Generalizing to Unseen Domains via Adversarial Data Augmentation Istituto Italiano di Tecnologia, Stanford
- domain generalization
- data-dependent regularization
- experiments on: digits, SYTHIA (season and weather change)
- TF1.6 py2.7
Algorithms and Theory for Multiple-Source Adaptation - NIPS18
- problem to solve: multi-source adaptation to new mixture target domain
- proposed method: distribution-weighted combination, DC-computing
Conditional Adversarial Domain Adaptation - NIPS18 (CDAN)
- problem to solve: to align multi-mode domains, adaptation of a layer is not sufficient to bridge domain shifts
- proposed method: conditional adversarial network
- conditional domain discriminator conditioned on feature representations and classifier predictions
- Caffe PyTorch 0.4
Adversarial Multiple Source Domain Adaptation - NIPS18
- problem to solve: a generalisation bound for multi-source domain adaptation
- proposed method: Multi-source Domain Adversarial Networks (MDANs)
Co-regularized Alignment for Unsupervised Domain Adaptation - NIPS18 (Co-DA)
Synthesize Policies for Transfer and Adaptation across Tasks and Environments - NIPS18
- problem to solve: simultaneous transfer of environment and task in reinforcement learning
- proposed method: new architecture and new training method (meta rule)
- adapt to new environment and task pair with as few seen pairs as possible
- a disentanglement objective: environment embedding and task embedding
- policy factorisation and composition
Domain Adaptation by Using Causal Inference to Predict Invariant Conditional Distributions - NIPS18
-
problem to solve: predict target variable distribution from measurements of other variables
-
proposed method:
Unsupervised Image-to-Image Translation Using Domain-Specific Variational Information Bound - NIPS18
Adapted Deep Embeddings: A Synthesis of Methods for k-Shot Inductive Transfer Learning - NIPS18
Hardware Conditioned Policies for Multi-Robot Transfer Learning - NIPS18
- problem to solve: transfer learnt policy to new robots
- proposed method: universal policy conditioned on hardware represented as a vector
Transfer Learning with Neural AutoML - NIPS18
- problem to solve: expensive cost to do Neural Architecture Search (NAS)
- proposed method: multi-task learning and transfer learning
- parallel multi-task training on NAS
- initialise controller parameters for new task with pre-trained controller, add a randomly initialised task embedding
GLoMo: Unsupervised Learning of Transferable Relational Graphs - NIPS18
- problem to solve: build richer and more versatile representations for transfer
- proposed method: unsupervised training of transferable relational graph
- asymmetric affinity matrix to model dependencies between paired data
- Graphs from LO-w-level unit MOdeling
- graph predictor: key CNN and query CNN and feature predictor:
Modeling Visual Context is Key to Augmenting Object Detection Datasets INRIA
Effective use of synthetic data for urban scene semantic segmentation (VEIS)
Multimodal Unsupervised Image-to-Image Translation - ECCV18 (MUNIT)
- problem to solve: failure to generate diverse outputs
- proposed method: multi-model, split domain-invariant content code and domain-specific style code (space)
Graph Adaptive Knowledge Transfer for Unsupervised Domain Adaptation - ECCV18
- proposed method: jointly optimise target labels and domain-invariant features, graph-based label propagation
- subspace learning, cross-domain graph, label propagation, EM-like alternating optimisation step
- top performance and experiments on Office only
Domain Adaptation through Synthesis for Unsupervised Person Re-identification - ECCV18
- drastic illumination variations across surveillance cameras
- new synthetic dataset containing various illumination conditions
- problem to solve: increase in representation power alone cannot overcome the domain gap
- proposed method: iterative self-training a latent variable loss minimisation problem
- class-balanced self-training, spatial priors, global and class-wise feature alignment
- SYNTHIA/GTA5 to Cityscapes, Cityscapes to NTHU
Open Set Domain Adaptation by Backpropagation - ECCV18
- problem to solve: unknown target samples should not be aligned with source
- proposed method: adversarial training tot separate unknown features from known
- gradient reversal for training feature generator (extractor)
- problem to solve: retrieve 3D shapes by sketches
- proposed method: cross-modality transformation network for feature transfer, adversarial learning
- 3D shape projected into multiple 2D and averaged
- importance-aware metric learning, batch-wise hardest sample mining
- class-aware MMD
- problem to solve:
- proposed method: CycleGAN pair generation
Meta-Tracker: Fast and Robust Online Adaptation for Visual Object Trackers - ECCV18
- problem to solve: offline meta-learning
AugGAN: Cross Domain Adaptation with GAN-based Data Augmentation - ECCV18
- problem to solve: image translation methods fail to preserve image objects
- segmentation subtask, weight sharing strategy, night-time vehicle detection (emphasize)
- experiments on detection task
- remarks: require segmentation mask for training?
Penalizing Top Performers: Conservative Loss for Semantic Segmentation Adaptation - ECCV18
- problem to solve: learn representations that are discriminating for the task and generalised
- proposed method: conservative loss that penalises extreme good (prevent from biasing towards source data) and bad predictions
- universally reversing gradients for all pixels is not suitable for structured prediction in segmentation
- performances on the source and target domain do not reach the best at the same time (close but not together)
- shared encoder, shared decoder as generator, separate discriminator
Zero-Shot Deep Domain Adaptation - ECCV18
- problem to solve: zero-shot adaptation/learning problem
- proposed method: task-irrelevant dual-domain pairs; cross-image-modality???
- new setting: task-irrelevant pair given, task-relevant target training sample not given, applied to sensor fusion
- test with target CNN trained on task-irrelevant target samples and source classifier
DCAN: Dual Channel-wise Alignment Networks for Unsupervised Scene Adaptation - ECCV18
- problem to solve: adaptation for pixel segmentation
- proposed method: channel-wise high-level feature maps alignment, normalisation of source and target images by matching channel-wise feature statistics
- adversarial training is heavy and difficult to train?
- problem to solve: generate segmentation annotations
- proposed method: new dataset, data preparation pipeline
Unsupervised Domain Adaptation for 3D Keypoint Estimation via View Consistency - ECCV18
- problem to solve: 3D key point prediction from single depth image
- proposed method: multi-view consistency, geometric alignment regularisation term. alternating optimisation
- generated large depth image dataset from ShapeNet and ModelNet
Partial Adversarial Domain Adaptation - ECCV18
- problem to solve: target label space is a subset of source label space
- proposed method: match features in shared label space, down-weigh outlier source class samples
DeepJDOT: Deep Joint Distribution Optimal Transport for Unsupervised Domain Adaptation - ECCV18
- problem to solve: poor scaling and poor representation to align in JDOT
- proposed method: simultaneous learning of joint representation and discriminative information
- Keras TF implementation
Real-to-Virtual Domain Unification for End-to-End Autonomous Driving - ECCV18
- problem to solve: merge data from different sources for model generalisation and interpretability
- proposed method: real-to-synthetic mapping, command prediction on synthetic data
- real-to-virtual mapping is actually not easy?
- every dataset has bias
- instance normalisation for image translation/generation tasks? Yes
Domain transfer through deep activation matching - ECCV18 (DAM)
- problem to solve: adaptation for semantic segmentation
- proposed method: layer-wise feature alignment (same as in CDAN)
- experiments on GTA to Cityscapes, SYNTHIA to Cityscapes, USPS to MNIST
- tensorflow code on digits based on ADDA
- assumption: all activation distributions are i.i.d. Gaussian?
- GradRev used GAN??
- image-to-image translation-> optimising symmetric confusion metric?
- ADDA: optimise inverted label objective?
- domain shift is anywhere inside network?
NAM: Non-Adversarial Unsupervised Domain Mapping - ECCV18
- problem to solve: unstable adversarial training
- proposed method: pre-trained generative model in the target domain, then source-to-target mapping
Choose Your Neuron: Incorporating Domain Knowledge through Neuron-Importance - ECCV18
- problem to solve: zero-shot learning
- proposed method: learn a mapping between class-specific domain knowledge and importance of individual neurons
- neuron importance as intermediate representation
- experiments on CUBirds (37.3k) and AWA2 (11.7k) generalised zero-shot learning benchmark
Learning Deep Representations with Probabilistic Knowledge Transfer - ECCV18
- PyTorch implementation
- problem to solve: knowledge transfer (from large network to smaller one), not just for classfiication
- proposed method: match feature space data distribution, embed teacher model feature space into student model
- KT applied to handcrafted features and text-to-image cross-modal transfer
Contour Knowledge Transfer for Salient Object Detection - ECCV18
- Caffe implementation
- problem to solve: save annotations required for salient object detection
- proposed method: multi-task network, contour-to-saliency transfer for ground truth generation, alternating training
- definition of saliency??
- convert a trained contour detection model into a saliency detection model
Zero-Annotation Object Detection with Web Knowledge Transfer - ECCV18
- problem to solve: zero-shot detection
- proposed method: multi-instance multi-label adaptation
- foreground attention and instance level adversarial adaptation, appearance transfer from web to target, pseudo labelling for transfer
- weakly supervised detector trained with image labels
- good rationing, explicitly pointing out the advantage with examples for reader to undetstand
Learning to Adapt Structured Output Space for Semantic Segmentation (AdaptSegNet)
- GTA5/SYNTHIA to CItyscapes
- PyTorch
Non-local Neural Networks CMU & Facebook
-
non-local operation: response at a position as a weighted sum of the features at all features
-
vanilla Gaussian
-
embedded Gaussian: similarity in an embedding space
-
dot-product: dot-product similarity
-
concatenation
Unsupervised Domain Adaptation with Similarity Learning (SimNet)
- 69.58% per-class average accuracy on VisDA17 validation set (ResNet-50), 72.9% with ImageNet pre-trained ResNet-152
- 88.6% classification accuracy on Office-31 A->W (ResNet-50)
Image to Image Translation for Domain Adaptation (I2IAdapt) - CVPR18
- generalised framework: classification loss, reconstruction loss, feature D loss, translation loss, cycle loss, translated classification loss
- experiments on digits, Office-31, GTA5 to Cityscapes
Generate to adapt: Aligning domains using generative adversarial networks (Gen2Adapt) - CVPR18
- train D, G to tune F to obtain target feature extractor??? F trained with source labels
- joint embedding learning
- experiments: digits, Office-31, CAD renderred to PASCAL VOC, VisDA
- comparison with other GAN methods: use GAN to obtain rich gradient to learn embedding??
- superior results compared to auto-encoder and disentangled embedding from adversarial learning
- works well when image generation is hard
- G & D trained with discriminative info from source labels
Collaborative and Adversarial Network for Unsupervised domain adaptation(CAN)
- problem to solve: domain information remains in lower layers but lost in top layers
- set of domain discriminators, iterative pseudo-labelled sample selection
- domain informative representation from lower blocks and domain uninformative representations from higher blocks
- experiments on Office and ImageCLEF-DA
Conditional Image-to-Image Translation - CVPR18 (cd-GAN)
- problem to solve: previous image translation is one-to-one deterministic
- proposed method: conditioning, bidirectional translation, reconstruction
Re-weighted Adversarial Adaptation Network for Unsupervised Domain Adaptation (RAAN)
- problem to solve: reduce feature distribution divergence and adapt classifier
- proposed method: minimise optimal transport based earth-mover distance; re-weighted source domain label distribution
- instance re-weighting
**Conditional Generative Adversarial Network for Structured Domain Adaptation - CVPR18
- problem to solve: synthetic-to-real segmentation adaptation
- proposed method: conditional synthetic-to-real feature generator, discriminator to fuse features in 2 domains
- learn a residual between feature maps from different domains
- condition on source images and noise label
- experiments on SYNTHIA/GTA5 to Cityscapes
Duplex Generative Adversarial Network for Unsupervised Domain Adaptation - CVPR18
- problem to solve: domain-invariant representation and domain transformation learning
- proposed method: encoder, generator and 2 discriminators
- problem to solve:
- proposed method:
**Maximum Classifier Discrepancy for Unsupervised Domain Adaptation - CVPR18 Oral
- PyTorch implementation
- problem to solve: class info lost in adversarial training; difficult to match feature distributions across domains completely
- proposed method: 2 task classifiers with different initialisation, no domain label, multi-stage training
- assumption: target samples far from source are likely to be classifier differently by different classifiers
- maximise discrepancy between classifiers, feature generator minimises discrepancy, adversarial training
- very clear writing
- notation of feature generator, any generative component???
- target samples near class boundary are likely to be mis-classified by classifier: better show examples
- figure and diagram: stand on its own without having to refer to text
- first stage: train with source labels
- second stage: train classifiers with fixed features to detect target far from source, same number of source & target samples, source classification loss is required to keep performance
- third stage: train feature generator with fixed classifiers
- training order does not matter, but classifier performance is the key
- experiments on self-built toy dataset, digits, VisDA17 classification, GTA5/SYNTHIA to Cityscapes
Boosting Domain Adaptation by Discovering Latent Domains - CVPR18
- problem to solve: source domain as mixture of multiple domains
- proposed method: discover multiple latent domains from source and align target with them
- DA layer: align source and target representation to a reference Gaussian distribution
- build on top of AutoDIAL: Automatic DomaIn Alignment Layers, extend to multiple latent domains
- experiments on Office-31 and Office-Caltech
- problem to solve: multiple source domain with different label spaces
- proposed method: target distribution modelled as weighted combination of source, alternating training
- multi-way adversarial training: minimise discrepancy between target and each source
- use pseudo-labelled target samples to update multi-source classifier and feature extractor
- experiments on Office-31, ImageCLEF-DA and Digits-five
*Residual Parameter Transfer for Deep Domain Adaptation - CVPR18
- problem to solve: same architecture for different domains - restrict domain gap; adaptation network - increase network parameters
- proposed method: auxiliary residual networks to predict target parameters with few labels
- experiments on SVHN to MNIST, Office-31
- evaluation protocol: Saenko ECCV10 paper? Long ICML15 paper?
- training and testing difference: use source classifier
- comparison with RTN - NIPS16?
- comparison with self-ensembling ICLR18?
- how to train???
Cross-Domain Weakly-Supervised Object Detection Through Progressive Domain Adaptation - CVPR18
- Chainer implementation
- problem to solve: cross-domain weakly-supervised detection problem
- proposed method: multi-stage training, new dataset
- first stage: supervised training on source set
- second stage: target training sample generation from source set using CycleGAN
- third stage: fine-tune source network on generated target-like images
- fourth stage: obtain pseudo instance-level annotation using fine-tuned network
Camera Style Adaptation for Person Re-Identification
*Adversarial Feature Augmentation for Unsupervised Domain Adaptation - CVPR18
- Tensorflow implementation
- problem to solve: learn domain-invariant feature representations
- proposed method: source feature augmentation, multi-stage training
- first stage: train source network
- second stage: train feature generator with noise and one-hot label input, and feature discriminator
- third stage: train feature extractor again with samples from both source and target (map both into common feature space)
- experiments on digits, NYUD RGB to depth
Deep Face Detector Adaptation Without Negative Transfer or Catastrophic Forgetting
Cross-Dataset Adaptation for Visual Question Answering
Fully Convolutional Adaptation Networks for Semantic Segmentation - CVPR18
- problem to solve: synthetic-to-real adaptation for segmentation
- proposed method: appearance and representation adaptation
- experiments on GTA5 to Cityscapes and BDDS dataset
- generate image pair in another domain
Importance Weighted Adversarial Nets for Partial Domain Adaptation
ROAD: Reality Oriented Adaptation for Semantic Segmentation of Urban Scenes - CVPR18
- problem to solve: synthetic-to-real urban scene adaptation for segmentation
- proposed method: ImageNet model to regularise synthetic feature learning
- experiments on GTA5/SYNTHIA to Cityscapes
Person Transfer GAN to Bridge Domain Gap for Person Re-Identification - CVPR18
- problem to solve: person re-id in viewpoint, lighting changes
- proposed method: new dataset
Partial Transfer Learning With Selective Adversarial Networks - CVPR18
- problem to solve: target label space is subset of source label space
- proposed method:
Taskonomy: Disentangling Task Transfer Learning Stanford,
Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning
Coupled End-to-End Transfer Learning With Generalized Fisher Information
CleanNet: Transfer Learning for Scalable Image Classifier Training With Label Nois
Instance Embedding Transfer to Unsupervised Video Object Segmentation - CVPR18
- problem to solve: video segmentation
- proposed method: transfer from image segmentation
Multi-Content GAN for Few-Shot Font Style Transfer
Revisiting Knowledge Transfer for Training Object Class Detectors - CVPR18
- problem to solve: weakly supervised target detector training from fully supervised source detectors
- proposed method:
Avatar-Net: Multi-Scale Zero-Shot Style Transfer by Feature Decoration - CVPR18
Separating Style and Content for Generalized Style Transfer - CVPR18
- problem to solve: current style transfer cannot generalize to new styles
- proposed method: separate style and content
- allow to transfer to multiple styles at the same time
- experiments on Chinese Typeface dataset
- domain separation networks for style transfer??
Feature Space Transfer for Data Augmentation - CVPR18 (FATTEN)
- problem to solve: model feature trajectory of various object poses
- proposed method:
- experiments on
Deep Cross-Media Knowledge Transfer - CVPR18
- problem to solve:
- proposed method:
Boosting Self-Supervised Learning via Knowledge Transfer - CVPR18
- problem to solve:
- proposed method:
Cross-Domain Self-Supervised Multi-Task Feature Learning Using Synthetic Imagery - CVPR18
- problem to solve:
- proposed method:
Domain Adaptive Faster R-CNN for Object Detection in the Wild - CVPR18
- problem to solve:
- proposed method:
Learning From Synthetic Data: Addressing Domain Shift for Semantic Segmentation - CVPR18
- problem to solve:
- proposed method:
Domain Generalization With Adversarial Feature Learning - CVPR18
- problem to solve:
- proposed method:
Efficient Parametrization of Multi-Domain Deep Neural Networks - CVPR18
Detach and Adapt: Learning Cross-Domain Disentangled Deep Representation - CVPR18
- problem to solve:
- proposed method:
Mutual Information Neural Estimation (MINE)
- dual representation of KL:
- Donsker-Varadhan: $D_{KL}(\mathbb{P}||\mathbb{Q})=\sup_{T:\Omega\to \mathbb{R}}\mathbb{E}\mathbb{P}[T]-\log (\mathbb{E}\mathbb{Q}[e^T])$
- $f$-divergence: $D_{KL}(\mathbb{P}||\mathbb{Q})\ge\sup_{T\in \mathcal{F}}\mathbb{E}\mathbb{P}[T]-\mathbb{E}\mathbb{Q}[e^{T-1}]$
- statistics network
$T_\theta,\theta\in\Theta$ :$I(X;Z)\ge I_\Theta(X,Z)$ - definition: $I_\Theta(X,Z)=\sup_{\theta\in\Theta}\mathbb{E}{\mathbb{P}{XZ}}[T_\theta]-\log(\mathbb{E}{\mathbb{P}{X}\otimes\mathbb{P}{Z}}[e^{T\theta}])$ (and a MINE-$f$)
- sample from
$(\bar{x},z),(x,\bar{z})\sim\mathbb{P}_{XZ}$
- unofficial pytorch
- follow-up: A Data-Efficient MINE for Statistical Dependency Testing ICLR20 submission
CyCADA: Cycle-Consistent Adversarial Domain Adaptation
Transfer Learning via Learning to Transfer
- problem to solve: automatically determine what and how to transfer
- proposed method: combine meta-learning and transfer learning
- comparison with traditional TL. multi-task learning, lifelong (meta) learning
Pseudo-task Augmentation: From Deep Multitask Learning to Intratask Sharing—and Back - ICML18
- problem to solve: multi-task learning from single task
- proposed method: task augmentation with pseudo tasks, multiple decoders for single task
Progress & Compress: A scalable framework for continual learning - ICML18
- problem to solve: sequential continual learning
- proposed method: knowledge base of previous solutions and active column for current task
- active learning and consolidation of new task into knowledge base
- experiments on sequential handwritten alphabets classification, Atari and 3D maze game
Detecting and Correcting for Label Shift with Black Box Predictors - ICML18
- problem to solve: detect and quantify label shift, correct classifier without target label
- proposed method:
Learning Adversarially Fair and Transferable Representations - ICML18
- problem to solve: representation learning for fair transfer
- proposed method: connect group fairness metrics (demographic parity, equalise odds, equal opportunity)
Learning Semantic Representations for Unsupervised Domain Adaptation - ICML18 (MSTN)
- problem to solve: category-aware feature alignment/mapping
- proposed method: use moving average centroid alignment to align labelled source centroid and pseudo-labelled target centroid
- few false pseudo labels can lead to extremely biased estimation in mini batch SGD training - moving average
Rectify Heterogeneous Models with Semantic Mapping - ICML18
- problem to solve: learn and use transferable (heterogeneous) models, for model reuse
- proposed method: meta information of features, rectify via heterogeneous predictor mapping
- 4 pages of mathematical proof
- notation, theoretical explanation (model reuse on homogeneous features and heterogeneous feature, meta feature representation and its encoding),
- optimal transport for model reused in heterogeneous feature space
- Bregman Alternating Direction Method of Multipliers to linearise loss function for faster optimisation
- experiments on general classification, user quality classification, academic paper classification
- discussion of extension to deep networks
JointGAN: Multi-Domain Joint Distribution Learning with Generative Adversarial Nets - ICML18
- problem to solve: multi-source domain joint learning, lack of learned sample mechanism for multiple domain marginal distrbutions
- proposed method: learn decomposed marginal and conditional distribution by adversarial training (combined together as joint distribution)
Augmented CycleGAN: Learning Many-to-Many Mappings from Unpaired Data - ICML18
- problem to solve:
- proposed method:
Adversarially Regularized Autoencoders - ICML18 (ARAE)
- PyTorch implementation
- problem to solve: training deep latent variable models on discrete input
- proposed method: use Wasserstein auto-encoder to model discrete sequences
- experiments on unaligned text style transfer, natural language inference
Video Prediction with Appearance and Motion Conditions - ICML18
- problem to solve: reduce uncertainty in predicting future frames with appearance and motion conditions
- proposed method: appearance-motion conditional GAN
- experiments on facial expression, human action
Importance Weighted Transfer of Samples in Reinforcement Learning - ICML18
- problem to solve: task transfer in RL need to consider discrepancies between task model
- proposed method: estimate relevance of source sample for the task (instance transfer)
Knowledge Transfer with Jacobian Matching - ICML18
- problem to solve: appropriate loss function for Jacobian matching in knowledge distillation/transfer
- proposed method: Jacobian matching = distillation with noise input
Towards Black-box Iterative Machine Teaching - ICML18
- problem to solve: cross-space machine teaching
- proposed method:
Improved Training of Generative Adversarial Networks Using Representative Features - ICML18
- problem to solve: trade-off between generated image quality and diversity
- proposed method: regularise the discriminator using features
MIXGAN: Learning Concepts from Different Domains for Mixture Generation - IJCAI18 Sun Yat-sen University
- mix content and style to generate samples in a new domain
Zero-Shot Visual Imitation - ICLR18
- problem to solve: imitation learning when agent first explores the environment without expert supervision
- proposed method:
A DIRT-T Approach to Unsupervised Domain Adaptation - ICLR18
- problem to solve: feature distribution matching is a weak constraint for high-capacity feature extraction function; model optimal on source are not optimal for target
- proposed method:
- non-conserve domain adaptation - no single classifier can perform well in both the source and target domains
- lens of the cluster assumption: decision boundaries should not cross high-density data regions
- virtual adversarial DA model, Decision-boundary Iterative Refinement Training with a Teacher
- experiments on digits, traffic sign and Wi-Fi activity recognition
Generalizing Across Domains via Cross-Gradient Training - ICLR18
- problem to solve: domain generalisation
- proposed method: domain-guided perturbation, data augmentation as sampling
- conclusion: domain-guided perturbation is better for domain generalisation; data augmentation is more stable and accurate than domain adversarial training
- experiments on google fonts character, handwritten character, Google Speech Command Dataset cross-speaker
Identifying Analogies Across Domains - ICLR18
- problem to solve:
- proposed method:
Minimal-Entropy Correlation Alignment for Unsupervised Deep Domain Adaptation - ICLR18 (MECA)
- Tensorflow implementation
- problem to solve: optimal alignment of second order statistics between source and target
- proposed method: align along geodesics instead of Euclidean, weighted source-to-target regularisation
- experiments on SVHN to MNIST, NYUD (RGB to depth), SYN digits to SVHN
Learning to Cluster in Order to Transfer Across Domains and Tasks - ICLR18
- problem to solve:
- proposed method: transfer pairwise semantic similarity (learnt from source)
- experiments on SVHN to MNIST, Office-31
- mean teacher variant of temporal ensembling, confidence thresholding and class balance
- 98.23 on MNIST$$\to$$ USPS, 99.54 on USPS$$\to$$ MNIST
- 99.26 on SVHN$$\to$$ MNIST
- 74.2% on VisDA17 validation set (ResNet-152 with minimal augmentation),
- Minimal augmentation - Gaussian noise
- TF - shift & flip & ?
- TFA - shift & flip & affine & ?
- Reduced augmentation - random crop, scaling, flip
- Competition - brightness scaling, rotation, de-saturation, colour space rotation & offset
Multi-Adversarial Domain Adaptation
Beyond sharing weights for deep domain adaptation EPFL
Adversarial teacher-student learning for unsupervised domain adaptation - ICASSP18
- problem to solve:
- proposed method: