Deep Learning
Deep Learning Subtopics
IMPORTANT INSTRUCTION: While Editing this Markdown make sure to switch off word wrapping. In case you are using GitHub GUI, use the No Wrap setting found in top right corner of text editor.
Domain Adaptation
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Contrastive Learning and Self-Training for Unsupervised Domain Adaptation in Semantic Segmentation (2021) | GDrive | Rohit | Author propose a contrastive learning approach that adapts category-wise centroids across domains. They extend the method with self-training, where they use a memory-efficient temporal ensemble to generate consistent and reliable pseudo-labels. |
| FixBi: Bridging Domain Spaces for Unsupervised Domain Adaptation (CVPR ‘21) | HackMD | Rohit | Propose a Unsupervised Domain Randomization method that effectively handles large domain discrepancies |
| Universal Domain Adaptation through Self-Supervision (NeurIPS 2020) | HackMD | Rohit | Author propose a universally applicable domain adaptation framework that can handle arbitrary category shift, called Domain Adaptative Neighborhood Clustering via Entropy optimization (DANCE) |
| Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation (ICML 2020) | HackMD | Rohit | Work tackles a practical setting where only a trained source model is available and investigates how we can effectively utilize such a model without source data to solve UDA problems |
| Unsupervised Multi-Target Domain Adaptation Through Knowledge Distillation (WACV 21) | HackMD | Rohit | Paper propose a novel unsupervised MTDA approach to train a CNN that can generalize well across multiple target domains. |
| Consistency Regularization with High-dimensional Non-adversarial Source-guided Perturbation for Unsupervised Domain Adaptation in Segmentation (AAAI ‘21) | HackMD | Akshay | They propose a bidirectional style-induced DA method (BiSIDA) that employs consistency regularization to efficiently exploit information from the unlabeled target dataset using a simple neural style transfer model. |
| Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR ‘21) | HackMD | Akshay | They propose ProDA which resorts to prototypes to online denoise the pseudo-labels and learn a compact target feature space. Using knowledge distillation to a self-supervised pretrained model further boosts the performance. |
| Discover, Hallucinate, and Adapt: Open Compound Domain Adaptation for Semantic Segmentation (NeurIPS ‘20) | HackMD | Akshay | This work investigates open compound domain adaptation (OCDA) for semantic segmentation which deals with mixed and novel situations at the same time. They first cluster the compound target data based on style (discover), then hallucinate multiple latent target domains in source using image translation, and perform target-to-source alignment separately between domains (adapt). |
| Domain Adaptive Semantic Segmentation Using Weak Labels (ECCV ‘20) | HackMD | Akshay | This paper proposes a framework for Domain Adaptation (DA) in semantic segmentation with image-level weak labels in the target domain. They use weak labels to enable the interplay between feature alignment and pseudo-labeling, improving both in DA. |
| DACS: Domain Adaptation via Cross-domain Mixed Sampling (WACV ‘21) | HackMD | Akshay | This paper proposes Domain Adaptation via Cross-domain Mixed Sampling which mixes images from two domains along with their corresponding labels. These mixed samples are trained on, along with the labelled data itself. |
| Learning Texture Invariant Representation for Domain Adaptation of Semantic Segmentation (CVPR ‘20) | HackMD | Akshay | This paper uses style transfer to enforce texture invariance in the model, followed by self training to adapt to the target domain texture for the semantic segmentation task. |
| Unsupervised Intra-domain Adaptation for Semantic Segmentation through Self-Supervision (CVPR ‘20 Oral) | HackMD | Akshay | This paper proposes a two-step self-supervised DA approach to minimize the inter-domain and intra-domain gap together. |
| Unsupervised Domain Adaptation with Residual Transfer Networks (NIPS ‘16) | HackMD | Akshay | A domain adaptation approach that can jointly learn adaptive classifiers and transferable features from labeled data in the source domain and unlabeled data in the target domain. |
| Phase Consistent Ecological Domain Adaptation (CVPR ‘20) | HackMD | Akshay | This paper introduces 2 criteria to regularize the optimization involved in UDA: (1) the map between 2 image domains should be phase-preserving and (2) to leverage regularities in the scene, regardless of the illuminant or imaging sensor. |
| FDA: Fourier Domain Adaptation for Semantic Segmentation (CVPR ‘20) | HackMD | Akshay | A simple method for UDA where the discrepancy between the source and target distributions is reduced by swapping the low-frequency spectrum of one with the other. |
| Domain Adaptation for Structured Output via Discriminative Patch Representations (ICCV ‘19) | HackMD | Akshay | This paper proposes a UDA approach that explicitly discovers many modes in the structured output space of semantic segmentation to learn a better discriminator between the 2 domains, ultimately leading to a better domain alignment. |
Semantic Segmentation
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Semi-Supervised Semantic Segmentation with Cross-Consistency Training (CVPR ‘20) | HackMD | Akshay | This paper proposes cross-consistency training, where an invariance of the predictions is enforced over different perturbations applied to the outputs of the encoder (in a shared encoder and multiple decoder architecture). |
| Gated-SCNN: Gated Shape CNNs for Semantic Segmentation (ICCV ‘19) | HackMD | Akshay | This paper presents a 2-stream CNN i.e. one stream is normal CNN (classical stream) while the other is a shape stream, which explicitly processes shape information in a separate stream. |
| ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation | HackMD | Akshay | This paper presents a network architecture which is faster and more compact, for low real-time inference times. |
| W-Net: A Deep Model for Fully Unsupervised Image Segmentation | HackMD | Akshay | This paper presents fully unsupervised semantic segmentation using deep networks and a soft version of Normalized Cut. |
| Understanding Deep Learning Techniques for Image Segmentation | HackMD | Akshay | This paper aims to provide an intuitive understanding of significant DL-based approaches to segmentation. |
| Recent progress in semantic image segmentation | HackMD | Akshay | This paper presents a review on semantic segmentation approaches - traditional as well as DL-based. |
Knowledge Distillation
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Distilling the Knowledge in a Neural Network (NIPS ‘14W) | HackMD | Raj | This paper is the first DL approach to transfer knowledge from a teacher network to a student network, and uses softened outputs of the teacher network for training the student network. |
| A Gift from Knowledge Distillation: Fast Optimization, Network Minimization and Transfer Learning (CVPR ‘17) | HackMD | Akshay | This paper formulates the knowledge to be transferred in terms of flow between layers, calculates it as the inner product between feature maps from 2 layers, and uses this for Knowledge Distillation. |
| Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer (ICLR ‘17) | HackMD | Akshay | This paper defines attention for CNNs, and uses it to improve the performance of a student CNN network by forcing it to mimic the attention maps of a powerful teacher network. |
| Active Mixup for Data-Efficient Knowledge Distillation from a Blackbox Model (CVPR ‘20) | HackMD | Akshay | This paper proposes to blend active learning (Gissin and Shalev-Shwartz, 2019) and image mixup (Zhang et. al. 2017) to tackle data-efficient knowledge distillation from a blackbox teacher model. |
| Data-Free Learning of Student Networks (ICCV ‘19) | HackMD | Akshay | The pre-trained teacher network is considered as a fixed discriminator and a generator generates training samples which can obtain maximum response from the discriminator. Simultaneously, a smaller network is trained using the generated data and the teacher network. |
Active Learning
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Cost-Effective REgion-based Active Learning for Semantic Segmentation (BMVC ‘18) | HackMD | Akshay | This paper introduces an active learning strategy for semantic segmentation that uses an information measure and an annotation cost estimate. |
| Variational Adversarial Active Learning (ICCV ‘19) | HackMD | Akshay | This paper introduces a pool-based active learning strategy which learns a low dimensional latent space from labeled and unlabeled data using a VAE. |
Feature Detection and Description
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Patch2Pix: Epipolar - Guided Pixel - Level Correspondences | HackMD | Arihant, Saurabh | This paper proposes a new method on determining pixel-level correspondences in a detect - to - refine manner. It follows a weakly supervised learning approach, guided by the epipolar geometry of input image pair. |
| Neighbourhood Consensus Networks (NeurIPS ‘18) | HackMD | Arihant, Saurabh | The paper proposes an end - to - end trainable CNN architecture identifying consistent matches by analysing neighbourhood consensus patterns. The paper also demonstrates the use of weak supervision as matching and non - matching pairs, rather than using manual annotations. |
| D2 Net - A Trainable CNN for Joint Description and Detection of Local Features (CVPR ‘19) | HackMD | Aniket Gujarathi | This paper introduces a Deep Learning based approach to solve the problem of local features detection and description using the detect-and-describe approach instead of the traditionally used detect-then-describe approach. |
Unsupervised Learning
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Deep Clustering for Unsupervised Learning of Visual Features (ECCV 2018) | HackMD | Rohit | A clustering method that jointly learns the parameters of a neural network and the cluster assignments of the resulting features |
| Augmented Autoencoders: Implicit 3D Orientation Learning for 6D Object Detection (ECCV ‘18) | HackMD | Aayush, Jayesh, Saketh | This paper presents a real-time RGB-based pipeline for object detection and 6D pose estimation, based on a variant of denoising autoencoder, which is an augmented encoder trained on views of a 3D model using domain randomization. |
| A Simple Framework for Contrastive Learning of Visual Representations (ICML 2020) | HackMD | Pulkit | This paper provides a simpler and a more efficient way of Contrastive Self-Supervised Learning without using specialised architectures. The paper uses design methods from various architectures and incorporates them to achieve SOTA performance on Representation Learning. |
Object Detection
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Fast R - CNN (ICCV ‘15) | HackMD | Arihant, Saketh, Siddharth | This paper is an extension of R - CNN, with 213 times faster test time than R - CNN and 10 times faster than SPPNet. |
| Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition (TPAMI ‘15) | HackMD | Arihant, Saketh, Siddharth | The paper proposes a workaround to feeding a fixed size input to CNNs. Resizing the input can lead to reduction in recognition accuracy for images/sub - images. Cropping and resizing often result in unwanted geometric distortions. The authors add a ‘spatial pyramid pooling’ layer after convolution layers to remove the fixed size constraint to network. |
| OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks (ICLR ‘14) | HackMD | Jayesh, Saketh | This paper presents a framework for classification, localization and detection of objects using a multiscale and sliding window approach. It can do mutiple tasks using a single shared network. Second important findings of this paper is explaining how ConvNets can be effectively used for detection and localization tasks. |
| Rich feature hierarchies for accurate object detection and semantic segmentation (CVPR ‘14) | HackMD | Jayesh, Saketh | This paper proposes a framework that handles object detection task in two steps, first being generation of region proposals in order to localize and segment objects and the second steps is about classifying these objects.This method improves mean average precision (mAP) by more than 30% relative to the previous best result on VOC 2012 achieving a mAP of 53.3%. |
| You Only Look Once: Unified, Real-Time Object Detection | GDrive | Rohit | Most popular object detection algorithm. It frames object detection as a regression problem to spatially separated bounding boxes and associated class probabilities. |
Curriculum learning
| Paper | Notes | Author | Summary |
|---|---|---|---|
| When do curricula work? (ICLR ‘21) | HackMD | Sharath | This paper conducted a large scale study on curriculum learning methods for supervised learning setting and made some interesting comments on when curricula is effective. |
Bayesian Neural Networks
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Radial Bayesian Neural Networks: Beyond Discrete Support In Large-Scale Bayesian Deep Learning (AISTATS ‘20) | HackMD | Sharath | This paper studies a famous soap bubble problem in high-dimensional probability spaces and how Mean field variational inference sufferes from this. As a work around for this, the paper proposes a new posterior approximation in hyperspherical coordinate system. |
Causality
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Recurrent Independent Mechanisms (ICLR’ 21) | HackMD | Sharath | This paper proposes a new recurrent architecture which takes into account the modularity and independence and shows how this helps in generalisation. |
Anomaly Detection
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Abnormal Event Detection in Videos using Spatio Temporal Autoencoder | HackMD | Raj | This paper proposes a new architecture for anomaly detection in videos. Their architecture includes two main components one for spatial feature representation, and one for learning the temporal evolution of these spatial features. |
Generative Adversarial Nets
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Generative Adversarial Nets(Neurips ‘14) | HackMD | Raj | This paper proposes a novel “adversarial method” for data generation. It is now considered as one of the classics in deep learning. |
| Disentangled Inference for GANs with Latently Invertible Autoencoder | HackMD | Vignesh | This paper proposes a novel generative model named Latently Invertible Autoencoder (LIA) which tackles the entanglement problem(which often occurs in GANs) and generates high quality images from disentangled latents. |
Recurrent Networks
| Paper | Notes | Author | Summary |
|---|---|---|---|
| Opening the Black Box: Low-Dimensional Dynamics in High-Dimensional Recurrent Neural Networks | HackMD | Sharath | This paper analyses the dynamics of RNNs leveraging some dynamical systems theory. |
Real-world DL applications
| Paper | Notes | Author | Summary |
|---|---|---|---|
| FaceNet: A Unified Embedding for Face Recognition and Clustering | HackMD | Muhammed Abdullah | This paper introduces the famous Triplet Loss function to training deep networks for face verification, recognition and clustering. It also achieved the state-of-the-art accuracy. |
| On Interpretability of Deep Learning based Skin Lesion Classifiers using Concept Activation Vector (IJCNN ‘20) | HackMD | Arihant, Prasad, Saketh | This paper states that the current methods for computer - aided diagnosis are not widely accepted due to their obscure nature. The main aim of this paper is to design a deep learning model that will be trained to make decisions similar to medical experts. Concept Activation vectors are used to map human understandable concepts to RECOD images. The results shows that the classifier learns and encodes human understandable concepts in its latent representation. |
| Concept Learning with Energy - Based Models | HackMD | Arihant, Saketh | Multiple aspects or hallmarks of human intelligence require the ability to convert experience into concepts. In this paper the aspect of representing these concepts in the form of an energy function is proposed. The entire framework is evaluated on learning visual, quantitative, relational, and temporal concepts in an unsupervised way. |