Skip to content

Latest commit

 

History

History
251 lines (156 loc) · 6.59 KB

README.md

File metadata and controls

251 lines (156 loc) · 6.59 KB

Deep Learning reading lists, 2011-present

Courses:

Basic Machine Learning:

back-propagation

Theory / math: http://www.deeplearningbook.org/contents/mlp.html see chapter 6.5

linear layers: https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/

conv net layers: https://grzegorzgwardys.wordpress.com/2016/04/22/8/ and https://www.slideshare.net/kuwajima/cnnbp and http://www.jefkine.com/general/2016/09/05/backpropagation-in-convolutional-neural-networks/ [you need to save the maxpooling indices for back-prop]

RNN - recurrent neural nets:

nice explanation

https://www.cs.ox.ac.uk/people/nando.defreitas/machinelearning/ http://karpathy.github.io/2015/05/21/rnn-effectiveness/

RNN code is here:

https://github.com/karpathy/char-rnn https://github.com/wojzaremba/lstm

##RNN Recurrent neural networks useful links

Graphs in Torch: You need this before attempting to start digesting the LSTM code. Exercise: https://www.cs.ox.ac.uk/people/nando.defreitas/machinelearning/practicals/practical5.pdf Nando de Freitas lecture

Videos: https://youtu.be/56TYLaQN4N8 (LSTM basics) https://youtu.be/-yX1SYeDHbg (Alex Graves’s hand-writer algorithm) Sides: https://www.cs.ox.ac.uk/people/nando.defreitas/machinelearning/lecture11.pdf Exercise: https://www.cs.ox.ac.uk/people/nando.defreitas/machinelearning/practicals/practical6.pdf

Soumith article http://devblogs.nvidia.com/parallelforall/understanding-natural-language-deep-neural-networks-using-torch/

Ilya Sutskever NIPS paper Sequence to Sequence Learning with Neural Networks http://arxiv.org/abs/1409.3215

Alex Graves (43 pages) paper Generating Sequences With Recurrent Neural Networks Prediction network – Long Short-term Memory Cell Text Prediction Penn Treebank Experiments Wikipedia Experiments Handwriting Prediction Handwriting Synthesis http://arxiv.org/abs/1308.0850

Alex Graves LSTM paper Speech Recognition with Deep Recurrent Neural Networks http://arxiv.org/abs/1303.5778

Wojciech Zaremba regularisation paper Recurrent Neural Network Regularization http://arxiv.org/abs/1409.2329

Hochreiter & Schmidhuber (the article) http://deeplearning.cs.cmu.edu/pdfs/Hochreiter97_lstm.pdf

Schmidhuber website Plenty of applications and references http://people.idsia.ch/~juergen/rnn.html

Schmidhuber LSTM tutorial Video: https://www.youtube.com/watch?v=JSNZA8jVcm4 Slides: http://people.idsia.ch/~juergen/deep2014white.pdf

Andrej Karpathy blog post Link: http://karpathy.github.io/2015/05/21/rnn-effectiveness/

Understanding LSTM modules http://colah.github.io/posts/2015-08-Understanding-LSTMs/

And its uses: http://colah.github.io/posts/2014-07-NLP-RNNs-Representations/

LSTM code torch explained by Adam: https://apaszke.github.io/lstm-explained.html

Augmented RNNs:

http://distill.pub/2016/augmented-rnns/

Attention:

https://blog.heuritech.com/2016/01/20/attention-mechanism/

Design neural nets:

VGG https://arxiv.org/pdf/1409.1556.pdf

PreLU http://arxiv.org/pdf/1603.05201v1.pdf

Inception v4 http://arxiv.org/abs/1602.07261

Inception v3 http://arxiv.org/abs/1409.4842

Inception v2 http://arxiv.org/abs/1512.00567

ResNet http://arxiv.org/abs/1512.03385

NiN http://arxiv.org/abs/1312.4400

systematic evaluation of modules:https://arxiv.org/abs/1606.02228

Xception: https://arxiv.org/abs/1610.02357

Unsupervised

Soumith DCGAN http://arxiv.org/abs/1511.06434

Solving Puzzles http://arxiv.org/abs/1603.09246

co-occurence patches https://arxiv.org/abs/1511.06811 http://graphics.cs.cmu.edu/projects/deepContext/

surrogate classes http://arxiv.org/abs/1406.6909

video LSTM http://arxiv.org/abs/1502.04681

learn to generate images from textual descriptions. https://arxiv.org/abs/1605.05396

sunsup / video prediciton

predicting next frames from video, MIT Torralba http://arxiv.org/abs/1504.08023

prednet Coxlab: https://github.com/coxlab/prednet

Alf extra refs: https://docs.google.com/document/d/1_t7_Q4RxeX_blEQQXYv3MTDXxsCo2ZghfH7zp1y4Hkk

better frame reconstruction by predicting transformations: https://arxiv.org/abs/1511.05440v6

Scrambling of video frames used to train unsup: https://arxiv.org/pdf/1611.06646v2.pdf

Reinforcement Learning

http://karpathy.github.io/2016/05/31/rl/

0-shot learning:

A. Frome, G. S. Corrado, J. Shlens, S. Bengio, J. Dean, M. A. Ranzato, and T. Mikolov. Devise: A deep visual-semantic embedding model. In NIPS, 2013 https://papers.nips.cc/paper/5204-devise-a-deep-visual-semantic-embedding-model.pdf

R. Socher, M. Ganjoo, C. D. Manning, and A. Ng. Zero-shot learning through cross-modal transfer. In NIPS. 2013. https://nlp.stanford.edu/~socherr/SocherGanjooManningNg_NIPS2013.pdf

review: https://arxiv.org/pdf/1703.04394.pdf

1-shot learning and adding new classes to pre-trained

http://fastml.com/good-representations-distance-metric-learning-and-supervised-dimensionality-reduction/

https://arxiv.org/pdf/1606.09282v2.pdf

https://arxiv.org/pdf/1606.04080.pdf

transfer learning:

this paper on using CNN transfer learning ability to reach state-of-art in a lot of other dataset (transferred from ImageNet training): http://arxiv.org/abs/1403.6382

This great paper also shows transfer from ImageNet to PASCAL VOC: http://www.di.ens.fr/~josef/publications/oquab14.pdf

And this paper from Bengio group also has a great analysis: https://papers.nips.cc/paper/5347-how-transferable-are-features-in-deep-neural-networks.pdf https://arxiv.org/abs/1411.1792

This work on image segmentation also use transfer learning of VGG CNN networks: http://arxiv.org/abs/1511.00561

and more details here in this Stanford course material: http://cs231n.github.io/transfer-learning/ http://cs231n.stanford.edu/reports2016/001_Report.pdf http://cs231n.stanford.edu/reports2016/313_Report.pdf

transfer learning vs fully trained for vehicle model http://cs231n.stanford.edu/reports/lediurfinal.pdf

transfer learning from demonstrations for robot trajectory /LSTM + attentions https://arxiv.org/abs/1703.07326

learn to perform similar task to an example video https://arxiv.org/abs/1707.03374

segmentation:

https://arxiv.org/abs/1511.00561

https://arxiv.org/pdf/1505.04597v1.pdf

https://arxiv.org/pdf/1611.10080v1.pdf

https://arxiv.org/abs/1611.09326

Artistic style

Super-resolution

other paper / reference lists: