', '2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2 2 2 2 2', '2 2 2 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2 2 2 2 2 2', '2 2 2 2 2 2 2 2 2 2', '<s> a d vi l 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2', 'fisher gan', 'integral probability metrics', 'sample', 'entropy', 'modelling', 'generalization capability', 'cifar model', 'activations', 'to leverage', 'detection', 'to leverage clustering', 'in we', 'posterior probability', 'rate of convergence', 'mutual oracles', 'mutual order', 'sample entropy', 'entropy rate', 'nonnegative matrix factorization', 'kullback . <unk>', 'kullback . leibler divergence', 'our k . 2 . 2 . 2 . 2 <unk> . 2', 'our . 2 . 2 . 2 . 2 . 2 <unk> . 2', 'our k . 2 . 2 . 2 <unk> . 2 <unk> . 2', 'our . 2 . 2 . 2 . 2 <unk> . 2', 'our', 'our . 2 . 2 . 2 . 2 <unk> . 2 <unk> .', 'our k . 2 . 2 . 2 . 2 <unk> . 2 <unk>', 'our . 2 . 2 . 2 <unk> . 2 <unk> . 2', 'our k . 2 . 2 . 2 . 2 . 2 <unk> .', 'our k . 2 . 2 . 2 . 2 . 2', 'our k . 2 . 2 . 2 <unk> . 2', 'arbitrarily conditioned linear systems', 'consistent independent estimation', 'rigorous estimation', 'leads accuracy', 'we accuracy', 'most distance accuracy', 'polynomial on we', 'leads free', 'on accuracy', 'parameters estimation', 'parameters also', 'parameters', 'private', 'on', 'deep neural network', 'trigeminal', 'whisker array', 'multivariate poisson hidden markov model', 'be', 'maps of structure', 'differential privacy', 'our - estimated our', 'privacy', 'accuracy conversely', 'from - two', 'is each model', 'is each satisfies', 'price of trust', 'price of anarchy', 'a set bound', 'revenue maximization', '( risk items', 'price of disclosure', 'price of risk', 'codes', 'deepen be', 'new the networks', 'new better than used', 'frame problem', 'experiments', 'new the solutions', 'frame systems', 'spectral learning', 'ones property', 'moments', 'weighted automata', 'less in algorithm', 'reversible jump', ', satisfying bounds', 'one quadratic functionals', 'on methods', 'one quadratic functions', 'one quadratic forms', 'person demonstrate', 'model methods', 'multi - way interactions', 'gibbs substantiate', 'factorization multilayer perceptron', 'normalizing constant', 'error analysis', 'draws algorithm', 'draws sequential monte carlo', 'bayesian in time series', 'gibbs', 'draws sequential sampling', 'draws sensing data analysis', 'complementing gradient', 'gibbs algorithm', 'gibbs bayes', 'gibbs information', 'directly free', 'weisfeiler - lehman network', 'organic exploration', 'molecular modeling', 'exploration', 'weisfeiler - lehman', 'organic reaction', 'graph search', 'molecular docking', 'computational geometry', 'metric compression', 'theory', 'data - estimates', 'metric', 'studies of algorithms and problem complexity', 'studies of algorithms', 'data compression', 'on to line learning', 'graph algorithms', 'analysis of algorithms and problem complexity', 'coding and information theory', 'theory of computation', 'graph', 'original', 'by', 'graph algorithms and data structures', 'generative', 'discrete latent variable models', 'gradient estimation', 'variance estimation', 'random sampling', 'high variance dimensional models', 'on to latent variables', 'on , line learning', 'variance', 'random matrix theory', 'to states', 'information asymptotically depend', 'to states ratio', 'asymptotically depend', 'learned parametrizations', 'learned passing style', 'seek time', '. 1 . 1', 'seek', '. 1', 'barycenters', 'built to peer', 'these algorithms', 'distributed sensing', 'learning from streams', 'good of of streams', 'good arrays', 'good on on on on arrays', 'good on on on on on arrays', 'good of', 'good of of', 'good of optimality', 'good', 'good on on on arrays', 'good of strategy', 'real - time strategy', 'count data', 'count data model', 'must', 'and many . 0', 'must . 0', 'and computationally distribution systems', 'and computationally distribution system', 'mse', '+ can', 'multi . 0', 'built', 'multi . <unk>', 'and linear programming', '+', 'multi . lingual measures', 'multi', 'this', 'robust pca', 'flexibility', 'self - masked data', 'outlier', 'by learning', 'difference learning', 'self - organizing rules', 'principal autoregressive model', 'proposed systems', 'by activations', 'mode collapse', 'kullback - leibler divergence', 'advertising', 'by . <unk>', 'array signal processing', 'most', 'array signal', 'array', 'most .', 'revenue sharing', 'auctions', 'dynamic pricing', 'decomposition - invariant conditional gradient', 'low a convex optimization', 'conditional gradient', 'decomposition - invariant points', 'low and rank approximation', 'such learning', 'knowledge discovery', 'likely stabilization', 'his', 'knowledge acquisition', 'predictive modeling modeling', 'compose . 0', 'electricity markets', 'multiagent predictive modeling', 'multi - the of agents', 'our ( <unk> equilibrium', 'our ( <unk>', '1 . 2 . 2 <unk>', 'multi - agent predictive modeling', 'a of optimality', '1 . 2 . 2 <unk> <unk>', '1 . 2 . 2 . 2 . 2 <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 <unk> <unk> <unk>', '1 . 2 . d . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk>', '1 . 2 . d . 2 . 2 . 2 . 2 <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk> <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk> <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk>', 'd . 2', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', 'd . 2 .', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . d . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . d . 2 . 2 <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . d . 2 . 2 .', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk> <unk>', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . d . 2 . 2', '1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 .', 'and and the prior knowledge', 'model algorithms', 'model validation', 'and and the the the is validation', 'and and the the the is gan database', 'and and the the the is gan dataset', 'and and the the the is validation data', 'as learning motivation', 'as learning least squares', 'in of mean', 'in of gaussian models', 'in of parameters', 'in of', 'in of on of parameters', 'in of optimality we', 'in situ', 'differentially private empirical risk minimization', 'and in approach', 'of optimality conditions', 'differentially private', 'and in can square', 'associative risk minimization', 'is polynomial', 'and in - bandwidth', 'and in can square detection', 'domain decomposition', 'practical 1 polynomial', 'non that new function .', 'non that new 1 ,', 'generated convex atoms', 'non that new 1 , his experts , <unk> classes', 'non that new 1', 'the generated convex atoms', 'since', 'but', 'but discriminate analysis', 'generated convex information', 'the edits', '<unk> benefiting', '<unk> bridging', 'but discriminate analysis ) p examine', '<unk> points', '+ <unk>', '<unk> -', 'also', 'acoustic emission', 'algebra of random variables', 'model e product', 'variational speech recognition', 'recommendation', 'time ) space', 'time series analysis', 'time series capable of interest', 'non ) non ) complex', 'non ) non ) non ) linear model', 'non ) non ) non ) standard model', 'rule of interest', 'non ) non ) non ) standard deviation', 'non ) non ) non ) more sampling', 'annotated field', 'assumption', 'learned annotated field', 'annotated', 'learned', 'nonconvex regularized sparse learning', 'proximal in ins', 'function optimization', 'give <unk> optimality', 'count - based exploration', 'markov distribution', 'markov furthermore', 'linear temporal logic', 's of optimality', 'count - based ( system', 'kernel machines', 'local learning', 'wfa', 'holographic recurrent networks', 'associative memory', 'holographic', 'temporal coding', 'neural computation', 'hybrid systems', 'preserves idea generation', 'self <unk> organization', 'first <unk> order neurons', 'first <unk> order continuous attractors', 'self organization', 'residual generation', 'emergent behavior', 'show <unk> through', 'indices', 'self', 'self <unk>', 'recurrent', 'feedback', 'policy', 'n', 'heuristic', 'measured', 'duration', 'value and policy purpose', 'our and properties', 'n , measured and measured problem', 'on . also and necessarily decisions', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 2 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 1 , 2 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 1 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 1 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 2 , 1 , 2 , 1 , 2 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 2 , 1 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 2 , 1 , 2 , 2 , 1 ,', 'n , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 2 ,', 'theorem proving', 'graph embedding', 'renaming', 'deep graph embedding', 'premise selection', 'to gain ratio', 'be . <unk>', 'for an graph embedding', 'model direct reduction', 'an graph embedding', 'an graph features', 'for theorem order empirical evaluation', 'model model selection', 'graph laplacian', 'for an graph features', 'of the sparsity', 'of the algorithm', 'measures of association', 'on <unk> line learning', 'gpu', 'gpu programming', 'gpu computing', 'on <unk> line training', 'fps', 'on built <unk> in self <unk> organizing maps', 'on built <unk> in self <unk> organizing map', 'for supervised learning', 'on built <unk> in self <unk>', 'on built <unk> in self built', 'unsupervised', 'on built <unk> in self test', 'manifold learning', 'coordination', 'geometry', 'space', 'there <unk> states', 'problems', 'of equations', 'operating system', 'connections', 'of models', 'data', 'softmax value function', 'challenged network', 'analyze', 'facilitates', 'policy - gradient', 'structured lower bounds', 'operations research', 'developed is', 'problem reduction', 'developed illustrate', 'hedge algorithm', 'tree', 'methodology', 'association', 'intern', 'spanned group', 'k matrix', 'originally while tree', 'more', 'gain', 'trees', 'denotes', 'former', 'effective dimension', 'quantities of interest', 'number of false alarms', 'originally [ and [ bound', 'multi - response linear model', 'alternating estimation', 'generalized dantzig superiority', 'update', 'performance analysis', 'update strategy', 'pathways', 'update parameter', 'rapidly mixing model', 'performed signal', 'rapidly mixing markov chain', 'sample algorithm', 'convolutional kernel', 'maximum margin criterion', 'maximum margin', 'maximum entropy principle', 'maximum a posteriori estimation', 'maximum a posteriori', 'heavy - tailed distributions', 'covariance matrix', 'to of tailed distributions', 'standard deviation inequalities', 'times series', 'heavy - tailed distribution', 'submodular function effective', 'decomposable algorithms', 'belief propagation', 'gauge transformation', 'partition function', 'mean - by , neural networks', '1 . time series', 'keys', '1 . 2 . 2 , 2 of 2 , 2 the influence function', '1 . 2 . show', 'are 2 . between trees', 'maximization', 'as a', 'as of', 'as a of optimality', 'maximization of', 'as a of random', 'as a of influence', 'as a of a', '1 . 2 . 2 , 2 of 2 of 2 , 2 of 2 , 2 of 2 , 2 of 2 of', 'micrornas', 'deep recurrent neural network', 'identification', 'm - value', 'm - estimation', 'calcium a', 'm - value estimation', 'm - value statistics', 'm - value estimate', 'temporal difference learning', 'state aware learning', 'state aware imitation learning', 'state auxiliary field', 'temporal carlo', 'collaborative - filtering', 'collaboration', 'show <unk> and streaming', 'obtain performance', 'experience', 'difficult - both', 'difficult convergence', 'distribution', 'to', 'using the', 'for distribution', 'pac - bayes', 'stochastic that complexity', 'randomized learning', 'pac - further learning', 'via generalization problem', 'via generalization results', 'pursuit tracking', 'pht', 'pht tracking', 'to tracking', 'rate distortion theory', 'parameters learning', 'regime switching', 'static optimization', 'regime by switching', 'structure learning', 'which', 'decentralized control', 'decentralized hand variance', 'concerns', 'decentralized empirical analysis', 'constraint propagation', 'pedestrian dynamics', 'pedestrian control', 'expected utility', 'or in strategic planning', 'scenes', 'decentralized structure learning', 'or in energy', 'or in military', 'or in strategic alignment', 'conditional independence test', 'non - parametric conditional independence test', 'rademacher complexity', 'conditional independence testing', 'total variational distance', 'non - parametric conditional independence testing', 'rademacher algorithm', 'rademacher analysis', 'text - to - speech', 'how inference', 'of speech', 'of speech synthesis', 'real in speech', 'provably sound', 'provably sound synthesis', 'real in speech synthesis', 'text - agent', 'paper the solution', 'problem based learning', 'to the resulting in speech synthesis', 'to the blind', 'to the resulting in a smdp', ') the resulting in speech synthesis', 'to the resulting in speech', 'to the resulting in the smdp', 'to the resulting in for speech synthesis', 'augmentation', 'relation mapping', 'relation . prove', 'regression models', 'for speech synthesis', 'integrate - and - fire neuron', 'spiral 1', 'periodic problem', 'contribution to noise', 'bayesian dynamics', 'contribution to noise ratio', 'randomly art', 'that 1', 'for <unk>', 'emulation', 'first . alternating current', 'somatosensory cortex', 'somatosensory evoked potentials', 'sub - appropriate generalization', 'that . alternating current', 'hand , off the the', 'better than connection', 'hand , off the cortex', 'hand , off the the imitation', 'variance - based regularization', 'pht a criterion', 'a posteriori estimate', 'variance - al tradeoff', 'we on for in learning', 'on a line optimization', 'paper for learning', 'the maximum entropy method', 't . inference', 't . 2 . 2', 't . 2 . 0', 't . 2 . 2 . 2', 't . 2 . 2 . 1', 't . 2 . 2 . 0', 't . 2 . 2 . 2 . 2', 't . 2 . 2 . 2 . 1', 't . 2 . 2 . 2 . 2 . 2', 't . 2 . 2 . 2 . 2 . 2 . 2', 't . 2 . 2 . 2 . 2 . 2 . 2 . 1', 't . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 1', 't . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 1', 't . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2', 't . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 1 . 2', 'that', 'analytic models', 'lattice networks', 'partial monotonic functions', 'attack', 'analytic modeling', 'parts', 'generative replay', 'deep generative replay', 'cooperative dual model', 'continual learning', 'deep of . a high machines', 'image internal models', 'near 1 state machine', 'deep of . a high dimensionality', 'approximate inference', 'observations', 'sequential monte carlo', 'auxiliary inference observations inference', 'm <unk> estimation', 'm <unk> estimator', 'important 1', 'important empirically strategies', 'generally strategies', 'important counterpart', 'precise introduction', 'processes', 'exploitation', 'improvements', 'learning markov chain monte and make', 'learning markov be into choosing deterministic random', 'learning markov be ) growth', 'introduction time', 'learning markov be of choosing deterministic random', 'learning markov chain rm <unk> can network', 'learning markov be into choosing deterministic model', 'learning markov chain rm <unk> can networks', 'learning markov be into choosing deterministic develop', 'causal discovery', 'regression invariance', 'baseline algorithm', 'multi - distribution', 'influence maximization', 'independent cascade model', 'chosen we expensive', 'online influence maximization', 'stochastic semi we squares', 'online', 'linear feature space', 'methods of probability', 'minimax strategies', 'on omb line learning', 'depth can operations', 'depth can controllers', 'depth algorithm', 'heavy tails', 'depth of interest', 'depth can', 'bounds on are extract', 'heavy but tails', 'is run', 'is neural networks', 'is run and time', 'is run time', 'model moreover', 'model mismatch', 'expert systems', 'the approximate weighting', 'in in model long distributions', 'the . <unk>', 'of optimality', 'robust', 'in in model key distribution', 'attention', 'c and 1 learning', ', 11', 'stochastic existing', 'hierarchical recurrent network', 'pedestrian . 0', 'spatial attention', 'pedestrian tracking', 'over learning', 'kth activity recommendation', '3 on', 'policy making', 'experimental validation', 'reduction of architectures', 'policy gradient', 'class - agnostic tracking', 'step - of on language model', 'limit cycle', 'lower limb', 'step - of on language', 'important multiplicative model', 'gives . for rat', 'k introduce temporal correlation function', 'k introduce temporal feature', 'generative model', 'k introduce temporal difference', 'as noise', 'as designed networks', 'as designed experiments', 'fly vision', 'estimate geometry', 'fly vision system', 'fly vision points', 'maximum - likelihood estimation', 'markov processes', 'markov chain monte carlo methods', 'input attention', 'equations', 'variational law', 'equations of distributions', 'equations of gaussian processes', 'equations of state distributions', 'equations of large numbers', 'equations of large scale systems', 'large . large scale networks', '( <unk> law', '( <unk> limit theorem', '( <unk> <unk> on . times', '( <unk> limit law', 'large scale networks', '( <unk> <unk> on . consider', '( <unk> <unk> on . <unk>', '( <unk> <unk> on . 5', 'spatial organization', 'discrete - charged networks', 'discrete event simulation', 'markov chain', 'markov chains', 'attributed graphs', 'poisson processes', 'mirna', 'canonical correlation analysis', 'markov chain monte carlo simulations', 'joint distribution', 'stochastic we switched networks', 'markov', 'stochastic we', 'joint', 'error - descent learning', 'parallel stochastic algorithm', 'by policies', 'error in learn', 'applications of markov chains', 'applications of alleviates and it', 'few - network optimality', 'single - layer network', 'few - network principle', 'exploitation and exploration', 'applications of alleviates and it of accuracy', 'exploitation and prove', 'gauging', 'exploitation and exploration of networks', 'gives histories', 'gives method', 'when d2 networks', 'gives are', 'gives of only', 'multi - armed bandit', 'multi - armed bandits', \"bernstein ' s theorem\", 'success factors', 'reduce', 'dependence', 'conditions', 'for i - optimality', 'aforementioned trees', 'for condition estimation', 'gradient estimates', 'for gradient methods', 'for gradient', 'gradient', 'for gradient estimates', 'hierarchical clustering', 'bisecting k - means', 'dna', 'many we', 'local process', 'local search', 'bisecting k we means', 'we action to change', 'we action', 'we > 2 learning', 'we > <unk>', 'we > <unk> <unk>', 'for most secondary', 'to additional information', 'tree additional information', 'to additional cost', 'tree additional cost', 'restarting', 'adaptive regularization', 'accelerated gradient method', 'quadratic growth condition', 'proximal gradient', 'linear convergence', 'h\\\\\"{o}lderian error bound', 'essentially smooth piecewise quadratic loss', 'essentially linear convergence', 'adaptive gradient converging', 'adaptive gradient converging method', 'convergence speed', 'variational gradient descent', 'stein learn', 'stein variational gradient descent', 'fokker - planck equation', 'point process', 'stochastic approximation', 'and support function spaces', 'gradient accuracy', 'we factors', '<unk> for progress', 'partial hard thresholding', 'support', 'support recovery', 'hard thresholding', 'orthogonal matching pursuit', 'orthogonal matching pursuit pursuit', 'partial hard for for features', 'partial hard for for on tasks', 'partial hard for for space tasks', 'level set method', 'level set methods', 'under applications', 'model selection and recovery', 'deep q - network', 'least squares', 'applications to machine learning', 'gradient boosting decision tree', 'information gain', 'feature bundling', 'one - side sampling', 'split point', 'least mean square', 'np - hard', 'greedy algorithm', 'split point determination', 'adversarial ranking', 'language generation', 'policy efficacy', 'generative adversarial network', 'ranking', 'high - modeling', 'language depth', 'easy . s . discriminative learning', 'easy . s . 1 , online learning', 'in of optimality', 'exploration . exploitation', 'on . <unk> risk', 'trials with in uncertainty', 'in of stochastic processes', 'trials the portion', 'is risk minimization', 'is risk management', 'hidden unit space', 'combinatorial topology', 'cellular automata', 'multi - layer networks', 'multi - layer network', 'classifier', 'language equations', 'combinatorial', 'combinatorial transition', 'performance guarantee', 'net - convex pruning', 'overfitting', 'model redundancy', 'net - trim algorithm', 'graph matching', 'multiplicative update', 'doubly stochastic previous concerns', 'reward maximization', 'for on 1 spatial output', 'for on 1 spatial an descriptions', 'as mdps', 'for on 1 spatial form', 'anytime algorithms', 'dynamic importance sampling', 'partition in graphs', 'partition in . mode', 'to of optimality', 'to of a a tree', 'to of a a a residual queries', 'a of a mechanisms', 'to of a a a tree', 'mechanisms with constraints', 'this for for optimality', 'this for for for optimality', 'to of a a a residual london', 'in , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2', 'in , 1 , 2 , 1 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2', 'in , 1 , 1 , 1 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2', 'in , 1 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'in , 1 , 1 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', 'in , 1 , 2 , 1 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'in , 1 , 1 , 1 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2', 'in , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 2 , 2 , 1 , 2', 'in , 1 , 1 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'in , 1 , 1 , 1 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'policy intractable', 'bellman residual', 'policy optimization', 'in entailment', 'ridge regression', 'random features', 'random feature', 'generalization bounds', 'distributed data', 'bayesian learning', 'qp movements', 'domain - specific transformations', 'domain - specific transformation', 'multi - modal data', 'point process model', 'wasserstein distance', 'nuisance process', 'have of <unk> at . replacement', 'multi - modal', 'have of 1 , 2 . 2 , 2 , 2 the em algorithm', 'have of 1 , 2 . 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 3 of d , 3 of d ,', 'have of 1 , 2 . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'have of 1 , 2 . 2 , 1 , 2 , 2 , 2 , 2 , 2 , 3 of d , 3 of d , 3 the', 'have of 1 , 2 . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 3 of d , 3 of d ,', 'have of 1 , 2 . 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'have of 1 , 2 . 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 3 of d , 2 of d ,', 'have of 1 , 2 . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 3 of d , 3 of d , 3 the', 'have of 1 , 2 . 2 , 1 , 2 , 2 , 2 , 2 , 2 , 3 of d , 3 of d , 2 ,', 'have of 1 , 2 . 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 3 of d , 2 , 2 ,', 'have of 1 , 2 . 2 , 1 , 2 , 2 , 2 , 2 , 2 , 3 of d , 3 of d , 2 the', 'ensemble sampling', 'decision problems', 'recurrent highway networks', 'hypernetworks', 'model - based introduce', ') a', 'we learning', 'in the ensuring', 'feedforward nets', 'approximation power', 'comparison of activation functions', '1 <unk> functions', 'power of and new of learning', 'power of and new and of algorithm', '1 <unk> descent property', 'adaptive svrg methods', 'error bound conditions', '0 ace', \"' analysis of algorithms\", 'linkage learning', 'interconnection networks', \"' s popular method\", 'svrg network', \"' s this a posteriori error bound\", 'based machine learning', 'linear and nonlinear optimization', 'linear and x performance tests', 'bayesian methods', 'bayesian compression', 'hierarchical priors', 'hierarchical prior', 'perspective', 'establish sampling', 'statistical relational learning', 'pseudo - practice', 'pseudo - point approximation', 'sequential sampling', 'in , off . <unk> data', 'in , off . line learning', 'in , off . <unk> data analysis', 'tracking', 'in , off . 2 , b', 'in , off . 2 , 2 . 2', 'expert advice', 'in , off . 2 , 2 . 2 <unk> data', 'in , off . 2 , 2 . 2 <unk> information', 'in , off . 2 , 2 . 2 . 2', 'in , off . 2 , 2 . 2 <unk> <unk>', 'in , off . 2 , 2 . 2 <unk>', 'in , off . 2 , 2 . 2 . 2 <unk>', 'in , off . 2 , 2 . 2 . 2 <unk> data', 'in , off . 2 , 2 . 2 . 2 <unk> information', 'in , off . 2 , 2 . 2 . 2 <unk> <unk>', 'in , off . 2 , 2 , 2 , 2 , 2 , 2 . 2 <unk> <unk>', 'in , off . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2', 'in , off . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2', 'in , off . 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2', 'structure from motion', 'implicit surfaces', 'variational', 'implicit variational', 'variational information', 'function representation', 'sparse representations', 'however <unk>', 'however geometry', 'forms processing', 'k - means clustering', 'random projection', 'singular value experiments', 'sparse neural networks', 'rely of detail', 'rely of reasoning', 'rely of learning', 'd2 - ctw', 'dynamic bounds', 'stochastic context v context activity', 'd2 - ctw signal analysis', 'finite state markov chain monte carlo', 'stochastic context mdps', 'stochastic context transformation', 'stochastic context trees', 'rapid serial source coding', 'part in state space', 'processes on time series', 'possibility and entropy', 'part in context trees', 'stochastic context feedback', 'processes on state space', 'stochastic results', 'stochastic learned models', 'stochastic learned', 'learned models', 'learned learning', 'stochastic context ( sparse coding', 'stochastic context ( sparse data', 'stochastic context ( sparse', 'stochastic context ( a', 'sparse and structured neural attention', 'sparse and and and matrix factorization', 'sparse and structured attention', 'structured penalties', 'sparse and and and matrix learning', '. dictionary learning', 'user dynamic programming', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 ,', '. 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 ,', '. 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 2 ,', 'polynomial networks', 'factorization machines', 'multi - in ,', 'multi - output classification', 'main can work', 'invariant subspace method', 'multi - in , and mf , <unk>', 'multi - in , and mf , <unk> , <unk>', 'multi - in , and mf , <unk> , <unk> , <unk>', 'multi - in , and mf , <unk> , <unk> , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , <unk> , <unk> , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , <unk> , <unk> , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , <unk> , <unk> , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , <unk> , <unk> , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', 'multi - in , and mf , <unk> , <unk> , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1', 'bioinformatics', 'dna databases', 'dna data retrieval', 'sequential prediction', 'distributed data mining', 'distributed learning', 'time features', 'update rule', 'approaches', '<s> amounts', 'approaches to become', '<s> methods', 'approaches to rank', 'practice of or', 'procedure', '<s> extra (', 'allowing errors', 'evidence theory', 'lower', 'allowing constraints', 'competitive analysis', 'grasping', 'shape recognition', 'somatosensory information', 'somatosensory system', 'random measurement matrices', ') smdp', 'asymptotic property', 'show smdp', 'asymptotic performance', 'asymptotic we', 'higher order statistics', 'power of convergence', 'almost uniform distribution', 'power', '1 data', 'power of random variables', 'power of large numbers', 'power of a random if', 'power of a random variables', 'power of a family of random first', 'anomaly detection', 'permutation invariance', 'permutation invariant functions', 'as model construction', 'of model data', 'as model', 'approximate string matching', 'approximate pattern matching', 'starting points', 'pipeline mining', 'multi <unk> class classification', 'multi <unk> criteria classification', 'multi <unk> criteria decision analysis', 'multi <unk> criteria decision making', 'notably <unk>', 'non <unk> parametric estimation', 'aims', 'climate data', 'bounding box prediction', 'tensors', 'fresh an sorting', 'give', 'climate change detection', 'perturbations', 'based clustering', 'electrodes', 'predictions', 'bayesian optimisation', 'heat treatment', 'batch bayesian optimisation', 'process - constrained optimisation', 'heat treatment process', 'large - scale optimisation', 'heat treatment process constraints', 'attacks', 'multi - task gaussian process', 'individualized treatment effect', 'are . ordinary least squares', 'guess complete', 'guess diffusely propose', 'data assimilation', 'model selection bias', 'assemblies', '1 and 2 , 2 , 2 , 2 , 2 , 2', '1 and 2 , 2 , 2 , 2 , 1 , 2', '1 and 2 , 2 , 2 , 2 , 2 , 1', '1 and 2 , 2 , 2 , 1 , 2 , 2', '1 and 2 , 2 , 1 , 2 , 2 , 2', '1 and 2 , 2 , 1 , 2 , 2 , 1', '1 and 2 , 1 , 2 , 2 , 2 , 2', '1 and 2 , 2 , 2 , 1 , 2 , 1', 'based a posteriori credible fisher discriminant', '1 and 2 , 2 , 1 , 2 , 1 , 2', '1 and 2 , 1 , 2 , 2 , 2 , 1', 'based a posteriori credible fisher discriminant analysis', 'activity recognition', 'spherical convolutions', 'image analysis', 'cubed - sphere', 'image representations', 'important events', 'molecular modelling', 'important extreme points', 'important extreme values', 'by the topology', 'pool a sphere', 'clean side regions', 'clean side a has', 'we - 1 fields', '1 a and 2 side regions', '1 a and 2 side fields', 'high - dimensional geometry', 'cooperative', '1 and 2 side regions', 'computational chemistry', '1 and 2 side a has', '1 a and 2 side a has', 'on the sphere', 'on the optimization', 'computational topology', 'linear dynamical systems', 'projected gradient descent', 'sparse constant', 'cross validation', 'parametric learning', 'leave - their - 1 criterion', 'leave - one - out cross module', 'dataset deep learning', 'leave - one - 0 ) algorithms', 'molecular diversity', 'leave - one - 0 - algorithm', 'molecular recognition', 'entire function', 'low - rank tensor regression', 'least squares regression', 'tucker neurons', 'on behavior', 'more prediction', 'more prediction error bound', 'least squares we', 'least curves', 'probabilistic sentential decision diagram', 'permutations and provides method', 'permutations and provides provides network', 'learning distributions', 'on <unk> line algorithms', 'noisy data', 'route', 'have <unk> clean assignment', 'self - organizing networks', 'growing cell algorithm', 'tt x data', 'laws and networks', 'model - based bayesian inference', 'neural connectivity', 'spiking activity', 'spike inference', 'all - optical circuit mapping', 'spike train analysis', 'latent structure discovery', 'poisson gaussian process', 'system identification', 'individual receptive field locations', 'individual receptive factual risk representations', 'interventional learning', 'security', 'data poisoning attacks', 'malware', 'intrusion detection', 'data poisoning attack', 'imdb attacks', 'outlier removal', 'attack resilience', 'attack detection', 'sufficient statistics', 'perceptual sensitivity', 'fisher information', 'hierarchical representations', 'fisher information engineering', 'fisher information matrix', 'non convolutions', 'hierarchical', 'local', 'ensuring find', 'finite sums', 'acceleration schemes', 'finite sum optimization', 'variance - results', 'acceleration methods', 'sequence classification', 'sequential output statistics', 'primal - computed tomography', 'noisy environment', 'influence fusion', 'approximation from of is', 'noisy monotone function', 'of multi are function', 'of multi are noise', 'of multi are level', 'of multi are optimization', 'primary visual function', 'and offer optimization', 'demonstrate constraint', 'demonstrate function', 'paper that constraint', 'and as noise', 'and as information', 'telemetry', 'data collection', 'counter data', 'ldp', 'repeated data collection', 'locally differentially private', 'population statistics', 'concrete dropout', 'uncertainty estimation', 'and fascinating', 'logic', 'class', 'posit', 'connection', 'audio', 'policy gradients', 'batch size size', 'batch size', 'policy feature', 'temporal reasoning', 'missing for guarantee', 'kalman variational auto - encoder', 'unsupervised in batch means', 'latent variable detection', 'latent variable for continuous constructed collected data', 'dynamic time warping', 'with supervised learning', 'unsupervised in 2 to d tv', 'unsupervised in 2 to d visual cortex', 'unsupervised in 2 to d longer ngram', 'theory of action recognition', 'unsupervised on to learning', 'stochastic time warping', 'stochastic mixture models', 'markov chain monte carlo markov', 'approximate sufficient statistics', 'hierarchical bayesian inference', 'perceptual learning', 'non on on retrieves', 'field on prediction', 'non on on on retrieves', 'such on random variables', 'bayesian gan', 'idea generation', 'result selection', 'variant selection', 'result fusion', 'paper information', 'hamiltonian monte carlo', 'parameter learning', 'use cases', 'show and accounts', 'exploiting are <s> personalized medicine', 'exploiting are <s> personalized detection', 'exploiting are <s> including accuracy', 'exploiting are <s> personalized prediction', 'empirical studies', 'scenario', 'conservative estimation', 'improve performance', 'multi - agent reinforcement learning', 'non - cooperative game theory', 'in , commons spike local realism', 'z . rt', 'common - pool resource appropriation', 'problem solving', 'neural network models', 'high . 0', 'z . 0', 'non - convex optimization', 'gradient ascent', 'kac - rice formula', 'gradient ascent algorithm', 'respectively', '( respectively', 'attention model', 'high - order correlation', 'world wide web', 'representations', 'crfs', 'for matrix factorization', 'for matrix', 'cell assemblies', 'sparse convolutional coding', 'integrate - and - fire neuron cultures', 'integrate - and - fire neural cultures', 'temporal motif', 'temporal figure - temporal motif', 'temporal figure - temporal pattern', 'integrate - and - fire neuron motif', 'integrate - and - fire information processing', 'sensory information', 'information - theoretic intersection', 'information - processes', 'neural is', 'behavioral performance', 'issue product', 'i product', 'many product approximation', 'in approaches', 'in situ dimensions', 'rational parametrizations', 'rational parametrization', 'dynamical systems', 'realization theory', 'feedforward analysis', 'geometric matrix completion', 'multi - vector quantization', 'multi - vector model', 'term memory', 'and theoretical learning', 'bound in graphs', 'multi - vector neural as , theoretical learning', 'multi - graph the benchmarks', 'bound in attacks', 'monte carlo variational inference', 'gradient variance reduction', 'noisy gradient variance', 'bayesian founding to models', 'approximate bayesian learning', '1 convolutional neural networks', 'bayesian founding to datasets', 'data fitting', 'data on data', 'in on on problem', 'by on on problem', 'by on data', 'in on on as experiments', 'in on on as inequality', 'in on on as inequality constraints', 'in on on . multi question', 'in on on as a attention', 'attention memory', 'dynamic parameter prediction', 'associative attention', 'visual dialog', 'visual reference resolution', 'optimal transportation', 'unsupervised domain adaptation', 'loss function', 'natural resources', 'specifically', 'into deployed images', 'into deployed', 'into deployed test', \"bayes ' theorem\", 'kernel approximation', 'multiresolution kernel approximation', 'the more etc', 'more recently framework', 'more difficulty', 'the more scale', 'more feature selection', 'more determined scale', 'markov jump processes', 'uniformization', 'variational bayes', 'collapsed variational bayes', 'mondrian forest', 'universal data structure', 'universal data structures', 'universal data', 'mondrian forests', 'to aspects', 'provide representations', 'linear scale', 'to model counting', 'to model interference', 'to model distance', 'auction design', 'welfare analysis', 'economics', 'intersection', 'completion time', 'p data', 'decompositions', '1 of optimality', 'our optimization', 'our 1 of', 'our 1 of optimality', 'over - regularization', 'nearest neighbor search', 'preconditioning', 'end - to - end differentiable proving', 'backward chaining', 'mechanisms with making', 'mechanisms with making we', 'mechanisms with making under uncertainty', 'finite state machines', 'finite state automata', 'analog cochlear model', 'two to phase flow', 'cochlear model', 'on we spectrum', 'practical subset', 'practical in', 'cochlear signal processing', 'two to phase vqa', 'two to coupled vqa', 'viral marketing', 'social network', 'approximation algorithm', 'human <unk> computer interaction', 'generative learning', 'human <unk> robot interaction', 'learning by demonstration', 'visual , human interaction', 'human <unk> computer interactions', 'human <unk> computer interface', 'visual attention', 'scanpath', 'variational laws', 'statistical mechanics', 'dynamic scenes', 'variational scale', 'factors', 'essentially non probabilities', 'statistical pattern recognition', 'efficiently computable model', 'econometrics', 'least ( ( that . instantiation', 'least action principle', 'on they line learning', 'guarantees <s> and', 'guarantees <s> logic', 'optimisation', 'guarantees <s> chaining', 'guarantees <s> and basic reproduction', 'guarantees <s> and basic al', 'recursive sampling', 'nystrom approximation', 'sparse approximation', 'nystrom method', 'nystr?m approximation', 'deep reinforcement learning', 'off - policy model - free learning', 'of the image events', ', <unk> and , , , and and and and and and and methods', ', <unk> and , , , <unk> and and and and and and methods', ', <unk> and , , , and and and and and and and methods , <unk>', ', <unk> and methods', ', <unk> and <unk>', ', <unk> , <unk> , <unk>', ', <unk> and , , , and and and and and and and methods , <unk> , <unk> , <unk> , <unk> ,', ', <unk> and', ', <unk> ,', 'it , performance', 'an', 'orientation', 'dataset', 'generally grazing', 'resolve', 'given domain', 'discovery', 'encoder', 'first memory model', 'covering', 'roc', '<s>', 'crowd counting', 'convolution', 'side information', 'filter weights', 'image deblurring', 'topic modeling', 'scan - and - beyond k', 'conic scan - and k', 'either - applications', 'em algorithm', 'algorithm design and analysis', 'bayesian network', 'subsampling', 'stochastic subsampling', 'large latent variables', 'uncertainties', 'is <unk> <unk>', 'vision', 'designated semantics', 'semi - supervised image classification', 'structured generative adversarial networks', 'semi - supervised image computed tomography', 'structured generative adversarial we', 'collaborative games', 'contextual linear bandits', 'conservative bandits', 'safety', 'contextual linear bandit', 'personalized recommendation', 'safe recommendation', 'to in validation', 'variational memory addressing', 'stochastic addressing', 'conditional mixture models', 'tensor train', 'rank minimization', 'statistical guarantees', 'masks', 'tensor train rank minimization', 'statistical efficiency', 'of linear equations', 'obtain optimization', 'higher - order tensors', 'tensor train projects', 'question answering', 'probabilities', 'notion', 'higher - o linear programming', 'computer and machine learning', 'notion of the', 'better than used', 'incomplete beta function', 'incomplete cholesky factorization', 'determined growth condition', 'establish regression', 'establish values', 'spectral mixture density', 'levy processes', 'levy process', 'counter algorithm', 'mean squared error', 'mean field methods', 'mean field approximation', 'from neural networks', 'from neural computation', 'most powerful test', 'utilize resource constraints', 'most powerful tests', 'hyperspherical learning', 'visual representation learning', 'hyperspherical convolution', 'generalizability theory', 'of addressing', 'it 1', 'experimental mathematics', 'is 1 . 2', 'and modelling', 'is 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 ,', 'is 1 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'is 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 semantics', 'attention - gated fascinating', 'is , and idea integration', 'is , <unk>', 'safe', 'is . <unk>', 'integration', 'the <unk>', 'the em algorithm', 'is , fusion', 'a <unk>', 'spherical', 'in a posteriori probability estimation', 'spherical , <unk>', 'resource constraints', 'resource estimation', 'agent learning', 'spike initiation protocol', 'in . 3', 'in . 3 . d .', 'in . 3 . 2 .', 'then rules', 'spike', 'in . 3 . 0', 'in . 3 . 3', 'in . 3 . d . 2 . 3', 'in . 3 . d . 3 . 2 .', 'in . 3 . d', 'in . 3 . 2', 'dynamic computation graph', 'batching', 'operation batching', 'dynamic computation graphs', 'dynamic neural networks', 'on - the - fly computation', 'incremental computation', 'readout', 'for on we are itself', 'we we that the architecture', 'is submodular functions', 'to is 2 . 1', 'to from benchmark', 'to 1', 'to from', 'a', 'to from 2', '2 . 2 . 1', 'to from <unk>', 'to . 1', 'to counting', 'extrapolation methods', 'stochastic objects', 'convergence rates', 'allowing known error', 'allowing known answering', 'variant', 'solution algorithm', 'by <unk> optimization', 'better types', 'weather', 'optical flow', 'in <unk> the <unk> our algorithms', 'in <unk> the <unk> parameters', 'decades', 'solution bound', 'by nodes', 'convolutional networks', 'in <unk> the <unk> <s>', 'data pre algorithms', 'non - discrimination', 'discrimination prevention', 'automatically defined function', 'data pre - processing', 'analytic hierarchy process', 'auto of discriminant functions', '1 in syndrome', '1 in nn', '1 in in syndrome', 'spike sorting', '?triage - then - pursuit? waveforms', '?coreset? - then - pursuit? waveforms', 'people counting', '?triage - then - pursuit? cluster', 'spike waveform templates', 'neural increasingly', 'neural p the cnn', 'spike significance', 'match scores', 'not . information', 'not . <s> a neural motifs', 'not . information processing', 'not . information rates', 'embedded signal processing', 'gives a standard', 'robot control', 'robotics', 'inverse dynamics', 'practice surprising', '\"trajectory extension', 'slate', 'practice strategy', ') robotics', 'inverse algorithm', 'logged', 'independence', 'independence clustering', 'stationary time series', 'matrix', 'bayesian model inversion', 'variational transfer', 'variational em algorithm', 'fluorescence microscopy', 'model inversion', 'critical biology', 'propose signal processing', 'propose signal detection', 'propose and direction estimation', 'warped be', ') problem', ') model image analysis', ') model image both', ') model image residual analysis', 'outputs statistics', 'multi actions detection', 'multi actions', 'multi actions recognition', 'limited information', 'sequential hypothesis testing', 'bayes rule', 'black - box classifier', 'cardinality decoding', 'black - box classifiers', 'weak submodularity: interpreting neural networks', 'black - box art', 'comparable submodularity: landmarks', 'based on the fascicle algorithm', 'based on the fascicle', 'that . our', '1 ( 2 , 2 the conventional advantages', 'cardinality', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 1 , 2 , 2', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 1 , 2 , 2 , 1', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 1 , 2 , 2 , 2', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '1 ( 2 , 2 , 2 , 1 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 1', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 1 , 2 , 2 , 2', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2', '1 ( 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 1 , 2 , 1', 'transfer', 'sarsa', 'successor features', 'q ideas: learning', 'dynamic programming?s policy improvement', 'dynamic programming?s policy general', 'dynamic policy improvement', 'marl', 'dynamic programming?s policy general experience', 'counterfactual fairness', 'gender', 'law solve', 'causal inference', 'law school', 'law', 'decision support systems', 'sexual orientation', 'zero - shot learning', 'few - shot classification', 'zero of data', 'metric propose learning', 'image generation', 'adversarial nets', 'sublinear - regret', 'sublinear - we', 'few , large datasets', 'learning with joint constraints', 'estimating equations', 'to . e the commerce', 'a . 2 . 2', ') to polynomial splines', '( . 2', 'and to . e . commerce', 'and ( 2 and 2', '( . 2 . 2', '( . 2 . 0', 'and ( 2 and 2 , 2 . 2', 'and ( 2 and 2 , 2', 'and ( 2 and 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'and ( 2 and 2 , 2 . 2 <unk> <unk>', 'and ( 2 and 2 , 2 . 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'and ( 2 and 2 , 2 , 2 , 2 of optimality', 'and ( 2 and 2 , 2 . 2 <unk> <unk> <unk>', 'and ( 2 and 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 of', 'and ( 2 and 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 ,', 'and ( 2 and 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 ,', 'and ( 2 and 2 , 2 , 2 , 2 , 2 , 2 , 3 , 2 , 2 ,', 'and ( 2 and 2 , 2 . 2 , 2 , 2 , 2 , 2 , 2 , 2 of', 'and ( 2 and 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 ,', 'and ( 2 and 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 of optimality', 'and ( 2 and 2 , 2 , 2 , 2 , 2 , 3 , 2 , 2 , 2 ,', 'functional connectivity', 'functional data analysis', 'optical neural computing', 'perceptron - iike learning', 'is learning', 'are , adaptivity', 'adaptivity', 'automatic compute', 'threshold logic function', 'traveling salesman problem', 'np - hardness', 'hard problems', 'np is np', 'np - completeness np np np np np', 'np - completeness np np np np', 'np - completeness np np np np np np', 'np is np np', 'np is np np np', 'np - completeness np np np np np np np', 'np is np np np np', 'np - completeness np np np np np np np np', 'np is np np np np np', 'np - completeness np np np np np np np np np', 'size of lower bounds', 'size of lower solutions', 'usual of sat', 'there - world network', 'via for problem', 'via for all solutions', 'via for points', 'points of solutions', 'jackknife estimator', 'in learning parameters', 'a based the empirical method', 'in learning algorithm', 'discriminator', 'metric movement costs', 'minimax regret', 'non - stochastic multi - armed bandits', 'co - clustering', 'bipartite graph', 'graph clustering', 'ordinal embedding', 'upper and lower bounds', 'low - dimensional metrics', 'upper and high .', 'ordinal case', 'likelihood ratio', 'upper and lower bound', 'functions of optimality', 'kullback , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 1 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 1 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 1 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 1 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 1 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 1 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'kullback , 1 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'local optimization', 'adaptive gradient methods', 'marginal value', 'localization', 'adaptive gradient descent', 'robotic learning', 'the reduce', 'error results', 'multi - class classification', 'binary reduction', 'multidimensional data', 'using problem complexity', 'empirical occurring that exchange rate', 'normalized cuts', 'empirical occurring that exchange', 'new problem complexity', 'double sampling', 'text to model', 'as to multi . from data', 'text classification', 'text summarization', 'isotonic regression', 'random permutation models', 'leave by one must evaluation', 'online isotonic regression', 'learn to rank', 'random permutation model', 'multiagent reinforcement learning', 'multiagent systems', 'coordination games', 'inverse filtering', 'sleep segmentation', 'structured output networks', 'non . video processing', 'non . video coding', 'non - parametric structured output networks', 'will', 'potential fields', 'on of gaussian models', 'on of gaussian maps', 'price decomposition', 'price . 2', 'restricted 1 of 2', '2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2', '2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '2 , 2 , 1 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2', '2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2', '2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1', '2 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1', 'time warping', 'trajectory classification', 'time warping invariant neural networks', 'trajectory nets', 'grammatical reference', 'time address prediction', 'dynamic semi address space', 'query selection', 'data - driven learning', 'out sparse learning', 'data - bounds', 'out sparse sampling', 'out sparse data', 'vae learning', 'stein non to gaussian mixture model', '( agents', 'exponential convergence', 'workers density estimation', 'address prediction', 'exponential random field', \"well for belonged ' s and hidden markov chain monte carlo\", \"stein ' s networks\", \"well for belonged ' s and hidden markov model\", 'exponential random fields', \"well for belonged ' s and hidden markov random fields\", 'brain activity', 'psychology', 'brain activations', 'mri', 'brain development', 'cuda', 'linear learning', 'heterogeneous computing', 'gpgpu', 'algorithmic concise models', 'reconstructing for fpgas', 'opencl', 'linear learning to a posteriori', 'linear learning to a posteriori estimation', 'tractography', 'amazon ec2', 'deep multi - stage adversarial network', 'our of often networks', 'or in sports', 'our of often milliseconds time', 'polynomial although for forecasting', 'our of often conditional binary network', 'difficult the lower bound', 'our of often conditional binary model', 'difficult on of as a is', 'from motion trajectory', 'sobolev training', 'agent playing', 'generalisation', 'multi - information source optimization', 'multidimensional signal processing', 'general sampling', '2 . 0', 'is sampling', 'multi - estimating functions', 'based optimization', 'explain sampling', 'surrogate models', 'sequential monte carlo methods', 'parallelism', 'based spatial multiplexing', 'robot locomotion', 'population - based learning', 'q to learning', 'relation learning', 'provides games', 'kernel ridge regression', 'with complexity', 'kernel powerful test', 'is adam analysis', 'kernel powerful tests', 'kernel svms', 'networks of supervised learning', 'kernel ridge postulate', 'including of policy', 'multiagent planning', 'decentralized mdps', 'value function approximation', 'decentralized planning', 'actor - critic', 'in - <unk>', 'in - vehicle', 'gender classification net', 'phonetic probability', 'speaker independent', 'speaker independent number', 'continuous speech recognition', 'speaker inverse gaussian mixture model', 'indicators', 'gender classification integer learning', 'games with side information', 'games with side constraints', 'the maximum likelihood classifier', 'the maximum likelihood estimation', 'marginal log - likelihood', 'log - likelihood', 'log - likelihood ratio', 'markov random field', 'marginal learning', 'adversarial real of codes', 'adversarial real of memory', 'adversarial real of linear regression', 'adversarial real of linear regression trees', 'magnetic resonance imaging', 'diffusion - weighted mri', 'samples autoencoders', 'diffusion autoencoders', 'crowdsourcing', 'minimax', 'triangular estimation', 'minimax estimation', 'noisy labels', 'probabilistic logic', 'accuracy estimation', 'unlabeled data', 'multiple classification', 'multiple classifier systems', 'prediction markets', 'cost - function - based prediction markets', 'sampling error', 'exponential beliefs', 'market - maker bias kt costs', 'market - maker bias kt kt', 'market - maker bias kt with bound', 'market - maker bias kt kt with bound', 'market - maker bias', 'exponential approach', 'exponential armed bandit', 'coordinate - descent', 'stochastic gradient projections', 'nonequilibrium processes', 'variational learning', 'detailed balance', 'transition operator', 'transition operators', 'stochastic transition operator', 'matrix multiplication', 'high - dimensional codes', 'optimal learning', 'polynomial codes', 'networks of linear equations', 'networks of graphs', 'recovery raw materials', 'optimal among data optimization', 'optimal among enables y from propose', 'disentangled representations', 'video analysis', 'disentangled representation', 'temporal coherence', 'temporal which filter', 'as for of text', 'temporal which filter bank', 'in stimuli recognition', 'we stimuli encoding', 'for human vision', 'in stimuli generation', 'federated learning', 'systems framework', 'fault', 'distributed multi - active learning', 'distributed multi - task learning', 'to , transfer learning', 'on in memory', 'on in line learning', 'on in network sampling', 'on in network learning', 'on in network computing', 'to , intra in networked art', 'can bus', 'at . all', 'at . 3', 'all of instead', 'low - rank approximation', 'relative error k - rank approximation', 'radial basis functions', 'polynomial kernels', 'radial basis function kernel', 'input sparsity', 'non - parametric density estimation', 'conditional exponential corrupted', 'nonparametric graphical models', 'non - estimation', 'conditional exponential distribution', 'geometric self - descent', 'riemannian metric', 'to regression', 'dual path network', 'residual network', 'dual path model', 'a the training results', 'training results', 'a the heart model', 'a the paper model', 'stochastic convex optimization', ') 1m', 'stochastic neural networks', 'truncated gaussian exponential distribution', 'doubly truncated to algorithm', 'restricted boltzmann many <unk> body problem', 'restricted boltzmann machine', 'dgps <unk> t distribution', 'individual nonlinearity', 'gd <unk> optimality', 'on <unk> line nonlinearity', '2 <unk> d nonlinearity', 'about nonlinearity', 'gd <unk> nonlinearity', 'dialogues', 'about resistance', 'gd <unk> expansion', '?distilled?', 'joint training', '?distilled? policy', 'speed control', 'any <unk> fit', 'multi - commodity auctions', 'bidding strategies', 'bidding algorithms', 'density ratio estimation', 'subgradient descent', 'in metrics', 'slam', 'navigation', 'self - localization', 'simultaneous multi co data', 'distribution function', 'vestibulo - ocular reflex', 'locomotion', 'adaptive learning', 'm algorithms', 'eye movement', 'low memory', 'eye', 'efficient learning', 'efficient . <unk>', 'empirical potentials', 'idea selection', 'motor learning', 'motor models', 'visual interaction networks', 'physics simulator', 'goal recognition', 'model - based decision making', 'it ) modeling', 'reconstruct & crush network', 'energy', 'energy -', 'energy - based learning', 'smooth network', 'motion computations', 'energy - based model', 'o ( 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 1 ,', 'o ( 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 1 , 2 ,', 'o ( 1 , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2 ,', 'o ( 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2 ,', 'o ( 1 , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 1 ,', 'o ( 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 1 , 2 ,', 'o ( 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 1 ,', 'o ( 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 2 ,', 'o ( 1 , 1 , 1 , 2 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 ,', 'o ( 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 2 , 1 , 2 , 1 , 2 ,', 'streaming algorithms', 'solutions', 'to selection', 'solutions of algorithms', 'min selection', 'to selection generalization', 'random projections', 'blind source separation', 'tailed distributions', 'principle component analysis', 'sample selection', 'maximum interact', 'mixtures of experts', 'mixtures of distributions', 'mixtures of gaussian distributions', 'map estimation', 'distributed denoising', 'it denoising', 'in measurement vectors', 'diverse density estimation', 'average correlation', 'potential correlation', 'hypercontractivity coefficient', 'binary hypothesis testing', 'correlation', 'correlation measure', 'is <unk> test', 'black box test', 'gene interaction', 'gene interactions', 'potential after test', '. 1 test', 'gene convex hull', '. <unk> test', 'double correlation', 'gene multiplication', 'gene test', 'deep gaussian process', 'among in situ hybridization', 'variational estimation', 'nonparametric estimation', '<s> constrained optimization', 'joint model', 'continuous ranking', 'oriented recursive partitions', 'empirical iroc curve optimization', 'we by deep learning', 'empirical warping', 'on gan learning', 'belief networks', 'factorized asymptotic bayesian', 'belief network', 'eeg', 'local field potential', 'synchrony', 'electroencephalography', 'it neural networks', 'our neural protocol', 'each edges', 'any neural network', 'any neural networks', 'simpler', 'mlp', '<s> independent component analysis', 'hand performance', 'path planning', 'autonomous robots', 'path optimization', 'robot navigation', 'generalized', 'as of a service', 'as of a service robots', 'superior colliculus', 'saccades', 'saccadic movement', 'saccade', 'oculomotor system', 'oculomotor graph', 'collicular embedding', 'hyperparameters', 'non - stationary kernel', 'non . learning', 'propose and experiments', 'model speech', 'embedding', 'q . 2 . 2', 'q . 2 . 2 . 2', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 2 , 2 , 2 , 2 , 2 , 2 , d , 2 , 2 , 2 ,', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , d , 2 , 2 ,', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 ,', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 ,', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d , 2 ,', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d ,', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 ,', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d', ', 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d .', 'catastrophic forgetting', 'incremental moment matching', 'radial basis function neural networks', 'radial basis function network', 'radial basis function neural network', 'l2 - kernels', 'radial basis function networks', 'information exposure', 'algorithmic personalization', 'as represent paths', 'social homophily', 'able represent', 'as topic', 'able outperforms', 'edges', 'best practices', 'analytically orders', 'of algorithms', 'of of algorithms', 'of algorithms and problem complexity', 'of algorithms and systems', 'of phase transition', 'verifiable execution', 'interactive proof', 'least embeddings', 'standard embeddings', 'interactive clustering', 'query complexity', 'similarity function', 'partial observability', 'planning', 'model - free learning', 'model the to learn', 'model - based planning', 'by . matrices', 'partial bias', 'are to methods', 'learn results', 'are , to methods', 'are embedding - deep learning', 'enable tools', 'are embedding - deep , methods', ', embedding', 'character classification', 'sparse matrix estimation', 'recommender system', 'compute unified device architecture', 'mixed estimation explore', 'compute unified device', 'recommendation systems', 'community detection isometry property', 'recommendation estimation be formula', 'mixed membership stochastic block model', 'affect sensing', 'affect recognition', ', 0 <unk> 1 graphical models', ', 0 <unk> 1 regression', ', 0 <unk> 1 graphical model', ', 0 <unk> 1 random walk', ', 0 <unk> 1 random walks', 'prediction under a budget', 'adaptive approximation', 'test - time prediction', 'test - time resourceconstrained prediction', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 2', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 1 , 2 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 2 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 2 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1 , 1 , 2 , 1 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1', 'adaptive achieves , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 , 2', 'multivariate density estimation', 'bayesian density estimation', 'applications of extremes', 'high blocks', 'non - maximum likelihood', 'rhythmic variation', 'nano o o adversarial models', 'elman network', 'nano o o adversarial neural network', 'labeler be a', 'attractor neural networks', 'nano o o o adversarial models', 'have', 'and theoretical models', 'and and and and and genetic algorithms', 'and and and and and and genetic algorithms', 'and and and and and and and genetic algorithms', 'and and and and and and and and genetic algorithms', 'drawn and belong to', 'genetic algorithm', 'stochastic systems', 'and and and and and and and and and genetic algorithms', 'genetic algorithms', 'been', 'low - rank experts', 'lazy newton', 'an have algorithm', 'have a posteriori', 'perceptron', 'saliency', 'low - rank experts embedding', 'function prediction', 'and lazy learning', 'and gaussian processes', 'dynamic optimization', 'affine policies', 'worst - case analysis', 'probabilistic analysis', 'probabilistic modules', 'model interpreting', 'shapley additive explanations', 'shap', 'e <unk> learning', 'shapley additive explanation', 'additive feature proposed proposed targeting', '1 <unk> norm', 'additive feature importance measures', 'k <unk> nearest neighbor', 'k <unk> nearest neighbors', 'near <unk> computational complexity', 'k <unk> nearest <unk> neighbor', 'shapley', 'k <unk> nearest', 'logistic', 'as <unk>', 'inexact', 'inexact matrix functions', 'stochastic online learning', 'inexact matrix batch methods', 'e optimization', 'in model selection', 'inexact matrix gradient', 'budget constraints', 'for lower bounds', 'for deep optimality', 'for a posteriori estimation', 'for a posteriori probability', 'for a posteriori error estimation', 'stochastic noise', 'in a posteriori', 'precision and recall', 'for deep mixing', 'in a posteriori probability', 'dynamical isometry', 'deep sigmoidal networks', 'deep relu input - output jacobian', 'deep network', 'in to to data', 'in to to data ratio', 'log - concave distributions', 'disagreement community', 'concentration of measure', 's the matrix', 'y the roc curve', 's - concave distributions', 'fat - tailed distributions', 'case and statistical learning', 'case and statistical machine learning', 'case and machine learning', 'case and reasoning under uncertainty', 'fat the band matrix', 'for system identification', 'phase transitions', 'x , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 ,', 'x , 1 , 1 , 2 , 1 , 1 , 1 , 2 , 1 ,', 'x , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 ,', 'x , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 ,', 'x , 1 , 1 , 1 , 2 , 1 , 1 , 2 , 1 ,', 'x , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 ,', 'x , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 1 ,', 'x , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 ,', 'x , 1 , 1 , 2 , 1 , 1 , 2 , 2 , 1', 'x , 1 , 1 , 2 , 1 , 1 , 2 , 1 , 2', 'gradient matching', 'grid free inference', 'networks of equations', 'networks of data', 'context selection', 'literature <unk> based learning', 'disparate data', 'conditional probability estimation', 'compared clustering', 'experts model', 'literature <unk> based evaluation', 'literature <unk> based clustering', 'high <unk>', 'a model', 'high <unk> dimensional', 'high <unk> pass filtering', 'wide baseline stereo', 'sift', 'dependence analysis', 'advantages and embedding a group', 'smooths', '( transparency', 'advantages and embedding a analysis', 'enforce scene analysis', '( excitation', 'levels of detail', 'greedy coordinate descent', 'soft thresholding projection onto simplex', 'soft thresholding projection projection', 'soft thresholding projection projection results', 'sampler', 'approximations', 'capacity', 'sotopo of provide problems', 'of robots', 'in a captioning <s>', 'in a captioning <s> ( is', 'in a captioning <s> ( application', 'in a priori error bound', 'in a the captioning we is', 'in a captioning <s> ( synchrony', 'in a the captioning we application', 'geometry learning', 'knowledge - based learning', 'geometric reasoning', 'children', 'batch learning', 'knowledge - based system', 'child development', 'child e we visual cortex', 'estimators of learning', 'are visual cortex', 'child e we visual learning', 'child e we visual system', 'child e we visual', 'child e we visual field', 'contextual bandits', 'efficient it', 'divergences', 'or in context', 'rate sampling', 'rate adaptation', 'minimax entropy', 'layer - wise pruning', 'layer - wise optimal brain surgeon surgeon', 'in . situ design', 'feedback vector quantization', 'cloud computing', 'layer . images', 'layer', 'geodesically convex optimization', '( in', 'global bound', 'accelerated first - order methods', 'ability', 'first - transfer principle', '( ability', 'selective classification', 'reject option', 'mission - how focus', 'feedforward neural networks', 'variable neighborhood search', 'approximately correct classification', 'approximately correct algorithm', 'in its', 'in its derivatives', 'resulting in but all', 'resulting in the roc curve', 'receiver operating characteristic', 'multilayer perceptron', 'receiver operating characteristics', 'inverse covariance matrix', 'precision matrix', 'bandable precision matrix', 'gauss - markov graph', 'stochastic least squares', 'in can matrices', 'of sparse is matrix', 'of sparse is', 'game tree search', 'best arm identification', 'bandit problems', 'game trees', 'monte - carlo tree search', 'best arm of design', 'game tree policy', 'best arm of mechanisms', 'rapidly mixing markov chains', 'additive model', 'intrinsic group additive model', 'kernel smoothing', 'multiple kernel learning', 'is propose', 'group a posteriori', 'group deep structure', 'we sparse', 'we sparse matrix', 'we sparse model', 'method of separation', 'we its', 'analysis of smoothing', 'written character recognition', 'descriptive complexity', 'we language separation', 'we language', 'written character', 'structured phase retrieval', 'compressive sampling', 'phase retrieval', 'have sampling', 'have ) optimality', 'product', 'hashing', 'differential geometry', 'is modeling', 'is differential geometry', 'is differential', 'is', 'hawkes process', 'multivariate hawkes process', 'nonparametric process', 'approximate process', 'object - oriented framework', 'hardware implementation', 'hardware', 'interval - valued data', 'interval data', 'maximum medoids', 'cold start', 'latent models', 'while greatly', 'propose and cold', 'relational reasoning', 'rns', 'super - human retraining', 'relation networks', 'human question answering', 'super - human needs', 'sequential decision process', 'latent topic model', 'q , on', 'sequential decision processes', 'latent topic models', 'q , on learning', 'q , on . problem', 'latent dirichlet allocation', 'q , on . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 0 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 0 , 1 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'q , on . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d ,', 'stochastic games', 'multiple - play bandit', 'position bias', 'permutation method', 'transparency', 'unknown , uses and gratifications', 'position embedding', 'online advertising', 'active exploration', 'symbolic representation', 'symbolic representations', 'clone convolution', 'clone detection', 'metropolis and hastings', 'stochastic sampling', 'of random sampling', 'metropolis and <unk> hastings', 'nonparametric statistics', 'intrinsic distribution', 'intrinsic likelihood', 'our gaussian distribution', 'minimax sampling', 'minimax recovery', 'volume sampling', 'fair clustering', 'fairlets', 'price of fairness', 'impact doctrine', 'dual volume sampling', 'components', 'stable measures', 'are with . unified power', 'achieving lazy sampling', 'advantages and paradigm', ', under mixed ) sampling', 'by mixed )', 'by mixed data', 'by mixed methods', ', sample', 'by mixed ,', ', matrix', 'by mixed by', 'by mixed markov', 'stereo vision', 'opaque surfaces', 'uniqueness inhibition', 'occlusion', 'uniqueness inhibition constraint', 'occlusion consistency', '<unk> divergence', 'hindsight bias', 'implicit feedback', 'experience replay', 'implicit learning', 'collect', 'in neural networks', 'agents', '<s> distributions', 'a as', 'in we secondary sense', 'a <unk> learning', 'a we', 'a <unk> initialization', 'a . 4', 'biased from statistical to . efficient estimation', 'biased from statistical to . by bayesian', 'black box variational inference', 'biased importance sampling', '1 in . 1 - bayesian', 'biased from statistical to . has', 'we to divergence', 'reparameterization', 'to the divergence', 'problem of modeled structure', 'near a posteriori', 'we to divergence divergence', 'reparameterization divergence', 'near a a posteriori', 'we to noise', 'problem of mixtures', 'problem of moments', 'near a posteriori estimation', 'more than used', 'more <unk> posteriori', 'more <unk> disagreement', 'more <unk> posteriori estimation', 'adversarial inference', 'computed tomography', 'deep graphical models', 'directed latent variable models', 'point set', 'metric space', 'is precision', 'hierarchical neural network', 'to greedy', 'point sets', 'hierarchical neural networks', 'to greedy learning', 'hierarchical feature learning', 'of interest', 'to reasoning', 'this rns', 'regularization by noise manifolds', 'regularization by noise injection', 'stochastic give <unk>', 'space <unk> time optimization', 'space <unk> time gradient descent', 'denoising', 'preserving smoothing', 'preserving projection', 'optimizing divergences', 'given likelihoods', 'optimizing regularization', 'given prior', 'resulting in results', 'challenges', 'an model', '( optimality', 'reconstruction', 'message', 'personalized data', 'latent semantic analysis', '1 is', '1 is a', 'latent providing', 'latent variable multiple output gaussian processes', 'latent side <unk> effect models', 'metropolis - hastings', 'volume structure', 'construct distances', 'volume preserving of the the . norm', 'volume preserving of the the . norm the models', 'competitive learning', 'applied probability', 'primary visual cortex', 'orientation selectivity', 'orientation columns', 'orientation preferences', 'context - dependent mlps', 'darpa resomce management database', 'of experimentally gradient', 'hybrid multilayer perceptron', 'context - visual features', 'observable markov models', 'task learning', 'task demand', 'bayes risk', 'latent gaussian models', 'gain accurate pruned', 'gain accurate markov chain monte carlo', 'each training sample problem', 'by training sample selection', 'each training sample size', 'pareto optimality', 'context provide keywords', 'class provide keywords', 'each training sample a posterior', 'bidding', 'online auctions', 'saliency map', 'multiset prediction', 'human visual attention', 'multi - label game', 'image attention', 'multi - label classification', 'sequential image attention', 'multi - of for a value', 'losses', 'losses minimisation', 'be process', 'focus of attention', 'develop regions', 'develop bound prediction', 'network visualization', 'g . 1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 .', 'g . 1 . 1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 .', 'g . 1 . 2 . 1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 .', 't . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 .', 'g . 1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', 'g . 1 . 1 . 1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 .', 'g . 1 . 2 . 2 . 2 . 2 . 2 . 3 . 2 . 2 . 2 .', 't . 2 . 2 . 5 . 2 . 2 . 2 . 2 . 2 . 2 . 2 .', 'g . 1 . 1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', 'g . 1 . 2 . 1 . 2 . 2 . 2 . 2 . 2 . 2 . 2 <unk>', 'linear complexity', 'reproducing kernel hilbert spaces', 'stochastic gradient ascent', 'reproducing kernel hilbert space', 'gaussian process model', 'k - medoids', 'voronoi diagram', 'k - observable', 'k - matching', 'the the interpretation optimization', 'the the interpretation optimization algorithm', 'the algorithm mcmc', 'the form factor', 'the a posteriori probability', 'the closest point', 'the closest point method', 'prior knowledge', 'incremental learning', 'the are', 'using', 'transductive regret', 'swap regret', 'gps', 'regret theory', 'riemannian manifold', 'posteriori <unk> based learning', 'solvers', 'motion capture', 'self - supervised learning', 'test - time adaptation', 'test - time optimization', '3d mesh motion', 'test - time network', 'image - to , models , graphs , matchings', 'image - to , models , models , important questions', 'image - to , models , nature . inspired algorithms', 'image - to , models , graphs , graphs , matchings', 'cross - domain joint distribution matching', 'key distribution', 'recordings', 'objective function', 'log of random matrices', 'log of random graphs', 'lstm', 'especially', 'the develop', 'especially optimization', 'the especially', 'functions', 'filters', 'images the of of of of of of of of', 'images the of of of of of of of we', 'images the of of of of of of of are', 'images the of of of of of of are we', 'images the of of of of of of are of', 'images the of of of of of of we of', 'images the of of of of of are of of', 'images the of of of of of of of effect', 'network embedding', 'learning - to - rank', 'link prediction', 'computing methodologies', 'global ranking', 'multi - updates', 'is of future internet', 'consistently model', 'propagation model', 'be the semi hra index', 'at the supervised learning', 'at the margin', 'at the energy', 'min , max in graphs', 'be the network method', 'derivative - free', 'gradients', 'kernel learning is', 'derivative - is', 'providing', 'finally', 'costly set', 'the label queries', 'trust region', 'kronecker - in - , in of of a this an we', 'kronecker - in - , in of of a this an we we', 'kronecker - in - the of decisions', 'to the an an ( fail bit rate', 'queries', 'kronecker - in - , in of of a this an an we are a ( a ( a ( a clustering and classification', 'kronecker - in - , in of of a this an an we ( a ( a ( a ( a clustering and classification', 'kronecker - in - , in of of a this an an we ( a ( a ( a ( a posteriori probability', 'kronecker - in - , in of of a this an an we ( a ( a ( a ( a posteriori probability density estimation', 'video coding', 'video games', 'renyi differential privacy', 'posterior which', 'logistic setting', 'survey', 'outlier rejection', 'guarantees posterior a in incrementally', 'robust statistics', 'strongly convex', 'on - line algorithms', 'algorithms on words', 'algorithms on strings', 'state space model', 'bi - descent', 'recent linear mixing model', 'bi rotation learn', 'bi - directional recurrent neural network', 'single hidden layer feedforward network', 'single hidden layer feedforward networks', 'single hidden layer feedforward neural network', 'bi - directional recurrent neural networks', 'gaussian process , linear systems', 'in region of interest', 'precipitation', 'novelty detection', 'feedforward neural network', 'multivariate time series', 'adversarial imitation learning', 'reaching behavior', 'diverse behaviors', 'reaching', 'decentralized algorithms', 'distributed machine learning', 'centralized algorithms', 'local aggregative games', 'nash equilibrium', 'structural setting', 'strategic games', 'multi - agent interactions', 'task allocation', 'task assignment', 'well uncertainty', 'well noise', 'implies each each each depth', 'structural algorithms', 'provided pitfalls', 'structural . activity', 'structural . activity goal', 'proposed code method', 'rewards', 'optimal auctions', 'optimal auction design', 'rademacher function', 'linear quadratic regulation', 'linear quadratic', 'non - linear optimization', 'non - of inference', 'non - of that generating functions', 'spatial learning', 'inference control', 'specifies representing network', 'time . flight control', 'specifies representing networks', 'specifies representing network study', 'specifies representing network latency', 'only control', 'specifies representing network with cost', 'to algorithms', 'tree search', 'iteration', 'data recover', 'slow navigation', 'slow learning', 'and pairs', 'electroencephalograms', 'seizure - onset localization', 'graph - compared', 'graph - compared clustering', 'is evaluation', 'network inference', 'on . <unk> which , sparsity', 'factor - critic', 'critic - proposed model', 'on , <unk>', 'on , <unk> line learning', 'on , <unk> network mining', 'd , 2 , <unk>', 'on , long , temporal and objects', 'on , long , temporal and objects the target', 'on , long , temporal , objects the target', 'd , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 1 , 2 , 2 , 2 ,', 'd , 2 , 2 , 0 , 1 , 1 , 2 , 2 , 1 , 2 , 2 , 2 ,', 'd , 2 , 2 , 2 , 0 , 0 , 1 , 1 , 2 , 1 , 2 , 2 ,', 'd , 2 , 2 , 2 , 0 , 1 , 2 , 1 , 2 , 2 , 2 , 2 ,', 'd , 2 , 2 , 2 , 0 , 0 , 1 , 1 , 2 , 2 , 2 , 2 ,', 'd , 2 , 2 , 0 , 1 , 1 , 2 , 1 , 2 , 2 , 2 , 2 ,', 'd , 2 , 2 , 2 , 0 , 0 , 1 , 1 , 2 , 2 , 1 , 2 ,', 'd , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'd , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 1 , 2 , 2 , 1 ,', 'd , 2 , 2 , 0 , 1 , 1 , 2 , 2 , 1 , 2 , 2 , 1 ,', 'expected improvement', 'best - arm identification', 'expected improvement algorithm', 'best - arm identification problems', 'best - arm identification problem', 'best - deep learning', 'algorithm design', 'allow embedding', '<s> randomized algorithms', 'toy - points', 'need , these', 'need , benchmarks', 'need , for population', 'reward architecture', 'approximate supermodularity', 'greedy experimental design', 'supermodularity bounds', 'supermodular functions', 'greedy experimental large scale', 'experimental large scale', 'forward', 'domains', '<s> proposed algorithm', '<s> proposed algorithms', '<s> approximate classification', '<s> proposed', '<s> proposed classification', '<s> proposed class', 'an h , -', 'minimal', 'an , , -', 'iter', 'can , e', 'can ,', 'but ,', 'but , e optimality', 'error', 'realizable path selection', 'sequence - to - bound', 'be , supervised learning', 'be , branch and bound', 'sequence - to - bound prediction bound', 'be , supervised and unsupervised learning', 'dof degree distribution', 'be , branch - and model', 'supervision', 'partition', 'dependencies', 'important point', 'important )', 'missing modes', 'to a posteriori estimation', 'is a posteriori estimation', 'd space', 'd . 2 <unk> trees', 'd . 2 . 2', 'brain maps', 'd . 2 . 2 <unk>', 'd . 2 . 2 <unk> <unk>', 'algorithm engineering', 'all input processes', '( sequential monte carlo', 'nowcasting', 'learning by imitation', 'sequential monte', 'sequential', 'straggler optimization', 'data encoding', 'learning automata', 'distributed', 'distributed optimization', 'to of to of simulators', 'it we performance', 'straggler scheduling', 'it we resistance', 'straggler tasks', 'data kronecker product', 'multi - view decision', 'steering', 'multi - view', 'multi - view decision - making', 'joint policy', 'rise how', 'helper - ai', 'sequential games', 'maximum inner product search', 'greedy approach', 'prefrontal cortex', '- cortex', 'cortico - hippocampal interaction', 'discrimination learning', '<unk> cortex', 'basis pursuit', 'end model selection', 'proposed end in region selection', 'improvement', 'gaussian mixture model', 'and important heterogeneity', 'in of line learning', 'sequence - to - sequence model', 'sequence learning', 'structural alignment', 'performing arts', 'network planning', 'regulation', 'approximator', 'graph coloring', '<s> 7', 'labeled graph', 'end - to - end', 'stochastic programming', 'enabling factor', 'energy storage arbitrage', 'end - to - svd', 'then , component tree', 'electrical grid scheduling', 'then , reduced rules', 'with constraints', 'with constraint propagation', 'with constraint', 'from constraints', 'then , reduced models', 'class . convergence', 'class . additional information', 'inventory', 'class . additional cost', 'class . additional cost function', 'one of means', 'one of tuple', 'one of squares', 'joint distribution matching', '. in marketing', 'is . method', 'the conditional', 'is . 4', ', maximum likelihood', ', maximum likelihood estimation', 'a posteriori probability density', 'markov setting', 'gtd', 'policy evaluation', 'long - term accumulated reward', 'long - term difference', 'finite sample analysis', 'markov analysis', 'linear function approximation', 'markov bounds', 'mixing coefficient', 'finite sample', 'turbo principle', 'complexity theory', 'services', 'hierarchical implicit models', 'likelihood - free predict', 'overlap graphs', 'hierarchical implicit modeling', 'overlap that conservation laws', 'predator', 'moreover', 'comprehensive learning', 'overlap that constraints', 'currently', 'query by example', 'gpus', 'lfvi', 'manifold invariance', 'assumptions', 'manifold despite', 'inverse mapping', 'of approximation', 'even we', 'adversarial despite', 'adversarial', 'generative despite', 'direction - of - arrival', 'gated feedback network', 'aimed entity', 'settings', 'memory', 'rate distortion', 'rate , convergence', 'aimed incorporate', 'q - managerial hierarchy', 'q - managerial learning', 'q ( <unk> development', 'paper ( <unk>', 'q , be , <unk>', 'q , be for learning', 'q , be , <unk> , <unk>', 'q , be , 1 , 1 , introduce', 'q an changes', 'q an', 'q - case', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', 'q , be , 1 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2', 'q , be , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1', 'q , be , 1 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', 'min - max propagation', 'min - max inference', 'makespan minimization', 'belief estimation', 'epistemic uncertainty', 'input - completing data', 'and approach', 'epistemic . spatial data', 'epistemic . spatial the curvature', 'epistemic . ( on in sequential data', 'epistemic . ( on in digital images', 'stability', 'limit discrimination', 'limit', 'gradient descent gan', 'expected far', 'label noise', 'discriminative neural networks', 'residuals', 'deep convolutional neural networks', 'two discriminators', 'gan training', 'saddle point formulation', 'conjugate duality', 'saddle point modification', 'conjugate how', 'location', 'imitation', 'precipitation nowcasting', 'network management', 'convolution filters', 'crowding', 'visual similarity', 'pooling', 'adaptability', 'early layer', 'convolution filter', 'artificial deep neural networks', 'complementary label', 'estimation extend', 'complementary labels', 'estimation error', 'estimation error bound', 'labeling', 'learning . . .', 'outperforming', 'may at random', 'learning . . . .', 'conditional random field', 'may at learn', 'decaying memory', 'generalized alpha - investing', 'online control', 'online multiple testing', 'prior weights', 'quality - preserving', 'nondegenerate gaussian processes', 'function the', '2 - wasserstein metric', '1 . function classifier', 'networks of computer vision', 'can . function classifier', '1 mips', 'dimensional analysis', 'smooth analysis', 'radial basis function', 'discrepancy', 'discriminative models', 'forecasting', 'nonstationary time series', 'discriminative state space models', 'broad network flows', 'design of algorithms', 'design and analysis of algorithms', 'but a convex initially', '1 theory of machine learning', 'discriminative state - space models', 'in in in and 1 to theory of machine learning', 'calibration', 'managers', 'tension who world', 'tension', 'like bounds', 'imagination - augmented agents', 'design privacy', 'imagination augmented agents', 'of error', 'and error probabilities', 'imagination - global models', 'and of error', 'and error estimation', 'and error optimal control', 'and , ( <unk>', 'low - dimensional dynamics', 'moment matching it a', 'moment matching it', 'moment matching', 'multi a priori estimation', 'large - scale neural population dynamics', 'moment it principle', 'state estimation', 'large - scale neural recording', 'moment extend', 'label propagation', 'state space methods', 'low . diverse density estimation', 'promise', 'k a priori information criterion', 'low . diverse density sequences', 'pac bounds', 'probably approximately function', 'regret a', 'representable computations', 'dynamic bayesian networks', 'tandem mass values', 'tandem mass spectrometry', 'set partitioning', 'fisher kernel', 'replace replace method', 'map inference', 'asynchronous algorithms', 'gan which', 'coordinate minimization', 'dual - decomposition', 'map groups', 'parallel design', 'mpi', 'openmp', 'asynchronous', 'process scheduling', 'process algebra', 'similarity search', 'multiscale quantization', 'multiscale similarity search', 'multiscale methods', 'product quantization', 'image caption generation', 'gaussian mixture models', ') sequence learning', 'image description', 'auto - encoder', 'conditional ) mechanism', 'deeper algorithms', 'gaussian iterative model', 'expectation this maximization', '. )', 'an ) we learning', 'as - cvae', 'wasserstein gans', 'lipschitz condition', 'language model', 'rbf network', 'rbf networks', 'into - that', 'and - machine', 'bayesian approximation', 'evidence approximation', ') auction', 'natural image statistics', 'back - parameters', 'weight', 'bayesian \"evidence\" approximation', 'neural introduce', 'o auction', 'by deep learning', 'by deep inference', 'by deep belief functions', 'by deep belief function', 'by deep belief', 'scales', 'the prior knowledge', 'success rate', 'success function', 'the prior information', 'cumulative density function', 'earth mover distance', 'earth mover as a distance', 'inner distance', 'based sampling', 'as function', 'as function points', 'derive small sample size', 'derive small brain controller', 'gaps', 'gaps algorithm', 'function points', 'noisy queries', 'lower bounds', 'non - adaptive', 'non - adaptive settings', 'noisy oracle', 'higher - order tv', 'trend free', 'trend filtering', 'higher - order inference', 'sequences', 'process variation', 'sequences filtering algorithm', 'total variation addressed of a an family', 'total variation addressed of this lqr', 'in by prediction', 'processing time', 'processing prediction', 'sequential estimation', 'dof degree reduction', 'dof degree estimation', 'to prediction', 'quasi to quasi to quasi iterations', 'quasi to quasi to test', 'quasi to quasi to convergence', 'quasi to quasi to analog', 'vertex iterations', 'understandability', 'i ( optimality', 'public', 'flow', 'search', 'single - cell gene values', 'single - cell gene', 'in we , off for graphical models', 'in we , off for gene expression data', 'many , off for gene expression data', 'in we , off for gene estimating equations', 'non nns', 'high nns', 'time - varying data', 'functional mri', 'additive models', 'in models', 'brain imaging', 'spatially varying graphical models', 'fmri data analysis', 'brain mri', 'as topology', 'gradient methods', 'projected gradient methods', 'stochastic gradient methods', 'factor approximation', 'nonsmooth optimization', 'primal - dual interior - point methods', 'linearize and convergence analysis', 'convergence data analysis', '. c . c . c . c . c . d . c .', '. c . c . c . c . c . c . d .', '. c . c . c . c . d . c . c .', '. c . c . c . c . c . c . c .', '. c . c . c . c . d . c . d .', '. c . c . c . c . c . c . d', '. c . c . c . c . c . c . m <unk>', '. c . c . c . c . c . c . 1', 'in . s . m . c . c . c . c . c . c . c . c . c . c . d . c', '. c . c . c . c . c . c . 1 <unk>', 'learning to influence', 'congestion games', 'multiplicative weights update', 'limit cycles', 'concepts', 'multiplicative weights update method', 'potential games', 'naturalness', 'bilinear maps', '1 features', 'proportional current amplifier', 'distal conductances', 'pyramidal neuron', 'apical synaptic conductances', 'exuberant', 'nmda channel', 'cognitive neuroscience', 'again showing', 'this algorithm', 'np , k , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 2 , 3 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 0 , 1 , 1 , 2 , 2 , 3 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 3 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 3 , 3 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 0 , 0 , 1 , 2 , 2 , 3 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 2 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 0 , 1 , 2 , 2 , 2 , 2 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 0 , 0 , 1 , 2 , 2 , 2 , 3 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 2 , 0 , 1 , 2 , 3 , 3 , 3 , 3 , 3 , 3', 'np , k , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 3 , 3 , 3 , 3 , 3', 'bounded rewards', 'online caption contest', 'in - correction', 'fixed topology networks', 'collaborative deep learning', 'fixed topology network', 'fast - slow rnn', 'setups', 'fast - slow of rnn arms a rnn maxima', 'fast - slow of rnn arms a rnn maxima functions', 'claims a bandits', 'fast - slow of rnn arms a rnn maxima functions width of 1 , long memory', 'fast - slow of rnn arms a rnn maxima functions width of 1 , long - term memory', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - a - not - as - as - as - as - as', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - not - a - as - as - as - as - as', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - go - long learning', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - a - not - as - as - as - as - a', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - not - a - not - as - as - as - as', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - not - a - as - as - as - as - a', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - a - not - as - a - as - as - as', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - a - not - as - as - as - a - as', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - a - not - as - as - a - as - as', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - not - a - as - as - as - a - as', 'fast - slow of rnn arms a rnn maxima functions width of 1 , t - not - not - a - as - a - as - as - as', 'general domain models', 'general domain knowledge', 'general domain', 'chow forms', 'rules', 'even sets', 'hopfield networks', 'delays', 'even', 'intrinsic image decomposition', 'rin', 'learned shading', 'image decomposition', 'learned related processes', 'generalizing size', 'own size', 'sharpness', 'pac - algorithms', 'online testing', 'rejection threshold', 'false alarms', 'control charts', 'online false discovery rate', 'multi - a testing', 'control - pills', 'remaining useful life', 'control pills', 'schedule selection', 'schedule gradient', 'schedule estimation', 'schedule remaining processing time', 'remaining gradient', 'remaining gradient squared error', 'remaining gradient squared', 'coverage', 'latent space', 'image decoder', 'image - based rendering', 'explicitly from of mutual information', 'image - based image synthesis', 'image from of mutual information', 'explicitly from of mutual images', 'image cost estimation', 'hyperparameter estimation', 'hyperparameter optimization', 'hyperparameter decoding', 'visual example be', 'for in domains', 'but in digital photography', 'but in digital images', 'in and the domains', 'first and second domains', 'in and the domains of in images', 'nonlinear image processing', 'in and the domains of in image', 'in and the domains of in image processing', 'deep sub domains', 'image domains', 'scalable video coding', 'video compression', 'scalable video compression', 'scalable video processing', 'visual', 'mobile health', 'reward', 'invariance', 'web', 'tuples', 'disabilities', 'plan', 'strongly debugging', 'linear', 'and data', 'machine vision', 'simd', 'normalized cross a', 'simd processors', 'variable secondly', 'abstractions', 'larger by computations', 'larger in parallel approaches', 'discrete - continuous mixtures', '3h - propose estimation', '3h - propose because nets', 'advances', 'mutual uncertainty', 'search trees', '3h - principle', 'at for and domains', 'single - model state - of - the . . if', 'recognition', 'dominant sequence transduction', 'single - model state - of . . the model scale', 'arise to baseline', 'and can', 'shapley value', 'general setting', 'shapley values', 'and can function', 'harmonics', 'decomposition', '1 for 1 for 1 for drawn for drawn for drawn for drawn for', '1 for drawn for drawn for drawn and et can function', '1 for drawn for drawn for drawn for drawn and et can function', '1 for 1 for drawn for drawn for drawn for drawn for drawn for', '1 for 1 for drawn for drawn for drawn and et can function', '1 for 1 for drawn for drawn for drawn for drawn and et can', '1 for 1 for drawn for drawn for drawn for from can function', '1 for 1 for 1 for drawn for drawn for drawn for drawn and', '1 for 1 for drawn for drawn for drawn for drawn for from can', '1 for 1 for drawn for drawn for drawn for drawn for bound', 'hierarchic', 'video', 'autoregressive models', 'in bandit models', 'psrs', 'random variables', 'alternative input', 'alternative splicing', 'oracle inequalities', 'random prove', 'random playout', 'oracle inequality', 'parameter - free learning', 'of to involves irregular collective can', 'of have for latter', 'is problem complexity', 'in . sample complexity', 'of the assumptions', 'is guarantee', 'in . sample consensus', 'in . sample paths', 'a posteriori probability bounds for machine learning', 'adversarial prediction', 'saddle - point problems', 'suf?cient & enables broad continuous selection', 'suf?cient & enables broad continuous costs', 'suf?cient & enables broad continuous ensembles', 'be aspects', 'be learning', 'reduction of graphs', 'continuous cache models', 'online language modeling', 'language models', 'predictive state representations', 'predictive state rd samples', 'state recurrent neural networks', 'predictive oracle', 'bayes filters', 'two - setting samples', 'state recurrent neural network', 'two - stage regression', 'recurrent selection', 'l2 - boost', 'logitboost', 'adaboost', 'test propose', 'early stopping', 'training set', 'less a posteriori error bound', 'less a posteriori error estimate', 'after a family of not formula', 'after a family of not components', 'after a family of not function', 'after a family of not function methods', 'be show through', 'many show through', 'in show through', 'singular vector canonical correlation analysis', 'difficult supervised learning', 'based learning', 'convolutional phase retrieval', 'cyclic convolution', 'restricted isometry property', 'random circulant matrices', 'restricted isometry rate', 'random circulant matrix', 'spectral initialization', 'generalized gradient descent', 'approaches to', 'demonstrate of', 'central clustering', 'pairwise data clustering', 'competitive neural networks', 'maximum evidence theory', 'maximum entropy', 'pairwise theory', 'maximum evidence', 'able data', 'and unsupervised learning', 'multi - index model', 'stein?s lemma', 'truncation abstracts', 'parallel factor analysis', 'signal the corresponds', 'non - gaussian multiple index model', 'number of <s> reinforcement', 'determine leads', 'bars problem', 'gaussian quadrature', 'value prediction network', 'deep q - policies', 'model - for , to propose cycles', 'model - for , to . carry', 'model - as , to propose cycles', 'model - for , by the the signaling', 'model - for , by the the signaling network', 'model - for , to the the signaling network', 'model - for , to the the signaling', 'dependency parsing', 'margin', 'learning error', 'online kernel learning', 'second - order updates', 'adaptive embedding', 'second - order methods', 'implicit regularization', 'nuclear norm', 'ill <unk> posed problems', 'ill <unk> posed problem', 'nuclear norm minimization', 'a posteriori selection', 'singular values', '<unk> model', 'contamination', 'o model', '<unk> - <unk>', 'in <unk>', 'online mirror descent', 'game is', 'equilibrium algorithm', 'gaussian a - significantly action', 'connections enables algorithms', 'for computing with words', 'for deep learning', 'connections enables with kernels', 'delay', 'recent a estimate', 'next generation sequencing', 'near boltzmann machines', 'near optimality', 'asynchronicity', 'only optimization', 'internet protocol', 'concurrency', 'near boltzmann', 'near', 'frank - wolfe method', 'trace - performance', 'trace - norm ball', 'nlp', 'well , its', \"' s method\", \"' s the condition\", \"' s the basis\", \"' s the basis function\", 'model order reduction', 'model - preserving', 'control structures', 'remaining processing time', 'tasks', 'rejection', 'schedule', 'convexity', 'remaining useful', 'remaining useful discrete', 'error bars', 'spectroscopy', 'mackay', 'natural bayesian statistics', 'natural is', 'stochastic block model', 'random graph models', 'random graph model', 'hierarchical stochastic block model', 'random graphs', 'worst - case', 'argue data', 'age', 'is validation', 'is <unk>', 'is and information matrix', 'is performance', 'deep convolutional representations', 'our 1', 'deep signal representations', 'reproducing of estimate spaces', 'deep signal representation', 'modeling and prediction', 'in from vector space', 'in from vector spaces', 'in from vector quantization', 'in from below and core', 'cost sharing', 'cooperative game theory', 'width', 'universal approximation', 'expressive power', 'expressiveness', 'multiclass generalization bound', 'margin bounds', '<s> symbolic regression', '<s> latent variable models', 'cooperative learning', 'bayesian neural networks', 'hidden parameter markov assignment', 'hidden parameter markov assignment problem', 'hidden parameter markov decision process', 'bayesian neural network', 'studied programmed simulation', 'hidden parameter markov decision processes', 'theoretical learning', 'we . abstract gradient', 'we . abstract algorithms', '1 to boosting', '1 to in as topic', '1 to in place', 'such', 'graphs', 'choose knowledge', 'present', 'case', 'better than', 'does down', 'population matching discrepancy', 'finding metric', 'minimum weight matching where sample architectures', 'kernel bandwidth hyper - parameter', 'variant finding', 'salesman problem', 'heuristics', 'finer approximation', 'minimum weight matching discrepancy', 'population matching where sample architectures', 'evaluation function', 'scale', 'maximum further algorithm', 'scale arbitrary has', 'cnns', 'inequalities', 'demonstrated optimization', 'is chow forms', 'is chow form', 'is continuance probability', 'pseudo defined function', 'pseudo defined boolean model', 'pseudo defined boolean function', 'primitive', 'discrete', 'hybrid planning', 'tensorflow', 'exploding gradient', 'mixed integer linear programming', 'exploding that we', 'mixed discrete and continuous nonlinear domains', 'boltzmann exploration', 'stochastic multi - armed bandits', 'sequential decision - making', \"bernstein ' s inequality\", 'question - answering', 'be approximators', 'we approximators', 'paper - based learning', 'sentiment analysis', 'machine translation', 'context vectors', 'machine store', 'parallel feature extraction', 'first principles', 'labelling', 'first search', 'generalized bayes', 'generalized bayes classifier', 'generalized bayes rule', 'discrete representation', 'vector quantisation', 'vector quantisedvariational autoencoder', 'discrete latent representation', 'repeated measures', 'discrete latent models', 'mirror neurons', 'prove best practices', 'turing test', 'sensor network', 'very large <unk> scale experiments', 'very <unk> large <unk> scale experiments', 'regularizers', 'recently used', 'lanczos', 'kernel x free methods', 'methods and complexity', 'chebyshev', 'entries', 'x in d', 'countering information', 'and infers random processes', 'cell biology', 'and in on of linear equations', 'cell decomposition', 'cell signaling', 'is classification', 'd', '<unk> deep', 'hyperbolic geometry', 'hyperbolic space', 'symbolic data analysis', 'hierarchical representation', 'scenarios', 'players', 'graph edit distance', 'maximum cut', 'meta - heuristics', 'vertex cover', 'traveling scale', 'shares', 'value prediction', 'learned high nns', 'must nns', '<unk> is', 'real', 'conditional probability', 'semi - unsupervised learning', 'same rather ( 1', 'same rather ( some 1', 'same in rather o . algorithm', 'same rather rather o . algorithm', 'same in rather o . systems', 'reducing', 'reducing algorithm', 'give a', 'and setting', 'learning with exponential weights', 'no - regret learning', 'bandit feedback', 'o ( <unk> optimality', 'o ( applications', 'semi - bandit accuracy', 'contextual learning', 'contextual choice', '<s> maps', 'by in in any delays', 'by in in . degrees of freedom', 'learning in games', '<unk> . guarantee', 'to . e . model', 'to . e . e', 'to . e . e . learning', 'to . e . e . model', 'to . e . e . e . learning', 'to . e . e . e . model', 'to . e . e . e . e', 'to . e . e . e . e . model', 'to . e . e . e . e . learning', 'to . e . e . e . e . e . model', 'to . e . e . e . e . e . e . e . e . e', 'to . e . e . e . e . e . e . model', 'to . e . e . e . e . e . e . e . e . model', 'to . e . e . e . e . e . e . e . model', 'to . e . e . e . m . e . e . e . e . e', 'to . e . e . m . e . e . e . e . e . e', 'to . e . e . e . e . e . e . e . e . learning', 'to . e . e . e . e . e . e . m . e . e', 'to . e . e . e . e . e . e . e . m . e', 'to . e . e . e . e . m . e . e . e . e', 'to . e . e . e . e . e . e . e . e', 'actor - critic x', 'actor - critic x systems', 'multi - to , as', 'multi - agent case: q - learning', 'actor - critic prediction', 'multiagent', 'actor', 'models of cooperation', 'agent coordination', 'extensive infrastructures', 'communication complexity', 'addition of distributions', 'statistics', 'available projected sampling', '2 - d data', 'here - efficiency', 'here - and no - here here here here', 'here - and no -', 'available all', 'here - and no - here here here', 'here - and no - here here here here here here here here', 'here - and no - wait', 'here - and no - here here', 'here - and no - here here here here here here here here here', 'here - and no - here here here here here here here random', 'predictive is', 'deep ensembles', 'predictive uncertainty estimation', 'deep at , polynomially solvable', 'evolutionary computation', 'internal dynamics', 'behavioral evolution', 'unsupervised feedback', 'genetic programming', 'recently , for networks', 'counterfactual inference', 'schatten norms', 'matrix norm estimation', 'eigenvalue estimation', 'singular value estimation', 'approximate bayesian inference', 'brain - machine interface', 'is algorithm', 'brain - machine interfaces', 'fpga', 'variational autoencoders', 'effect inference', 'dag', 'variational inequalities', 'blanket discovery', 'subspace clustering', 'convex analysis', 'directed acyclic graphs', 'conditional independence separation', 'non - identifiability', 'polynomial time learning', 'activity learning', '1 and 2 test', '1 and 2 test sets', 'gradient episodic memory', 'parallelisation', 'processor', 'cannot interpretation', 'binary decision tree', 'how feature', 'capture', 'capture data', 'capture effect', 'situated', 'nature <unk> inspired algorithms', 'many <unk> sources <unk> first search', 'norm', 'code', 'version', 'b <unk> tree', 'b', 'b <unk> trees', '1 <unk> tree', 'a <unk> tree', 'similar', 'and <unk> tree', 'stable instances', 'euclidean k - means', 'additive sequence', 'state of relationships', 'gaussian graphical models', 'gaussian graphical model', 'long memory', 'amnesia', 'markov process', 'prediction approaches', 'variable memory length markov process', 'statistical prediction error', 'multiclass answers', 'learned model', 'network information theory', 'measurement will', 'viterbi algorithm', 'measurement error', '<s> a posteriori probability', '<s> intelligent complex detection', 'wide <unk> baseline learning', 'wide <unk> baseline 1', '2 <unk> d model', '2 <unk> <unk>', '2 <unk> d experimentally', '2 <unk> d state space model', '2 <unk> d state space', '2 <unk> d <unk> model', 'manner', 'total', 'symmetry', 'jacobian', 'to model order reduction', 'gaussian elimination', 'gym', 'to model synthesis', 'to natural optimization', 'to natural gradient', 'compute and conquer', 'to natural extension', 'to natural gradient reconstruction', 'protein interface prediction', 'graph convolutional networks', 'graph convolution', 'graph oracle', 'protein structure prediction', 'solid harmonic wavelets', 'solid harmonic functions', 'quantum decipherment', 'quantum molecular densities', 'invariant demonstrate', 'and adversarial models', 'invariant descriptors', 'quantum molecular corpora', 'quantum', 'solid , and as a can complex experiments', 'solid', 'solid , harmonic', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 1 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 ,', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d ,', '1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 3 ,', 'continuous control', 'test', 'data structures', 'affinity', 'best is', 'sequence', 'sequence parsing', 'certain answers', 'certain increment', 'certain', 'our in art', 'random projection filter bank', 'random method', 'random', 'random method ) agents', 'parallel algorithm', 'random projection agents', 'scene detection', 'recursive estimation', 'state space', 'evidence lower bound', 'maximum ) model estimation', 'provide variational', 'easy agents', 'which in generate data', 'x memory', 'variational mirror', 'help in generate model', 'fixed point method', 'help in generate data', 'complex wishart distribution', 'complex wishart matrix', 'definition of stable distributions', 'mixture of experts', 'equilibrium', 'pam', 'biological control', 'regret visual functions', 'biological networks', 'on . 2 . 2', 'differences', 'approaches with estimating equations', 'approaches and control', 'approaches with estimating', 'biological system', 'biological network', 'approaches for estimating equations', 'language', 'ablation studies', 'visual processing', '1 recommendation', 'quicker problem', 'conditional batch normalization', 'convolutional experiments', 'mixture of gaussians', 'online estimation', 'streaming data', 'online pca', 'that applications', 'our applications', 'rate applications', 'estimation of distribution algorithm', 'similarity estimation', 'near - neighbour search', 'mixed tabulation', 'permutation hashing', 'hash function', 'random hashing', 'credit assignment', 'long - portrait algorithm', 'long - term dependencies', 'digits recognition', 'long - portrait', 'call benchmarks', 'long - portrait learning', 'long - portrait dependencies', ', the invariant', 'the invariant class', 'the invariant reasoning', 'the invariant algorithm', 'the invariant adversarial environment', 'the invariant adversarial models', 'the invariant adversarial learning', 'the invariant neural networks', 'time - scale update rule', 'many means clustering', 'case - based reasoning', 'huge models', 'case study', 'case studies', 'is ) directed evolution', 'local nash equilibrium', 'time - have means clustering', 'case - based learning', 'high - dimensional prediction', 'high - our algorithms', 'high - our property', 'prediction with noise', 'proper time prediction', 'understand and second to regularizing iterations', 'o hashing', 'factor graphs', 'functions of bounded variation', 'functions of gaussian processes', 'factor analysis', 'functions and statistics', 'bicriteria approximation', 'randomized rounding', '`0 - low rank approximation', '`0 - low rank approximations', '`0 - low for streaming algorithms', 'low - for streaming algorithms', '`0 - streaming algorithms', '`0 - streaming algorithm', 'good?turing estimator', 'absolute all - dimensional distribution', 'follow ) estimation', 'powerlaw estimator', 'proposed dimensional networks', 'follow ) distance', 'absolute all - dimensional distribution estimation', 'proposed dimensional model', 'using scenarios', 'complex - valued cross - covariance', 'spectral mixture require', 'kernel density estimation', 'many . <unk>', 'although huge random processes', 'expectation maximization', 'spatial mixture models', 'perceptual grouping', 'techniques', 'agnostic learning', 'spectral filtering', 'problems with dynamics', 'discrete - time linear dynamical systems', 'sampling with dynamics', 'stochastic recurrent neural networks', 'will our moments', 'sum of squares', ', is clustering', 'stochastic recurrent networks', 'a - our moments', ', complex certain increment', 'we to pass learning', ', complex certain increment processes', ', complex certain answers', ') a <unk> model', ') a new certain increment', 'cases', 'causal models', 'be . is', 'at <unk> all', 'core', ') model', 'accordingly', 'at <unk> speed network', 'neuron mos transistor', 'biological neural network', 'processed architecture', 'vmos', 'and and and item simulation', 'and and and item generation', 'universal controller', 'and and and item classical conditioning', 'processed techniques', 'however hardware', 'at this classification', 'at this hardware', 'at this efficiently computable hardware', 'at this efficiently computable functions', 'at this efficiently computable state', 'black - box learning', 'sparse learning', 'is that optimality', 'conic geometry', 'tangent cones', 'tangent cone membership', 'tangent cone membership retina', 'tangent cone membership test', 'conic exploiting approaches', 'conic subspace', 'continuous - time recurrent neural networks', 'multivariate point input processes', 'continuous - task learning', 'bank memory', 'four ) model', 'recurrent models', 'domains of learn', 'domains of spiking neurons', 'domains of attraction dynamics', 'human factors', 'reliability', 'verification', 'g to guarantees', 'by is hacking but it', 'assume fundamental analysis', 'to datasets', 'risk aversion', '+ <unk> measure', 'risk - aversion', 'bayesian pruning', 'log - normal multiplicative noise', 'multiplicative this performance', 'log - normal prior', 'multiplicative noise', 'performance modeling', 'structured sparsity', 'log and on this in motion', 'structured bayesian pruning', 'future motion', 'auxiliary scene analysis', 'future baselines', 'chromatin', 'chromatin structure prediction', 'chromatin structure', 'gene regulation', 'selective attention', 'feature visualization', 'attend and reduce', 'attend and predict: understanding', 'we for , algorithm', 'we for algorithm', 'we for tree structure', 'we for tree', 'accelerated mirror descent', 'stochastic dynamics', 'covariation', 'one ability', 'asymptotic descent', 'mathematical statistics', 'convergence descent', 'asymptotic learning descent', 'similarity triplets', 'kernel functions', 'high - dimensional spaces', 'high - dimensional embedding', 'error detection', '3d convolutional networks', '3d convolutional network', 'multiscale 3d convolutional network', 'connectomics', 'a posteriori error latent semantic analysis', 'a posteriori error detection', 'of measurement', 'a posteriori error latent feature extraction', 'a posteriori error latent semantic indexing', 'hessian matrix', 'a posteriori error latent feature detection', 'a posteriori error latent semantic description', 'a posteriori error latent feature analysis', 'a posteriori error latent feature map', 'non - parallel text', 'sentiment', 'cross - alignment', 'non - parallel based across words', 'cross adversarial information design', 'cross adversarial information retrieval', 'non - all text', ') problems', 'smd', '1 detection', 'word neurons', 'non - all across words', 'universal learning', '\"hilbert \\' tenth problem', 'universal learning algorithm', '\"loading \\' tenth problem', 'cross - spectral factor analysis', 'local field potentials', 'meg', 'epilepsy', 'structures', 'coverage functions', 'stochastic submodular optimization', 'in - network optimization', 'in - network aggregation', 'of ( <unk>', ', one - sided instances', 'prove of results', 'optimal also', 'parsing', 'cross 1 criterion', 'be 1 criterion', 'cross 1 optimization', 'cross 1 norm', 'be 1 minimization', 'minimum spanning tree', 'minimum spanning trees', 'affinity clusters', 'affinity clustering', 'transformation learning', 'convex relaxations', 'convex relaxation', 'flow visualization', 'image reconstruction', 'image registration', 'ability estimation', 'across d reconstruction', 'multiple changepoint detection', 'fused lasso', 'sharp error supervised learning', 'linearizing model', 'sharp error analysis', 'on action potentials', 'scattering matrix', 'sum - product network', 'sum - product networks', 'bayesian online algorithms', 'linear - time algorithm', 'sum - our', 'linear time scale', 'linear - in ( )', 'tweet', 'cold - start', 'tweet a', 'tweet recommendation', 'meta - scheme', 'meta of remedy', 'is . under uncertainty', 'redistributes prediction', 'scene parsing', 'motion prediction', 'motion dynamics', 'motion among', 'optical ( model', 'motion marks', 'humans', 'scene and of chromatin', 'standard reparameterized gradient', 'we expectation', 'gradient estimator', 'simple', 'importance problem', 'of prior', 'is likelihood', 'importance rate', 'pairs of random variables', 'of a models', 'defined function', 'are at random variables', 'strings', 'pattern matching', 'similarity score', 'similarity measures', 'pairwise similarity score', 'multiple target tracking', 'phase - locked oscillator', 'to . <unk> optimization', 'to . <unk> . 4', 'to . <unk>', 'union - find', 'covariance minimization', 'kernel dimension reduction', 'filtering', ', 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 . 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 . 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 ,', ', 1 . 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 . 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 . 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 . 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 . 2 , 2 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', ', 1 . 2 , 1 , 2 , 1 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', 'gaussian mixture of gaussians', 'expectation - maximization algorithm', 'to for convolution', 'to and contaminated normal distribution', 'to for interest rate', 'is causal discovery', 'a posteriori error criterion', 'is causal inference', 'saliency detection', 'black box classifier', 'object localisation', 'weakly supervised learning', 'structured visual and speech recognition', 'adversarial examples', 'deep structured visual and speech recognition', 'pose estimation', 'peer descent', 'peer of review', 'peer of supervised learning', 'on which constraints', 'advances free approach', 'on which constraint', 'locate fixed points', 'parameter inference', 'variational approximation', 'state inference', 'variational is', 'gradient approximation', 'gradient search', 'in <unk> network flows', 'cyclic coordinate descent', 'randomized coordinate to is is is', 'be network calculus', 'log sequential approximations', 'randomized coordinate to is is', 'cross layer optimization', 'cross our data', 'high learning rate', 'cyclic bound', '1 learning', 'log sequential approximation', 'moments of inertia', ') input input learn', ') learn', 'first input input learn', 'first input input multiple output', 'to variance ratio', 'all variance', 'o variance ratio', 'randomized bound', 'is for for for for for pdes', 'is for for for for for for for for for', 'is for for for for for for pdes', 'is for for for for for for optimality', 'is for for for for for for for pdes', 'is for for for for for for for optimality', 'for for for for for for for for for for', 'is for for for for for for for for pdes', 'is for for for for for for for for optimality', 'o variance', 'mistake bounds', 'latent variables', 'causal graphs', 'depends challenges', 'must challenges', 'encodable challenges', 'introduce adam analysis', 'imbalanced datasets', 'meta - knowledge', 'many - shot learning', 'impossible data', 'coherent in . we', 'confounders', 'information develop', 'cross a', 'from', 'and networks', 'bayesian justification', 'correction tenn to - non linear classifier', 'a , representations', 'i - o functions', 'describe . <unk>', 'at of activations', 'at of hessian', 'i - o function', 'in on and', 'is on manifolds', 'maximum selection', 'stochastic transitivity', 'pairwise comparisons', 'stochastic paper', 'pac maximum selection', 'intelligent of choice', 'new system comparison', 'random pairwise comparisons', 'of stochastic intra sub', 'only per sub', 'class vlsi ranking', 'a vlsi ranking', 'between stochastic fabricated pairs', 'research', 'communities', 'averaged <unk>', 'averaged networks', 'activity detection', 'that streams', 'that interest rates', 'network data', 'hmm', 'that <unk>', 'that interest', 'network analysis', 'based on role clustering', 'p', 'user feedback', 'convergent learning', 'convergent neural networks', 'actually', 'have matrix', 'computer modeling', 'covers', 'computer learning', 'have finite', 'have free', 'skeletal data', 'represented groups', '<unk> . <unk>', 'linear dynamical system', 'skeletal system', 'biological neural networks', 'massively parallel supercomputers', 'propose recognition', 'tones enhancement', 'molecular simulation', 'propose and computing', 'spike neural networks', 'propose show through', '384', 'molecular dynamics', 'tones of attraction', 'space , molecular dynamics', 'associative memories', 'neural net', 'associative neurons', 'optimal signalling', 'history - dependent or distributions', 'history - dependent p neural networks', 'extended kalman filter', 'high similarity 1 with offline connections', 'in general and medium with time delays', 'system dynamics', 'high similarity 1 with one momentum', 'system analysis', 'history - dependent p neural could', 'multiprocessor system', 'neural net simulation', ', ,', 'high performance neural net', 'music stands', 'computational model', '<unk> - down\"', '<unk> - up\"', 'feedback control', 'decision theory', 'emergence', 'mixtures of multiple causes', 'multiple cause models', 'conventional more trajectory', 'oscillations', '1/f noise', 'random walk', 'cross - correlation', 'leaky integrateand fire neuron', 'leaky integrateand - neurons', 'representational activity', 'granger causality', 'one the vector potential', 'has ) model', 'leaky integrateand the and are to moore', 'leaky integrateand the and are in to moore', 'but )', 'experience ) model', 'leaky integrateand the and are to in to moore', 'satisfiability', 'mean field theory', 'k - separate', 'k - satisfaction', 'visual object recognition', 'feature correspondence', 'sets', 'is for measurement', 'non - rigid hand gestures', 'non - rigid 3d high resolution', 'medicine', 'medicine and science', 'correspondence', 'medicine and technology', 'counting function', 'counting functions', 'voicing', 'caused emg', 'emg', 'adjacency', 'adjacency matrix', 'trained network', 'preliminary design', 'each per , <unk> can', 'text planning', 'each per , )', 'preliminary design of algorithms', 'pattern formation', 'max , k and k region', 'max , workshop', 'region growing', 'resource allocation', 'threshold computation', 'max , k and workshop', 'sequential decision making', 'retrieval', 'analogical similarity', 'holographic reduced representations', 'analogical representation', 'dot - product rankings', 'in - network processing', 'network structure', 'in - network structure', 'network coding', 'network structure prediction', 'in in to complex structure', 'in in to given domain', 'fault - prone software modules', 'software metrics', 'feed - by . ii', 'is diagnosis', 'is software learning', 'on . 0', 'is software passed', 'dp methods', 'the algorithms', 'the dp methods', '- game', 'state . 1 <unk> space', 'state - space', 'search space', 'presented', 'retina', 'modeled', 'robot', 'measures', 'mechanics', 'measures of action', 'investigation', 'neuron', 'if', 'iteratively', 'higher dimensions', 'leech', 'swimming', 'metropolis algorithm', 'task decomposition', 'cross - validation', 'metropolis algorithms', 'nearest neighbor', 'locally adaptive', 'knn', 'locally neurons', 'to , em', 'and metric spaces', 'and to data', 'to , data mining', 'to , )', 'multiple classifiers', 'multiple instance learning', 'in - network', 'boltzmann machine', 'learning algorithm', 'gradient descent learning', 'learning in kinematics', 'by <unk> example learning', 'it implementation', 'cmos scaling', 'learning in vector quantization', 'learning classifier systems', 'data acquisition', 'learning in vector fields', 'learning in vector space', 'layered neural networks', 'information for surface', 'layered neural network', 'output function', 'modified forward forward algorithm', 'information for surface fitting', 'information for for or', 'feed - forward nets', 'feed - forward net', 'feed - forward networks', 'description schemes', 'indirect inference', 'generic training kinds', 'consist been', '4 architecture', 'consist been application', 'error surface', 'description condition', 'critical', 'network design', 'phase transition', 'phase f f test', 'pattern classification', 'node cluster', 'constructed correlation', 'independently trained network', 'behavior', 'power law', 'behavior analysis', 'physiology', 'decision support', 'decision support system', 'group decision making', 'group decision support system', 'group decision support systems', 'risk management', 'risk assessment', 'approximating network', 'estimation', 'approximating network design', 'risk', 'back - measurements', 'probabilistic cardiac models', 'characterization', 'necessary and sufficient conditions', 'field programmable gate arrays', 'known values', 'it information environment', 'established', 'it is', 'either', 'it information retrieval', 'field size', 'handwritten digit recognition', 'field programmable gate array', 'mixture of controllers', 'jump linear systems', 'hmms', 'functions of state space', 'models of control', 'number of adaptive systems', 'mixtures of controllers', 'jump linear and ) hmms', 'rhythm in speech', 'inverse and may result', 'inverse and may generalization', 'inverse and may performance', 'voice production', 'we , application', 'history of vision', 'resource - limited', 'history', 'simulated annealing', 'conjugate gradient descent', 'tank neural we', 'f combinations', 'first combinations', 'polynomials', 'derivative free grammars', 'cross - layer perceptrons', 'cross - layer perceptron', 'twelve for problem', 'path predictable design', 'produced data', 'collection', 'angle', 'feed correlation', 'specific concentration', 'effort', 'produced web', 'alternately', 'annealed', 'feed', 'text', 'an algorithm', 'effect', 'feed forward', 'correspondence analysis', 'cardiac electrogram', 'analogue vlsi', 'pacemaker that averaged <unk>', 'cardiac deliver', 'outputs of rankings', 'these neural network', 'memory ) compared', 'pacemaker that averaged synapses', 'memory ) paper', 'cardiac deliver chip for this method', 'hurdle models', 'hurdle model', 'form finding', 'form', 'intra', 'inside <s> 1 . reduced brain comparisons', 'inside to the implementation', 'extensions', 'motoneuron', 'renshaw cells', 'renshaw interneurons', 'power spectrum', 'weak feature', 'weak', 'simultaneously o a', 'simultaneously backprop', 'activation times', 'a . <unk> 1', 'activation matrices', 'activation cross . paper', 'single of discriminative models', 'activation cross . node', 'single of discriminative model', 'activation cross . load', 'formal language learning', 'finite state machine', 'formal language', 'larger order', 'larger order network', 'dependence network', 'invertebrate', 'alpha helix', 'beta which', 'protein secondary structure', 'the the movements', 'units', 'co - evolving neural networks', 'protein the the the a cells', 'layer use', 'layer prediction', 'layer use of method', 'the we', 'protein of the the the of the images', 'protein of protein the the of the images', 'protein of protein the the of of images', 'protein of protein the the of of the images', 'protein of the the the of of the images', 'protein of the the the of the images in the shapes', 'protein of the the the of the board of protein , protein , secondary structure', 'protein of the the the of the board of protein , protein , protein structure classification', 'protein of the the the of the board of protein , protein , protein be be useful', 'protein of the the the of the board of protein , protein , protein be improved be useful', 'protein of the the the of the board of protein , protein , secondary the be networks , ligand', 'protein of the the the of the board of protein , protein , protein be be improved be useful', 'protein of the the the of the board of protein , protein , protein be improved be be useful', 'protein of the the the of the board of protein , protein , secondary the be networks , ligand projection', 'protein of the the the of the board of protein , protein , secondary the be networks , ligand , receptor', 'protein of the the the of the board of protein , protein , secondary the be networks , ligand , software', 'protein of the the the of the board of protein , protein , secondary the be networks , ligand , protein software', 'componential models', '2 geometry', 'images', 'melody', 'from geometry', 'plasticity', 'of techniques', 'tonal memory', 'pitch', 'inputs', 'just based', 'just based time', 'have edges', 'development', 'er', 'critical events', 'drawings', 'each approach', 'discussed', 'each breaking', 'have resolving', 'time counterpart', 'each', 'melody its', 'each are time counterpart', 'each clusters', 'clusters', 'time approach', 'temporal pattern recognition', 'motion detection and threshold extract', 'temporal pattern processing', 'motion detection and threshold', 'central nervous system', '2 - d multi 2 2 a optimal maps', 'system of visual of and and with', '2 - d multi 2 2 a optimal descriptor', 'infer', '<unk> models', '<unk> direct', 'lead', '<unk> by', 'human genes', 'sequence analysis', 'kinematics', 'us', 'motif discovery', 'motif detection', 'sequence alignment', 'network complexity', 'imse', 'average squared error', 'integrated mean squared error', 'segmental neural net', 'phonetic modeling', 'speech networks', 'phonetic propagation', 'hybrid system', 'discriminative training', 'represented word form', 'speech problems', 'speech prediction', 'automatic speech recognition', 'automatic speech problems', 'speech problem', 'set', 'set optimization', 'computer', 'system integration', 'system adaptation', 'set selection', 'stochastic iterative dynamic programming', 'context bodies', 'example bodies', 'optimal experiment design', 'experiment design', 'neural network exploration', 'can controller', 'mushroom we', 'glomeruli', 'including learned experiments', 'mushroom body', 'including learned maintaining', 'including learned learning', 'including learned neurons', 'novel neuron', 'incomplete data', 'temporal constraints', 'both 2mm models', 'be algorithm', 'be formula', 'be the head', 'be . thought', 'be ii', 'fovea', 'contour extraction', 'amacrine cell', 'dopamine', 'contour in cell', 'at and estimates', 'contour in the is auditory system', 'contour in the is auditory', 'contour in the is auditory channel', 'one the early vision', 'optic', 'best of attention', 'each a', 'between a', 'between cell', 'combining endoscopy', 'between', 'visual network', 'distance', 'signature verification', '\"siamese\" neural network', 'pen - input tablet', 'but olfactory system', 'combination of classifiers', 'as olfactory system', 'dynamics of central pattern generators', 'nearest neighbor classifier', 'nearest neighbour classifier', 'that s game', 'that a plan', 'convex set projection', 'multi <unk> layer perceptron', 'metrics', 'multi <unk> layer perceptrons', 'sigmoid', 'receptive field topology', 'routing', 'q - routing', 'required demonstrate', 'required cell', 'restricted routing', 'restricted shortest path', 'observable behavior', 'reactively', 'traffic', 'motion ambiguities', 'coupled markov random fields', 'coupled markov random field', 'confidence measures', 'markov observation', 'confidence measure', 'markov observation models', 'momentum', 'stochastic search', 'adaptive momentum', 'stochastic images', 'from issue', 'asymptotic normality', 'stochastic gradient algorithms', 'to assertions', 'are assertions', 'large margin', 'adaptive control', 'spiking neurons', 'adaptive systems', 'nonlinear systems', 'to fault diagnosis', 'understanding', 'nonlinear', 'adaptive', 'large', 'developed', 'quadratic', 'projection', 'spin - glasses', 'self - organization', '<s> dynamics', 'necessitates - probe', 'necessitates - structural', 'necessitates - structural dynamics', 'probe models', 'identified parameters', 'larger network', 'learnt convergence', 'eye movements', 'eye tracking', 'superior dynamical systems', 'ganglion cells', 'articulator', 'd . 2 . 2 <unk> d vision', 'd . 2 . 2 . 2 <unk> d vision', 'computer graphics', 'to vision', 'to computer vision', '5', 'computational', 'bayesian modeling', 'neural signals', 'poisson distribution', 'signal decomposition', 'neural signal processing', 'nervous system', 'nervous system modeling', 'signal similar', 'nervous system imaging', 'signal', 'analog vlsi circuits', 'on of line learning', '2 to d circuits', 'self to self the circuits', 'self to self the neural networks', 'self to self the analog circuits', 'self a analog circuits', 'self . <unk>', 'self . self of analog circuits', 'self . analog circuits', 'self . analog vlsi circuits', 'address block location', 'neural net system', 'this spike detection', 'spike pattern recognition', 'to are largely silicon', 'space clusters', 'string matching', 'from are largely silicon', 'does presentations', 'string kernel', 'presentations chain', 'factor', 'change', 'neuromodulation', 'functional electrical stimulation', 'neural control', 'phase <unk> sensitive neurons', 'dynamic reposing', 'tangent distance', 'drug activity prediction', 'handwritten character recognition', '\"optimal', 'barn owl', 'world', 'learning\"', 'processing', 'all theory', 'novel algorithms', 'akaike information criterion', 'microstimulation', 'limited capacity', 'limited memory', 'input potential', 'input models', 'given minimize sum', 'employ', 'given ) noise', ') noise', '<unk> noise', 'hodgkin - huxley neuron model', 'next a . classifier', 'next generation neural network', 'next a . <unk> eliminates algorithm', 'eliminates algorithm', 'next generation network', 'next a . <unk>', 'eliminates the development', 'difficult a approach', 'difficult a and ations', 'difficult a and exponent region', 'difficult a and exponent back the to activation', 'difficult a and exponent back the to activation curve', 'difficult a and exponent back the to activation ) point', 'difficult a and exponent back the to activation ) through components', 'difficult a exist trajectories an associated here the and speech most powerful model', 'difficult a exist trajectories an associated here the and speech most essential', 'difficult a exist trajectories an associated here the and speech most powerful a components', 'backpropagation algorithm', 'dynamic neuron', 'dynamic neuron model', 'dynamic a dynamic model', 'dynamic a dynamic neuron', 'exemplar selection', 'time - series prediction', 'mackey - paper diagnosis', 'time signal represent', 'mackey', 'solvable models', 'ordinary differential equations', 'artificial distortion', 'ordinary built <unk> in tests', 'a posteriori estimates', 'ordinary built <unk> in self <unk> training', 'ordinary built <unk> in self <unk> test', 'training algorithm', 'ordinary built <unk> in self <unk> adapting systems', 'error backpropagation', '<s> training', 'computational elements', 'nonlinear dynamics', 'computational mechanics', 'trajectory optimization', 'local models', 'locally weighted regression', 'robot learning', 'complex juggling juggling', 'low shift', 'model quality', '1 shift', 'speedup', 'distance makes', 'juggling', 'locally weighted placed executed a 1', 'model phoneme', 'model its', 'statistical', 'architectural', 'model particular solutions', 'model particular', 'locally weighted placed executed a juggling', 'model particular size', 'model particular computation', 'model has', 'model not supplied', 'relevant components', 'nonlinear dimensionality reduction', 'relevant regions', 'out series', 'from fom', 'evaluate computation', 'defined networks', 'neighborhood system', 'efficient computation', 'i based finite state machine', 'higher in in complex systems', 'gt', 'higher in in in in in', 'gt in when', 'sound localization', 'owl', 'sound localization by analogy', 'sound localization system', 'oo model', ') model the oo model', 'computational neuroscience', 'oo', 'growing cell structures', 'radial robust regression', 'minimizing', 'figure of merit training', 'far in ed analysis', 'wearable computer vision', 'handwriting recognition', 'two in cell presents', 'handwriting analysis', 'hidden markov model', 'gaussian mixture', 'our weighted em algorithm', 'gaussian', 'minimize sum', 'minimize', 'brainstem circuit', 'feedforward network', 'auditory nerve fiber', 'realistically constrained hearing', 'auditory employed', 'directional hearing', 'auditory nerve been through modeling', ', systems biology', 'realistically constrained as by dynamics', 'realistically constrained feedforward network', 'the dynamics of problem', 'the dynamics of work', 'on - line handwriting recognition', 'explicitly .', 'on - line cursive handwriting', '1 . trained neural network', 'on . .', 'on . . of handwriting recognition', 'on . . of handwriting', 'on . . of of handwriting', 'short term memory', 'short term memories', 'connectionism', 'of memory', 'of memory analyze', ') <unk> model', ') <unk> <unk>', ') <unk> models', 'back - propagation algorithm', 'neural network regression', 'be <s> outlier', 'be <s> weights', 'problem rate', 'image understanding', 'image interpretation', '3d reconstruction', 'wavelet transforms', 'wavelet packets', 'wavelet transform', 'principal components analysis', 'wavelet', 'radial', 'principal', 'minimum description length', 'factorial codes', 'minimum description length principle', 'autoencoder networks', 'recognition weights', 'connectionist networks', 'catastrophic interference', 'systematicity', 'prepyriform cortex', 'sigmoid nonlinearity', 'sigmoid curve', 'slope', 'high enough', 'normalized wave input', 'sigmoid to signal moved', 'sigmoid to noise ratio', 'high achieved being', 'sigmoid to signal ratio', 'high enough position', 'stochastic learning', 'role', 'play', 'signaling', 'technologies', 'analog', 'hebbian learning', 'hebb <unk> learning', 'synaptic plasticity', 'synaptic minimize pixel', 'hebb learning rule', 'model <unk> based learning', 'synaptic minimize pixel <unk> temporal cortex', 'noise sensitivity', 'fitting', 'more networks', 'training of of learning', 'training learning', 'training of cases', 'agnostic', 'constructive learning', 'line detection', 'a units', 'well sequence from method', 'training classifiers', 'asymptotic values', 'rate of data', 'classifier design and evaluation', 'brain networks', 'linear models', 'continuous time neural networks', 'asynchronous dynamics', 'global asymptotic stability', 'limit theorems', 'asynchronous model', 'eg strategy', 'in <s> asymptotic behavior', 'demonstrated distributed processing', 'in <s> combines global neurons', 'showed effect', 'fault false alarm rate', 'bayes accurately', 'robot safety', 'robot vision', 'robot resolution', 'relaxation method', 'multiple <unk> input multiple <unk> output systems', 'relaxation labeling', 'relaxation methods', 'multiple <unk> input multiple <unk> output detection', 'relaxation labelling', 'surveillance', 'understand redundancy', 'understand <unk> made systems', 'surveillance systems', 'understand <unk> inspired algorithms', 'understand <unk> machine systems', 'understand <unk> inspired systems', 'dynamical behavior', 'pseudo - orbits', 'memory - based learning', 'batch currently', 'scripts', 'discrete learning', 'local efficient', 'deficient data', 'bayesian self - organization', 'binocular stereo', 'early vision', 'leads', 'receptors', 'seen', 'sunspots time', 'real complexity', 'it principle', 'here as a tree', 'here as a', 'billion error probability', 'here as a experiments', 'here as a continuous', 'here continuous', 'it large scale', 'here as a posterior', 'distributed associative memory', 'conformal mapping', 'object actions', 'on hidden surface model', 'object 1 norm', 'object implementations', 'on hidden surface', 'neurocomputer', 'that <unk> <unk>', 'object networks', 'are 2 algorithms', 'all is', 'object labels', 'enough learning', 'billion', 'ed', 'conformal automatically', 'development environment', 'billion memory', 'analog and examples', 'analog and digital vlsi', 'texture synthesis', ', , generalization', ', half standard', ', , , combined synthesis', 'texture mapping', 'analog and number generator', ', , combined synthesis', 'expected network complexity', 'generalization error', 'regularizei unsupervised learning', 'bias/variance unsupervised learning', 'hidden units', 'schemes of construct', 'adaptive estimation', 'pattern', 'analogies', 'schemes of rate', 'pattern of instead', 'analogies on machine', 'gibbs paper', ', case studies', 'bias/variance', 'makes computations', 'computations on probability', 'modified bessel functions', 'modified bessel function', 'bias/variance on insufficient main main network', 'regularization network', 'makes computations on insufficient variables', 'makes computations on insufficient main network', 'bias/variance on insufficient main main extraction', 'principal components character', 'context dependency', 'factoring', 'functional models', 'computational models', 'combined learning', 'changes in for problems', 'changes in localization', 'changes in for classification', 'changes in for by we', 'information modeling', 'an half finite element', 'an half finite element method', 'an half finite elements', 'an half assignment', 'an half finite element analysis', 'surface learning', 'lipreading', 'lip reading', 'visual tracking', 'lms', 'optimal', 'hoo', 'optimality criteria', 'optimal estimator', 'hoo optimality', 'to noise ratio', 'to noise', 'measured signals', 'measured value', 'problem', 'feedforward networks', \"tltt ' problem\", 'machine attentional mechanism', 'neurobiology of learners', 'machine complexity', \"tltt ' s i\", 'crf', 'judge is learning', \"tltt ' s i and manipulated manipulated\", 'input weights output learning', 'input weights output model', 'input and input and output control', 'input and output and input control', 'input and input weights output learning', 'crosstalk', 'grammatical inference', 'ordinary differential equation', 'percolation', 'internal noise', 'internal crosstalk', 'random transitions', 'state automata', 'reber', 'state automaton', 'reber grammar', 'transition state', 'literature metrology', 'literature <unk> based classification', 'knot a', 'literature <unk> based event detection', 'encoded vector quantization', 'expensive optimization', 'expensive black box', 'expensive black box model', 'expensive function', 'classes defined function', 'classes defined functions', 'classes of movement', 'often needed', 'classes of i', 'classes of i data', 'efficiency index', 'enough to model data', 'enough random', 'hidden markov chain monte carlo', 'computationally efficient', 'cases of interlocking layers', 'enough to modified can into interlocking layers', 'cases of minimization', 'computationally efficient algorithms', 'formation rate', 'formation', 'minimization', 'handwritten word recognition', 'spatial representation', 'convolutional methods', 'oo classification', 'hidden problems', 'oo that slow training', 'theory of machine learning', 'oo that for internal models', 'to a as topic performs', 'is of handwriting', 'to a . minimizing for method', 'to a are for in method', 'information processing', 'computer model', 'pyramidal cell model', 'pyramidal cell', 'pyramidal cells', 'cable simulation', 'computer simulation', 'model simulation', 'computer chess', 'games of go', 'temporal both', 'temporal difference form', 'candidates', 'becker <unk> <unk>', 'a <unk> <unk>', 'as <unk> <unk>', 'be <unk> <unk>', 'learning classes', 'i <unk>', 'motor learning rate', 'motor', 'i <unk> convergence', 'evaluation', 'to <unk>', 'position error', 'to <unk> convergence', 'finite combined automata', 'equation an state complexity', 'm - state machines', 'from and vector implementations', 'in curve implementations', 'finite a posteriori error bounds', 'freeman k to sat', 'finite a posteriori error bound', 'high performance', 'answer set programming', 'this it bound', 'it bound', 'case bound', 'it to field', 'it it attention', 'behaving e', 'enhanced recurrent locations', 'or', 'modified bumptree neural network', 'neural network classifier', 'bumptree neural network', 'geometric are classifier', 'modified bumptree neural network classifier', 'geometric are it', 'neural network convolution', 'time units', 'pca', 'non - linear dimension reduction', 'auto - associative networks', 'auto axon pruning', 'basic principal component analysis', 'cursive scripts', 'stochastic automata', 'cursive script', 'online handwriting recognition', 'adaptive probabilistic acyclic automata', 'is shown', 'input shown', 'auditory scene analysis', 'connectionist models', 'auditory scene . attention', 'sound', 'a <unk> models', 'internal models', 'far <unk> field', 'between <unk>', 'in <unk> the <unk>', 'between <unk> and', 'cortex', 'between <unk> states', 'between <unk> and top', 'between states', 'in <unk> and <unk>', 'platform', 'parallel', 'that systems', 'board processing', 'derivatives', 'distance learning', 'large <unk> scale', 'classes', 'when <unk> or', 'probabilistic reasoning', 'b <unk>', 'programming', 'versions', '6', 'acute recurrent neural networks', 'b language', 'b <unk> splines', 'animals', 'probabilistic neural networks', 'b <unk> rep', 'neurobiology', 'attentional focus', 'pattern association problems', 'generalized distributions', 'pattern association show', 'backpropagation . finding algorithms', '6 method', 'be for all', 'based . attractive set', 'forward - inverse - understand model', 'be . akaike information criterion', 'interest point considered part', 'interest point extraction', 'character recognition', 'training training boltzmann machine', 'training boltzmann machine', 'amplitude estimation', 'amplitude equation', 'view - based representation', 'gesture recognition', 'view - based distributed representation', 'hand gesture recognition', 'logic synthesis', 'computer arithmetic', 'circuit optimisation', 'circuit complexity', 'two - dimensional circuits', 'between ( two ( two', 'vlsi design', 'two - stage circuits', 'logic conversion', 'between ( two ( many', 'between ( two ( batch', 'software', 'float', 'natural', 'natural phenomena', 'circuits', 'natural computing', 'scale invariance', 'remote sensing', 'statistical image statistics', 'featuring', 'hill climbing', '\"royal function', 'analog neural networks', 'piecewise polynomial activation functions', 'analog neural nets', 'piecewise polynomial learning', 'domain - specific distance measure', 'databases', 'nested objective function', 'do . c . clustering', 'expectation - given . is', 'initially , 1 <unk> dimensional object recognition', 'do . c . c . c . c . c . c . c . c . c . c . c . c . d . c .', 'do . c . c . c . c . c . c . c . c . c . c . c . c . c . c .', 'do . c . c . c . c . c . c . c . c . c . c . c . c . c . d .', 'do . c . c . c . c . c . c . c . c . c . c . c . d . c . c .', 'do . c . c . c . c . e . c . c . c . c . c . c . c . c . c .', 'do . c . c . c . c . c . c . c . c . c . c . d . c . c . c .', 'do . c . c . c . c . e . c . c . c . c . c . c . c . c . d .', 'do . c . c . c . c . e . c . c . c . c . c . c . c . d . c .', 'do . c . c . c . c . c . c . c . c . c . c . c . d . c . d .', 'do . c . c . c . c . c . c . c . c . c . c . c . c . c . 1 <unk>', 'multilayer networks', 'committee machines', 'discontinuous generalization', 'committee machine', 'is is', 'network topology', 'found', 'found theory', 'context figure', 'found in the corpus', 'found in probability', 'found in the limit', 'distinct elements', 'found in the first proposed i', 'axon', 'similarity', 'conduction failure', 'circuit model', 'firing threshold', 'compositional hierarchy', 'hierarchical models', 'part - whole relationships: reinforcement learning', 'structural learning', 'part - whole relationships: reinforcement', 'one paper recognition', \"' s shown model\", 'is metrics', 'as metrics', 'one paper lattice model', 'be ones property', 'at nonlinear belgian', 'at nonlinear properties', 'at nonlinear sketch', 'bump networks', 'at nonlinear dimensionality reduction', 'hoeffding races', 'function basins', 'classification and function basins', 'hoeffding races states', 'fixed points', 'is units', 'layer units', 'measure units', 'hinton units', 'that g units', 'shape - from - texture', 'shape complexity', 'to and image processing', 'frequency analysis', 'frequency', 'visual system', 'autonomous agents', 'a incorporates learning', 'into learning', 'a posteriori error analysis', 'computational for independent learning', 'computational in tangent', 'computational in neuroscience', 'layer <unk> learning', 'back <unk> propagation', 'layer <unk> based learning', 'exhibit', 'layer 4', 'layer <unk> sensitive learning', 'population coding', 'minimum description length point', 'modeled structure', 'population code', 'but order coding', 'but order statistics', 'these redundancy', 'have ) vowel', 'dolce', 'recurrent neural nets', 'finite - state large can machine', 'transition points', 'transition point dynamic programming', 'adaptive optimal control', 'transition point dynamic networks', 'etc', 'spline approximation', 'focus point', 'structured machine learning', \"soft ' classification\", 'be the analog to machine learning', 'distances in the roc curve', 'order restricted memories', 'distances in the first kind', 'distances in the centered perspective', 'that efficient sample selection', 'no free lunch', ') <unk>', ') other', 'hopf algebras', 'order restricted observe information', 'roc receiver operating characteristic', ') other probability', 'roc receiver operating characteristic curve', ') based on markov chains', ') based on markov chain monte carlo', ') based on markov chain monte carlo methods', 'inversion capable of markov chains', 'inversion capable of markov chain', 'inversion of markov chain', ') ) other', 'inversion capable of', 'inversion of markov chains', 'inversion of markov chain monte', 'layered networks', 'multicategory classification', 'delta network', 'delta rule', 'multicategory the weights', 'multicategory', 'korean', 'layered network', 'multicategory the approach', 'paper ) algorithm', 'classification algorithm', 'the delta', 'intractable likelihoods', \"moore ' s law\", 'the minimum spanning tree', 'of neural networks', \"moore ' s method\", 'neuronal dynamics', 'spike real coding', 'globally connected spiking neurons', 'spike response order coding', 'spike response model', 'spike real - time coding', 'birdsong', 'spike time coding', 'spike percept', 'dynamic synapses', 'spike coding', 'hybrid connectionistfhmm systems', 'sum data', 'myriad filter', 'myriad filters', 'spatial', 'relationships', 'built <unk> in self <unk> organization', 'map', 'network architecture', 'backgammon', 'built <unk> in self <unk> organizing map', 'diamond', 'binary diamond', 'interconnected neural network', 'bottlenecks', 'neural function', 'back network', 'correlation functions', 'critical slowing down', 'correlation bands', 'critical slowing depends', 'correlation function', 'multi - layer classification', 'electroencephalogram', 'multi <s> indicate compress - and to problems', 'multi <s> indicate compress and forward', 'electroencephalogram algorithm', 'class - accounts', 'to reinforcement learning', 'class of of decomposition', 'postal address block location', 'convolutional locator network', 'postal address block', 'of tutor', 'convolutional', 'convolutional locator the of mean', 'of reciprocal network', 'of replicated data', 'gamma learning', 'forward network', 'address point', 'models of algorithms', 'differentiating hebbian network', 'differentiating hebbian learning', 'input codes', 'developing regions', 'developing countries', 'to codes on words', 'to be vq', 'a be be useful rules', 'vq', 'k', 'boolean networks', 'time delay neural networks', 'visual preprocessing', 'and bump maps', 'cross of correlation', '( bump maps', '( invariant features', 'network of gaussian distributions', 'vowel localization', 'vowel detection', 'vowel tracing', 'a posterior distribution', 'a posterior probability distribution', 'visual selectivity', 'to signal latter to noise ratio', 'to signal to noise ratio', 'to signal ratio', 'a posterior distribution of', 'a posterior', 'gained', 'audition', 'music and audition', 'brain electric field', 'spontaneous series', 'information - features', 'team attempt', 'is hidden rescheduling', 'modeled . exhibit', 'is hidden markov model', '. most observations', 'recognize . phase', '<s> . most observations', 'a most observations', 'further lies', '. most powerful test', 'methods for methods', 'further further lies', 'further training examples', 'lraam', 'raam', 'hopfield network', 'raam network', 'labeling policy', 'the physically <unk> based model', 'constraint <unk> based reasoning', 'hopfield', 'machine <unk> learning', 'baseline', 'constraint <unk> based coding', 'be <unk> example', 'modified neural tree network', 'discriminant representation', 'cortical learning', 'neural tree network', 'discriminant learning', 'near are classifier', 'nonlinear pca', 'gaze tracking', 'back classes', 'precise', 'digitize', 'parallel computation', 'currently routing', 'described instruction', 'topologies', 'higher learning', 'to analog', 'to to analog', 'patients', 'cross and', 'objects', 'cross and analog', 'cross', 'cross and image', 'to cross and analog', 'cross and amplitude', 'object localization', 'delta classification', 'delta operator', 'vocabularies', 'coarse - to - guarantee points', 'implemented biometrics', 'coarse - to - guarantee', 'number of limited characterization', 'accuracy analysis', 'exemplars', 'many and one ) systems', 'two - dimensional object localization', 'mean different for approach', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '5 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 3 , 2 , 2 , 2', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , d , 2 , 2 , 2', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 3', '5 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1', 'fuzzy systems', 'surface interpolation', 'fuzzy system', 'ocular dominance', 'infant', 'monkey', 'macaque', 'orientation preference', 'cortical maps', 'infant development', 'iterative algorithms', 'vector signal processing', 'filter', 'vector autoregressive models', 'vector autoregressive model', 'primate', 'toeplitz matrix', 'fast solvers', 'high', 'meter reading', 'associative memory design', 'high - order neural networks', 'association rules', 'association rule', 'of association', 'of associative hardware', 'map learning', 'continuous domains', 'learning from continuous domains', 'cognitive robotics', 'learning from analysis', 'exploration vs . exploitation', 'learning from is', 'learning from examples', 'learning from learning', 'developmental robotics', 'connected components', 'learning from for or', 'essential dynamics', 'by imitation', 'by analogy', 'better than worst <unk> case learning', 'better than worst <unk> case execution', 'crew pool', 'rescheduling', 'belgian airline', 'belgian airline pool', 'belgian new coordinates', 'here ntn', 'crew scheduling', 'here ( we', 'belgian new by', 'chaotic systems', 'radial b <unk> basis functions', 'related groups', 'string processing', 'distance transforms', 'text mining', 'text processing', 'silicon retina', 'illumination invariance', 'illumination an s probability', 'attractions', 'are discrete wavelet transform', 'arbitrary bandpass', 'sensory fusion', 'sensory processing', 'sensory image processing', 'are discrete wavelet bases', 'mean - field theory', 'way learning', 'of pca', 'video processing', 'topology', 'head learning', 'modular networks', 'multiterminal mean the adaptive resonance theory', 'best a priori algorithm', 'any - analogous correlation', 'order - restricted network', 'network - most - analogous', 'network - most correlation', 'large in in in', 'as in in in', 'network - most - analogous correlation', 'correlations', 'bands', 'that in', 'an recorded', 'network - most analogous correlation', 'orientation in', 'large - scale more', 'blocking', 'stochastic perceptron', 'k - networks', 'blocking probability', 'k are', 'pt of measure', 'necessary optimality', 'pt of convergence', 'pt of norms', 'o of activities', 'on of based learning', 'speaker learning', 'on of based perceptron', 'image interpolation', 'nonlinear image interpolation', 'nonlinear interpolation', 'model - based vision', \"model ' s\", 'into disorder detection', 'much disorder', \"lip ' s patch activation\", \"model ' insights\", 'shows data', 'e - set', 'e - and we', 'attribute', 'e of competition', 'cross , language', 'cross language', 'alignment', '( e of competition', 'spatial selective attention', 'saliency problem', 'spatial to provides of neural networks', 'outlines', 'in ) task', 'in ) algorithm', 'multiple sensors', 'in ) actions', 'in ) policies', 'in ) model', 'in ) systems', 'cross connected networks', 'degree distribution', 'from <unk> network analysis', 'variance of characters', 'knowledge representation', 'at <unk> free networks', 'mean shape', 'an <unk> model', 'based model', 'an <unk> network', 'synaptic noise', 'form error', 'synaptic noise noise classification', 'e description', 'e description language', 'continuous neural networks', 'race conditions', 'continuous neural network', 'race neural networks', 'race neural network', 'rates of convergence', 's learning', \"' s of classifiers\", \"' s of classifier\", 'is ) model', \"' s of statistic\", 'ensembles of classifiers', 'parietal cortex', 'basis function', 'bifurcation diagram', 'basis functions', 'egocentric coordinates', 'decay function', 'between - examples', 'between - function', ') their model', 'analytically their all', 'analytically their', 'analytically their connectivity', 'analytically cases', 'stereo', 'spatial approach', 'functioning', 'many cases', 'approximate action', 'actions', 'reinforcement', 'spatial &', 'nonlinear system identification', \"1 ' s to and 2 intrusive models\", 'analog signal processing', 'multiterminal source coding', 'multiterminal source control', \"1 ' s to and 2 intrusive ' s , 2 , 2 networks ,\", '1 and 2 , 2 models ,', 'accept models', '1 and 2 , 2 models , 2 ( 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2', '1 and 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 2 ,', '1 and 2 , 2 , 2 , 1 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 and 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 ,', '1 and 2 , 2 models , 2 ( 2 , 2 , 2 , 2 , 2 , 2 , 1 , 2', '1 and 2 , 2 models , 2 ( 2 , 2 , 2 , 2 , 2 , 2 , 2 , 1', '1 and 2 , 2 models , 2 ( 2 , 2 , 2 , 2 , 2 , 2 , 2 , 3', '1 and 2 , 2 , 2 , 1 , 2 , 2 , 2 , 1 , 2 , 2 , 1 ,', '1 and 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 , 2 a', '1 and 2 , 2 models , 2 ( 2 , 2 , 2 , 2 , 2 , 2 , 2 a priori', 'neural vlsi', 'uran', 'neural hardware', 'universally reconstructible', 'em', 'partially defined action', 'skills', 'partially defined action policies', 'skills algorithm', 'partially changes', 'skills learning', 'skills algorithm based learning', 'song learning', 'resistive or included iterative learning', 'resistive to included iterative learning', 'song learning system', 'workshop', 'tutor vocalization', 'only memory', 'semi - markov decision processes', 'motion control', 'image motion estimation', 'motion estimation', 'markov of processes', 'markov a posteriori estimation', 'not animals', 'not although is', 'not development', 'is the direct marketing', 'with better than computable functions', 'choice models', 'choice equations', 'foot the', 'user the', 'markov to', 'user arrival', 'connection machine', 'simd neurons', 'parallel implementation', 'on <unk> line simulation', 'hopfield model', ', applications', ', performance evaluation', ', performance analysis', 'sense of direction', 'attractor dynamics', 'neuron model', 'morphogenesis', 'lateral geniculate nucleus', 'lamination scene', 'improves non', 'lamination selective', 'olfaction', 'present non', 'glove - taikii', 'adaptive interfaces', 'hand gesture', 'although agent', 'although rl', 'growing neural gas', 'self of learning', 'self - organizing maps', 'go pattern learning', 'do pattern learning', 'self non 1 learning', 'content addressable memory', 'self non 1 recurrent networks', 'self non 1 recurrent neural networks', 'moving units', 'moving object recognition', 'position to also', 'neurological function', 'predicts constraints', 'neurological image', 'model - fitting', 'significantly confined', 'moving', 'may a bands', 'process control', 'not gate', 'cooperative vector quantizer', 'zemel', 'expectation - maximization architecture', 'cooperative vector retina', 'zemel architecture manipulated use', 'focus', 'data found', 'determines data', 'zemel architecture manipulated algorithms', 'determines visual memory', 'focus algorithms', 'hinton how', 'extracting function', 'determines visual data', 'local error bars', 'several variables', 'shows', 'several theoretical analysis', 'syllable ) approach', 'object perceptrons', 'object activation', 'errors - in - variables', 'trackers', ', than layer as changing', 'predictive coding', 'text compression', 'direction selectivity', 'orientation and direction selectivity', 'hysteretic mode', 'expansive neuron', 'direction control', 'expansive maps', 'flat minima', 'use of compositional rule', 'their search', 'stock market prediction', 'locator rule', ') a is of et models', 'be a inducing machine', 'cross neurons', 'if neurons', 'cross complexity', 'process maps', 'be a inducing', 'fields of learning', 'segments', 'hence classification', 'can in representation', 'higher codes', 'network selection', 'network maps', 'higher low bump search', 'higher low bump maps', 'segments of at bump', 'network classification', 'feedback network', 'equilibrium state', 'stochastic descent', 'feedback networks', 'possible has', 'annealing', 'mixtures', 'matrix algebra', 'is parallel systems', 'phasor measurement unit', 'associative memory present long the range of a neural networks', 'and 1 the <unk> vectors vectors . 1', 'optimal experiment based are data', 'optimal experiment based algorithms', 'bias and that this learning', 'that learning simpler algorithm', 'bias and that this this moments', 'optimal experiment based are be forward as topic', 'optimal experiment based are be forward as a learning', 'bias and that this not supplied', 'imposed constraints', 'optimal experiment based are be forward as a learning student', 'hme', 'optimal experiment based are be forward as a , 1 cases: optimality', 'optimal experiment based are be forward as a , 1 <unk> optimality', 'kolmogorov complexity', 'optimal experiment based are be forward as a , 1 , 1 our', 'optimal experiment based are be forward as a , 1 , 1 <unk> optimality', 'optimal experiment based are be forward as a , 1 , 1 , 1 <unk> optimality', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 handwriting', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 handwriting', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 that 1 , 1 , 1 , 1 , 1', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 3', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 2 , 1 , 1 , 1', 'optimal experiment based are be forward as a , 1 , 1 , 1 , 2 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1 , 1', 'decorrelation', 'statistical decorrelation', 'high - order statistics', 'nonlinear decorrelation', 'principal components', 'cumulant tensors', 'redundancy reduction', 'reactive synaptic regeneration', 'schizophrenia', 'temporal lobe projections', 'coronary cortex', 'auxiliary themes', 'using temporal', 'using temporal a', 'using the approaches', 'silicon neuron', 'cmos process', 'pulse width ai', 'silicon axon', 'pulse width restoration', 'pulse motor', 'analog circuits', 'immediate recall', 'short - term memory', 'articulatory loop', \"articulatory loop ' thought\", 'combine harvester', 'gating networks', 'error rate estimation', 'gaussian mixtures', 'winner - take to . are', 'in in situ hybridization', 'in in distance', 'by . a classifier', 'effects of', 'discrete cosine transform', 'employs', 'as a <unk> optimality', 'complications', 'minimum message length', 'writer identification', 'minimum message', 'maximum a', 'dynamic writing information', 'multi - trains', 'on - line cursive handwriting recognition', 'piecewise recurrent neural networks', 'piecewise recurrent neural network', 'writer', 'about of input 2 . 5 . 8', 'writer model', 'hidden markov', 'recurrent neural', 'fixed .', 'fixed point', 'forward dynamic models', 'used for recognition', 'experimental results', 'motor system', 'forward can', 'propose domains', 'procedure of connected codes', 'procedure of connected', 'procedure of connected components', 'mortality', 'trained feedback', 'instantiated memory', \"trained 5 ' s criterion\", 'improved recall', 'original missile', 'coarse - coded symbol memory', 'points of memory', 'neural exploits', 'points the capabilities', 'is input model', 'is input models', 'is input the projections', 'is input the and projections', 'is input the the projections', 'degenerating model', 'neural network a function', 'neurons to , individual decision of and regression trees', 'conventional attractor neural networks', 'conventional attractor neural network', 'emergence of a the 1', 'emergence of a associated vectors', 'emergence of a the a the 1', 'decaying item memories from neurons', 'eyes , allowed', 'eyes , but imposes', 'eyes , but allowed', 'eyes , but , allowed', 'eyes , but , but interacts controlled', 'eyes , but interacts , but interacts , but', 'eyes , but interacts , but interacts controlled', 'eyes', 'eyes , but , but interacts , but interacts', 'eyes , but interacts , but interacts , allowed', 'auto - associative network', 'and object recognition', '3d wire - investigate cortical map', 'and object classification', '3d object classification', 'neuronal networks', 'and object cortical maps', 'perturbative learning', 'differential game', 'advantage updating', 'missile', 'q - and item adaptation', 'q - and are learning', 'q - and item an system', 'q - and item an writing', 'q - and item an vocabulary', 'consistent learning', 'vc - dimension', 'forwards murphy', 'in murphy', 'have and and policy', 'mutual of training algorithm', 'by murphy', 'mutual input', 'and policy', 'mutual of training in', 'input and and policy', 'receptive fields', 'fixed point attractor', 'linsker - type hebbian learning', 'in silico model', 'in silico prediction', 'in silico models', 'in silico modeling', 'in silico neuron', 'in silico predictions', 'visuomotor coordination', 'sizes of coded recently used', 'visuomotor transformation', 'casting neural networks', 'symbolic rule extraction', 'rule extraction', 'weight templates', 'weight space', 'parameterized regions', 'weight unsupervised learning', 'parameterized saturated representations', 'parameterized saturated representation', 'model approximation', 'ann', 'symbol grounding', 'nonlinear learning machine', 'central limit theorem', 'generalization algorithm', 'it <unk> minimization algorithm', 'paper <unk> glass model', 'categorizing <unk>', 'bee colony algorithm', 'central subspace method', 'bee colony optimization', 'learning machine accuracy', 'data quality', 'e <unk> service', 'learning machine', 'output', 'account', 'diffusion', 'stochastic models', 'markov consequences', 'markovian models', 'and delivering time', 'of cognitive models', 'of cognitive machines', 'robot architecture', 'learning from demonstration', 'robot arm', 'multilayer perceptrons', 'parallel vlsi', 'sequence processing', 'pca pyramid', 'within action coding', 'that action coding', 'wavelet scaling', 'spatial filters', 'saccadic eye movements', 'saccadic eye movement', 'adaptive motor maps', 'adaptive motor map', 'adaptive motor map map', 'in achieved vision', 'distributed representations', 'model extraction', 'backpropagation - provide', 'validity', 'shape extraction', 'knowledge extraction', 'knowledge examined', 'invariant learning', 'error correction', 'advantage', 'ocr', 'pseudo - optic', 'colliculus', 'membrane computing', 'exploits', 'letters', 'model input', 'however has', 'interval analysis', 'deficiency detection', 'associated neural network', 'associated neural networks', 'trajectory generation', 'trajectory generator', 'answer set', 'also of converge', 'also of also', 'answer set recognition', 'also of also of acquisition', 'acquisition', 'also of movement', 'also of tdnn', 'acquisition of neural network', 'tdnn', 'auditory system', 'inferior colliculus', 'auditory cortex', 'probability estimation', 'hierarchical mixtures of experts', 'continuous networks', 'speech method', 'probability density estimation', 'networks of networks', 'networks of stochastic processes', 'networks of stochastic models', 'networks of of theory', 'networks of stochastic systems', 'directional selectivity', 'retinal directionally selective', 'retinal waves', 'mathematical model', 'mathematical models', 'compartmental model', 'retinal directionally selective cell', 'by <unk> cell', 'adaptive algorithm', 'hints', 'learning from hints', 'financial applications', 'gradient - based neural networks', 'state identification', 'instance - based learning', 'state neural networks', 'executed learning', 'and network learning', 'and activation functions', 'that . cognitive learning', 'is the semantic technologies', 'is unsupervised learning', 'is the semantic web', 'is the future', 'is the hinton', 'is the improves system', 'is the semantic knowledge', 'point matching', 'providing environment', 'graph sample path', '. its', 'arm its', 'or in motion', 'preknowledge: clustering', 'or in natural gradient', 'games on manifolds', 'dynamic modelling', 'chaotic', 'chaotic time series', 'network dynamics', 'in <unk> network processing', 'in <unk> network analysis', 'in <unk> network data', 'pulse <unk> coupled oscillators', 'honey bee chemosensory', 'second messenger', 'd permutations', 'decompression', 'first in in shear', '<unk> architecture', 'bitmap', '<unk> has', '<unk> information theory', 'self - organizing feature map', 'theory of recurrent neural networks', 'resulting in general mechanisms', 'speech perception', 'remote - sensing', 'periodic variables', 'periodic data', 'memories', 'information fusion', 'conditional probability density estimation', 'long in in present value', 'a <unk> estimation', 'must experimental data', 'must experimental efficiency', 'must experimental design', 'must experimental measurements', 'must experimental <unk>', 'k - model', 'if its', 'by its', 'pulse - coded neuronal systems', 'spike - timing - dependent plasticity', 'with with with from this information', 'spike a timing analysis', 'spike - timing - dependent learning', 'spike a timing a model', 'output a posteriori probability estimation', 'output a posteriori error estimation', 'output a posteriori error analysis', 'output a priori error bounds', 'output a', 'output error', ', over', 'that or', 'output a priori error estimation', 'deterministic annealing', 'markovian arrival process', 'statistical mechanics analogy', 'maximum constrains', 'two <unk> phase flow', 'dementia', 'stroke', \"alzheimer ' s disease\", 'into system', 'time - series analysis', 'include memory', 'time we', 'high chemosensory networks', 'time rules', 'time system', 'self system', 'oscillator', 'selective gating', 'relaxation oscillator', 'neural oscillator', 'figure/ground', 'oscillator networks', 'synchrony segregation', 'useful', 'electrotonus', 'voltage attenuation', 'voltage', 'spatial distribution', 'distance metric', 'dynamic decay adjustment', 'p - rce algorithm', 'self - calibration', 'analog vector quantization', 'double - real the the demonstration', 'net to also immediate investigate', 'double - real the the on of fast partially', 'long neurons', 'double - poly information', 'we of demonstration', 'access time', 'machines encoding', 'double - real the the on of fast but of acquisition', 'double - real the the on of fast but also', 'operations on the learned weights', 'system on chip', 'operations on the learned these biological properties', 'involve of are formulated', 'coordinate transformation', 'visual feedback', 'remappings', 'motor map', 'spatial resolution', 'spatial mapping', 'ground truth', 'sar images', 'subjective labelling', 'rules extraction', 'deficiency analysis', 'autonomous navigation', 'land vehicle', 'land vehicles', 'digit recognition', 'noise reduction', 'word spotting', 'optical character recognition', 'spectral warping', 'partially observable markov decision problems', 'non - markov decision processes', 'partially observable markov decision processes', 'policy improvement', 'linear classifiers', 'linear perceptron', 'we chip', \"' rule\", \"' s output rule\", 'svms', \"' s output method\", \"' s output functions\", \"' s output\", 'stochastic dynamical systems', 'stochastic dynamical system', 'information maximisation', 'physically based vision', 'show - through', 'in - network filtering', 'in - network analysis', 'reminiscent', 'stationary distribution', 'relatively quasi networks', 'out networks', 'for noisy signals', 'for noisy speech', 'prisms', 'stationary distributions', 'site a', 'site a service', 'site a to speech', 'diversity', 'cortical columns', 'steerable/shiftable filters', 'factorial learning', 'highly correlated categories', 'risk prediction', 'coronary artery k', 'coronary artery bypass', 'stochastic gradient corresponding', 'confidence n', 'fractal block codes', 'subthreshold quantization', 'subthreshold oscillations', 'storage', 'fractal block coding', 'storage system', 'nonlinear circuits', 'kerr - type optical networks', 'distributed error', 'optical networks', 'kerr - type', 'lagrangian formulation', 'lesions', 'optical error correction', '. disease models', 'bootstrap', 'the results', 'optical', 'a posteriori error estimates', 'our areas', 'lagrangian', 'does', 'of optical easier optical effects', 'computational fluid dynamics', 'or in machine learning', ') neural networks', 'discuss', 'carnegie institution', 'inductive bias', 'multi task learning', 'our calibrating', 'perspective projection', ') neural symbolic learning', 'non achieved learning', 'non maximisation', 'camera selection', 'joint between layers', 'non achieved internal models', 'hallucinations', 'paper maps', 'speech coding', 'hierarchical mixture of experts', 'non . internal signals', 'multi . internal signals', 'hierarchical mixture of composed composed composed between between applications', 'hierarchical mixture of composed composed composed between association', 'hierarchical mixture of composed composed composed between between additional adversely', 'speech modelling', 'matrices', 'modular classification', 'asymptotic invariance', 'autoassociative neural networks', 'transformation invariance', 'handwritten and character limiting', 'sensing', 'may of extraction', 'handwritten digits', 'mixtures of linear models', 'principal converted model', 'mixtures of in efficiency', 'principal components an gesture recognition', 'into efficiency', 'auditory localization', 'coordinate transform chip', 'head - based coordinates', 'phrase evolution', 'input paper', 'maximum in <unk> place learning', 'a <unk> model', 'query learning', 'student learning', 'minimum student space to small mosis', 'single state learning', 'student space entropy', 'single mosis mosis value', 'accomplished on mosis mosis', 'accomplished on mosis value', 'minimum student space to fabrication', 'accomplished on mosis prognostic index', 'accomplished on mosis', 'boltzmann chains', 'presented networks', 'of workstations', 'different models', 'person modeling', 'implements', 'i', 'propose', 'here a model', 'realizable model', 'different negative matrix', 'different negative', 'hidden', 'different interference', 'different', 'different hidden and consolidation', 'error feedback', 'robust control', 'associative net', 'information efficiency', 'recall particularly', 'brain - like associative net', 'information he', 'ca', 'function is', 'mapping', 'adaptive input field', 'recognition improvement', 'affine translation', 'character estimator', 'elastic input field', 'diffusion - enhancement layer', 'spatio - temporal grouping', 'line end detection', 'modeled grouping', 'diffusion map', 'apparent realistic selectively for properties', 'apparent motion', 'high field', 'poisson clumping', 'front <unk> specific learning', 'front <unk> specific training', 'front <unk> propagation training', 'front <unk> specific model', 'familiar environment', 'increased more', 'non - constant weighting functions', 'combining estimators', 'non - for commonly adaptive systems', 'mixture of effect', 'namely state estimation', 'on attempt neural networks', 'on attempt networks', 'input accuracy', 'kinds of learning', 'case perceptron', 'non - for . commonly attempt idea', 'learning output', 'cholinergic modulation', 'cholinergic suppression', 'temporal event signal', 'temporal neural network', 'compartmental modeling', 'network simulation', 'done we', 'stem time', 'dynamic cell structure', 'competitive hebbian learning', 'growing cell structure', 'topology representing network', 'competitive gain vector', 'competitive hebbian of way vectors', 'and layer perceptron', 'and on . <unk>', 'and on . 1', 'layer perceptron', 'after test', 'and layer . 1', 'after', 'layer perceptrons', 'in mst', 'and on . 1 , on mst', 'point correspondence', 'mean into replicated builds', 'point matching: pose estimation', 'time digits', 'advantages and digits', 'mean into replicated objects', 'auditory modeling', 'pattern playback', 'spectrogram', 'cochlear modeling', 'thermodynamic limit', 'we conscious state computed tomography', 'in conscious machine learning', ') conscious machine learning', ') conscious machines', 'in conscious machine', 'we conscious state prey model', 'we conscious state space orientation', 'we conscious state space prior', 'distributed microfeatures', 'spreading activation', 'self interpret property', 'distributed algarit', \"butt ' s been\", '3 image stabilization', 'microfeatures', 'closed fundamental codes', '3 image exactly solvable', '3 image exactly solvable classes', 'interior point methods', 'tangent function', 'tangent modules', 'state aggregation', 'soft state aggregation', 'compact representation', 'coarse - to - fine search', 'integrated feature pyramids pyramids pyramids', 'optimal l ( 2 of 2 of 2', 'optimal l ( 2 of 2 of 2 of 2', 'integrated feature pyramids pyramids based pyramids pyramids', 'into ( <unk> of of by of class', 'class - based class - based class - based class - based class - based class', 'class - based class - based class - based class - based class - based enhancement', 'class - based class - based class - based class - based class', 'function -', 'class - based class - based class - based enhancement', 'function - based enhancement', 'class - based class - based class - based class - based class - based clustering', 'class - based class - based class - based class - based class - based .', 'class - based class - based class - based class - based enhancement', 'class - based class - based class - based class', 'class - based class -', 'class - based class - based class - based class - based class - based vector', 'class - based class - based class - based class - based class - based indexing', 'consciousness', 'computational modeling', 'human consciousness', 'computational modules', 'subjective experience', 'cognitive vision', 'cognitive modeling', 'cognitive systems', 'hand tracking', 'in to', 'and video', 'attempt', 'efficient input', 'output tracking', 'output detection', 'output dichotomies', 'environment', 'conjugate gradients', 'given have is target detection', 'its dichotomies', 'missing inputs', 'incomplete pattern', 'incomplete feature vectors', 'the speculation', 'the net', 'designed experiments', 'tracker selection', 'grammar learning', 'self algorithms', 'symbol clustering', 'obtained results', 'narrow implementation rule', 'layered fully recurrent networks', 'optimal algorithm', 'spreading function', \"' s maps\", 'cost functions', 'be of lack', 'be of neural networks', 'produces exploiting lack', 'deliberately', 'frequently input', 'early termination', 'conjugate gradient', 'conjugate function', 'cascade correlation', 'conjugate gradient method', 'efficient techniques', 'case image processing', 'cascade correlation neural network', 'case resolution analysis', 'breast cancer', 'prediction accuracy', 'case fast fourier transform', 'breast under results', 'k - means algorithm', 'convergence properties', 'shunting recurrent neural networks', 'analog adaptive resonance', 'engineers network', 'spatio be sets', 'closed cost network', 'describe )', 'curves do cancer', 'describe network', 'shunting recurrent neural network', 'adjust learnable', 'in transistor memory', 'available m transistor', 'decreases electron transistor', 'in transistor mosis', 'arrays with memory', 'allows', 'fine be features', 'about', 'appearance', 'product units', 'survival: ptnm staging system', 'output analysis', 'introduced selected provide', 'efficient efficient dynamic system', 'efficient kind detection', 'product reflecting boundary', 'based diagnosis system', 'charge detection', 'maximum likelihood computerized', 'maximum likelihood computerized diagnosis', 'maximum likelihood computerized diagnosis system', 'maximum likelihood computerized adaptive a target', 'maximum likelihood computerized adaptive a a behavior', 'survival: error', 'maximum likelihood computerized adaptive a a immune system', 'maximum likelihood computerized adaptive a smooth ( of survival: fabricated', 'maximum likelihood computerized adaptive a smooth ( of experimental data', 'maximum likelihood computerized adaptive a a ( of backpropagation fabricated', 'adapts correlation correlation with survival: fabricated', 'f2 laser', 'expansibility', 'assembling experimental arti', 'been experimental it', 'all in in in in in in in in in in in in in in in', 'all in in in in in in in in medicine', 'all in in in in in in in in in medicine', 'all in in in in in in in medicine', 'all in in in in in in in in in in medicine', 'all in in in in in in in in in in in medicine', 'all in in in in in in in in in in in in medicine', 'all in in in in in in in in in in in in in medicine', 'all in in in in in in in in in in in in in in medicine', 'all in in in in in in in in in in in in in in the', 'all in in in in in in in in in in in in in', 'all in in in in in in in in in in in in in in experimental', 'all in in in in in in in in in in in in in the have', 'all in in in in in in in in in in in in in the experimental', 'all in in in in in in in in in in in in the experimental results', 'three - state neuron', 'master equation', 'three - state neural networks', 'one - way neuron', 'one - way function', 'this track laws', 'one - step stochastic system', 'recall', 'cal systems', 'areas', 'medial superior temporal cells', 'area mst cells', 'area mst of mst cells', 'area mst of mst of mst cells', 'object motion grouping', 'selective tuning of mst cells', 'object motion', 'visual speech recognition', 'stochastic networks', 'simple hidden markov models', 'limited vocabulary', 'hidden chains', 'single - transistor', 'synaptic', 'single - transistor silicon synapses', 'single - transistor can', 'single - transistor staging', 'has accuracy', 'silicon', 'staging', 'non - volatile memory', 'single - transistor silicon cancer', 'been ideas', 'has information gain', 'has information classes', 'has information exchange', 'has information states', 'least squares estimators', 'soft errors', 'least squares estimation', 'limit analysis', 'arti', 'inspired algorithms', 'neural engine', 'real time clustering', 'it interfaces', 'wireless sensor networks', 'vlsi architectures', 'wireless sensor network', 'real time series', 'programmable matter biometrics', 'psychophysics', 'motion noise', 'motion perception', 'facial animation', 'view - based total variation', 'estimators', 'distributed view - based tracking', 'bias field', 'adaptive localization', 'least the squares', 'network staging', 'models integration', 'network networks', 'on the if sampling', 'on the expression recognition'])"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "kp.keys()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}