Paper Reading List (Updated Regularly)

Metric Learning

1. SoftTriple Loss: Deep Metric Learning Without Triplet Sampling
   • Proposes a method that performs classification and metric learning with a single loss. It was proven that smoothing the triplet loss leads to cross entropy (they are essentially the same).
2. Visual Explanation for Deep Metric Learning
   • Visualization of metric learning models
3. Embedding Expansion: Augmentation in Embedding Space for Deep Metric Learning
   • By applying several simple rule-based optimizations to mining, accuracy improvements were confirmed across all distance loss functions.
4. Moving in the Right Direction: A Regularization for Deep Metric Learning
   • Compares regularization techniques for deep metric learning and discusses the risks of triplet loss.
5. Deep Metric Learning via Adaptive Learnable Assessment
   • Replaces mining rules with a learning-based approach and adopts an episode-based training scheme.

Video Tasks

1. Spatiotemporal Contrastive Video Representation Learning
   • Applies SimCLR to video classification tasks; would like to incorporate this.
2. Predicting Video with VQVAE
   • Achieves 65% on Kinetics-600; a teacher_forcing-like approach is possible.
3. Is Space-Time Attention All You Need for Video Understanding?
   • A Transformer-based video classifier with many novel aspects.
4. VideoMix: Rethinking Data Augmentation for Video Classification
   • Proposes VideoMix, a new data augmentation for video action recognition. The T-VideoMix method seems applicable.
5. TSM: Temporal Shift Module for Efficient Video Understanding
   • Addresses the problem of 3D CNNs being too heavy by inserting a TSM module into 2D CNNs as a substitute. TSM has zero parameters, so the complexity remains that of a 2D CNN.

Future Prediction

1. Improved Conditional VRNNs for Video Prediction
   • Predicts unknown video frames using a Variational Recurrent Autoencoder. A typical RAE that is very simple -- the go-to approach for generation.
2. Video Prediction via Example Guidance
   • Not finished reading; the first multimodal model for video future prediction.
3. Predictive Learning: Using Future Representation Learning Variational Autoencoder for Human Action Prediction
   • Two-stream approach using RGB and Optical Flow.

Training Methods

1. Invariant Information Clustering for Unsupervised Image Classification and Segmentation
   • An unsupervised training method that can directly output predictions; robust to noise.
2. Supervised Contrastive Learning
   • Performs supervised learning based on SimCLR.
3. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments
   • SwAV
4. A Simple Framework for Contrastive Learning of Visual Representations
   • SimCLR
5. AutoAugment: Learning Augmentation Policies from Data
   • Performs data augmentation using a learning-based approach.
6. What Makes Training Multi-modal Classification Networks Hard?

Others

1. Revisiting ResNets: Improved Training and Scaling Strategies
   • Training and scaling strategies for ResNet.
2. An annotation-free whole-slide training approach to pathological classification of lung cancer types using deep learning
   • Avoids image downscaling by using a Unified Memory (UM) mechanism and several GPU memory optimization techniques.
3. Prototypical Contrastive Learning of Unsupervised Representations
   • EM algorithm-based clustering that modifies the distance function to make clusters harder to converge, thereby suppressing overfitting.