一个ICLR 2020个人偏向的list

There is nerver nothing we can do.

Generative Model

• [promising results!] U-GAT-IT: Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation [Tensorflow]: Style transfer

• [Discriminator] Real or not real, that is the question [PyTorch]: Change the output of discriminator to a distribution with a vector size of 8 (followed by KL loss), which is similar to the extended soft label.

• [Understanding generative model] Do deep generative models know what they don’t know

• ⭐️⭐️⭐️⭐️⭐️ [Robust GAN] Optimal strategies against generative attacks [PyTorch]: GAN in the middle networks. Find the optimal solution to against attacks, with analysis of how to attack with leaked samples.

• [Loss function] Improving Adversarial Robustness Requires Revisiting Misclassified Examples : Mark

• [Comparison discriminator] Self-Adversarial Learning with Comparative Discrimination for Text Generation

• Controlling generative models with continuous factors of variations

• [Learning para as loss weight] You Only Train Once: Loss-Conditional Training of Deep Networks

• [Augmentation] Adversarial AutoAugment : RL

Generalization

• Identity Crisis: Memorization and Generalization Under Extreme Overparameterization: FC layers. Low level features show identity ability, while high-level features output constant information.

Representation Learning

• Target-Embedding Autoencoders for Supervised Representation Learning : VAE usually reconstracts X, but in this paper, it reconstracts Y(Y is high-dimensional). Application: multivariate sequence forecasting

• ⭐️⭐️⭐️⭐️⭐️ [Unsupervised clustering] Self-labelling via simultaneous clustering and representation learning [PyTorch]: random generated labels & optimization & update labels (representation learning) + linear program solving

• [Interpretation] Rotation-invariant clustering of neuronal responses in primary visual cortex

• [Multi-view information] Learning Robust Representations via Multi-View Information Bottleneck

Domain Adaptation

• Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification : re-identification, mean teacher, 4 models, triplet loss

• Domain Adaptive Multibranch Networks

Multi-modal (image/text)

• Latent Normalizing Flows for Many-to-Many Cross-Domain Mappings [PyTorch]: image caption
• From Inference to Generation: End-to-end Fully Self-supervised Generation of Human Face from Speech : using voice to generate face, interesting.

Graph Neural Network

• Deep Graph Matching Consensus [PyTorch]: Idea (neighboors should contain same information in two similar images) is good, which might be used in somewhere else.
• DropEdge: Towards Deep Graph Convolutional Networks on Node Classification [PyTorch]

Self-supervised Learning

• ⭐️⭐️⭐️⭐️ A crical analysis of self-supervision, or what we can learn from a single image : Self supervision with one image could learn low-level features with high quality

Semi-supervised Learning

• [Application] Automatically Discovering and Learning New Visual Categories with Ranking Statistics [PyTorch]: 应用流也能中ICLR了… self-supervision + supervised-learning + pesudo label + incremental learning

• Deep Semi-Supervised Anomaly Detection

• [Remove noise labelled data to semi-supervised learning] SELF: Learning to Filter Noisy Labels with Self-Ensembling

• [Mixmatch upgrade] ReMixMatch: Semi-Supervised Learning with Distribution Matching and Augmentation Anchoring

• [Semi-supervised] DivideMix: Learning with Noisy Labels as Semi-supervised Learning: filter noise label, then trest them as unlabelled data, apply pseudo label for mixmatch.

Weakly-supervised Learning

• Weakly Supervised Clustering by Exploiting Unique Class Count: predict the class No. within one images.

Active Learning

• [certainty and diversity] Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds [Oral]

Disentangle

• Weakly Supervised Disentanglement with Guarantees

Noise Lable

• [Remove noise labelled data to semi-supervised learning] SELF: Learning to Filter Noisy Labels with Self-Ensembling

• [Regularization] Simple and Effective Regularization Methods for Training on Noisily Labeled Data with Generalization Guarantee [Code]

• [Semi-supervised] DivideMix: Learning with Noisy Labels as Semi-supervised Learning: filter noise label, then trest them as unlabelled data, apply pseudo label for mixmatch.

Network Architecture

• [Semantic Segmentation] FasterSeg:Searching for Faster Real-time Semantic Segmentation [PyTorch]: network search + teacher/student knowledge distilation

Network Component Upgrade

• [Activation] Enhancing adversarial defence by k-winners-take-all [PyTorch]: For improving the robustness, similar to ReLU but with a ratio to keep the values instead of using threshold 0.

• [Curriculum Loss] Curriculum Loss: Robust Learning and Generalization against Label Corruption

• [Normalization] Mixup Inference: Better Exploiting Mixup to Defend Adversarial Attacks [PyTorch]: Adding permutations to the samples to improve the robustness.

• [Optimization] Don’t Use Large Mini-batches, Use Local SGD multi-GPU, before communication, more inference…

• [Optimization] Large Batch Optimization for Deep Learning: Training BERT in 76 minutes [TensorFlow]

Meta-Learning

• Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distribution Tasks [Code][Oral]

• A Meta-Transfer Objective for Learning to Disentangle Causal Mechanisms

Point Cloud

• Unpaired Point Cloud Completion on Real Scans using Adversarial Training [TensorFlow]: GAN + super-resolution

Mobile Network

• [Model Compression] Once for All: Train One Network and Specialize it for Efficient Deployment [PyTorch]