This repository is a collection of 70 short summaries of papers in Deep Learning. Summaries are based on papers which explore novel groundbreaking ideas or consist of theoretically rich concepts. Each short summary was typically added each week which explores essential aspects of the work, its technical innovation and new questions and ideas raised by the work. Length of each summary is 1 page. Each summary is based on a fixed set of guidlines which are given here.
Note- At the time of writing a sample, I had no relation to the authors or their respective organizations. Opinions and claims expressed in the text are solely my own.
| Summary Number | Paper Title | Author List | Summary Link |
|---|---|---|---|
| 1 | Action and Perception as Divergence Minimization | Danijar Hafner,Pedro A. Ortega,Jimmy Ba,Thomas Parr,Karl Friston,Nicolas Heess | link |
| 2 | Momentum Contrast for Unsupervised Visual Representation Learning | Kaiming He,Haoqi Fan,Yuxin Wu,Saining Xie,Ross Girshick | link |
| 3 | When to use parametric models in reinforcement learning? | Hado van Hasselt, Matteo Hessel, John Aslanides | link |
| 4 | Data-Efficient Image Recognition With Contrastive Predictive Coding | Olivier J. Henaff,Aravind Srinivas,Jeffrey De Fauw,Ali Razavi,Carl Doersch,S. M. Ali Eslami,Aaron van den Oord | link |
| 5 | A Learning Algorithm for Boltzmann Machines | David H. Ackley,Geoffrey E. Hinton,Terrence J. Sejnowski | link |
| 6 | Dependence Measures Bounding the Exploration Bias for General Measurements | Jiantao Jiao,Yanjun Han,Tsachy Weissman | link |
| 7 | On Variational Bounds of Mutual Information | Ben Poole,Sherjil Ozair,Aaron van den Oord,Alexander A. Alemi,George Tucker | link |
| 8 | Hindsight Credit Assignment | Anna Harutyunyan, Will Dabney, Thomas Mesnard, Nicolas Heess, Mohammad G. Azar, Bilal Piot, Hado van Hasselt, Satinder Singh, Greg Wayne, Doina Precup, Rémi Munos | link |
| 9 | Learning a Contact-Adaptive Controller for Robust, Efficient Legged Locomotion | Xingye Da, Zhaoming Xie, David Hoeller, Byron Boots, Animashree Anandkumar, Yuke Zhu, Buck Babich, Animesh Garg | link |
| 10 | Counterfactual Data Augmentation using Locally Factored Dynamics | Silviu Pitis, Elliot Creager, Animesh Garg | link |
| 11 | LEAF: Latent Exploration Along the Frontier | Homanga Bharadhwaj, Animesh Garg, Florian Shkurti | link |
| 12 | Cautious Adaptation For Reinforcement Learning in Safety-Critical Settings | Jesse Zhang, Brian Cheung, Chelsea Finn, Sergey Levine, Dinesh Jayaraman | link |
| 13 | Model-Based Reinforcement Learning with Value-Targeted Regression | Zeyu Jia, Lin F. Yang, Csaba Szepesvari, Mengdi Wang | link |
| 14 | Skill Transfer Via Partially Amortized Hierarchical Planning | Kevin Xie, Homanga Bharadhwaj, Danijar Hafner, Animesh Garg, Florian Shkurti | link |
| 15 | Evaluating Agents Without Rewards | Brendon Matusch, Jimmy Ba, Danijar Hafner | link |
| 16 | Conservative Safety Critics For Exploration | Homanga Bharadhwaj, Aviral Kumar, Nicholas Rhinehart, Sergey Levine, Florian Shkurti, Animesh Garg | link |
| 17 | Mastering Atari with Discrete World Models | Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, Jimmy Ba | link |
| 18 | Weighted QMIX: Expanding Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning | Tabish Rashid, Gregory Farquhar, Bei Peng, Shimon Whiteson | link |
| 19 | Continual Model-Based Reinforcement Learning with Hypernetworks | Yizhou Huang, Kevin Xie, Homanga Bharadhwaj, Florian Shkurti | link |
| 20 | Model-Based Inverse Reinforcement Learning from Visual Demonstrations | Neha Das, Sarah Bechtle, Todor Davchev, Dinesh Jayaraman, Akshara Rai, Franziska Meier | link |
| 21 | RODE: Learning Roles To Decomponse Multi-Agent Tasks | Tonghan Wang, Tarun Gupta, Anuj Mahajan, Bei Peng, Shimon Whiteson, Chongjie Zhang | link |
| 22 | The Act Of Remembering: A Study In Partially Observable Reinforcement Learning | Rodrigo Toro Icarte, Richard Valenzano, Toryn Q. Klassen, Phillip Christoffersen, Amir-massoud Farahmand, Sheila A. McIlraith | link |
| 23 | An Inductive Bias for Distances: Neural Nets that Respect the Triangle Inequality | Silviu Pitis, Harris Chan, Kiarash Jamali, Jimmy Ba | link |
| 24 | MAVEN: Multi-Agent Variational Exploration | Anuj Mahajan, Tabish Rashid, Mikayel Samvelyan, Shimon Whiteson | link |
| 25 | Visual Imitation Made Easy | Sarah Young, Dhiraj Gandhi, Shubham Tulsiani, Abhinav Gupta, Pieter Abbeel, Lerrel Pinto | link |
| 26 | “Other-Play” for Zero-Shot Coordination | Hengyuan Hu, Adam Lerer, Alex Peysakhovich, Jakob Foerster | link |
| 27 | Using Fast Weights to Attend to the Recent Past | Jimmy Ba, Geoffrey Hinton, Volodymyr Mnih, Joel Z. Leibo, Catalin Ionescu | link |
| 28 | γ-Models: Generative Temporal Difference Learning for Infinite-Horizon Prediction | Michael Janner, Igor Mordatch, Sergey Levine | link |
| 29 | Learning to Communicate with Deep Multi-Agent Reinforcement Learning | Jakob N. Foerster, Yannis M. Assael, Nando de Freitas, Shimon Whiteson | link |
| 30 | Bayesian Action Decoder for Deep Multi-Agent Reinforcement Learning | Jakob N. Foerster, H. Francis Song, Edward Hughes, Neil Burch, Iain Dunning, Shimon Whiteson, Matthew M. Botvinick, Michael Bowling | link |
| 31 | Improved Variational Inference with Inverse Autoregressive Flow | Diederik P. Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, Max Welling | link |
| 32 | Deep Reinforcement Learning from Self-Play in Imperfect-Information Games | Johannes Heinrich, David Silver | link |
| 33 | Variational Policy Gradient Method for Reinforcement Learning with General Utilities | Junyu Zhang, Alec Koppel, Amrit Singh Bedi, Csaba Szepesvari, Mengdi Wang | link |
| 34 | The Value-Improvement Path Towards Better Representations for Reinforcement Learning | Will Dabney, Andre Barreto, Mark Rowland, Robert Dadashi, John Quan, Marc G. Bellemare, David Silver | link |
| 35 | Fictitious Self-Play in Extensive-Form Games | Johannes Heinrich, Marc Lanctot, David Silver | link |
| 36 | Expected Eligibility Traces | Hado van Hasselt, Sephora Madjiheurem, Matteo Hessel, David Silver, André Barreto, Diana Borsa | link |
| 37 | Transfer in Deep Reinforcement Learning Using Successor Features and Generalised Policy Improvement | Andre Barreto, Diana Borsa, John Quan, Tom Schaul, David Silver, Matteo Hessel, Daniel Mankowitz, Augustin Zıdek, Remi Munos | link |
| 38 | A Theoretical and Empirical Analysis of Expected Sarsa | Harm van Seijen, Hado van Hasselt, Shimon Whiteson and Marco Wiering | link |
| 39 | Learning to Play No-Press Diplomacy with Best Response Policy Iteration | Thomas Anthony, Tom Eccles, Andrea Tacchetti, János Kramár, Ian Gemp, Thomas C. Hudson, Nicolas Porcel, Marc Lanctot, Julien Pérolat, Richard Everett, Roman Werpachowski, Satinder Singh, Thore Graepel and Yoram Bachrach | link |
| 40 | Neural Dynamic Policies for End-to-End Sensorimotor Learning | Shikhar Bahl, Mustafa Mukadam, Abhinav Gupta, Deepak Pathak | link |
| 41 | Learning By Cheating | Dian Chen, Brady Zhou, Vladlen Koltun, Philipp Krähenbühl | link |
| 42 | Your Classifier Is Secretly An Energy based Model Aand You Should Treat It Like One | Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Kevin Swersky, Mohammad Norouzi | link |
| 43 | No MCMC for me: Amortized sampling for fast and stable training of energy-based models | Will Grathwohl, Jacob Kelly, Milad Hashemi, David Duvenaud, Mohammad Norouzi, Kevin Swersky | link |
| 44 | Implicit Autoencoders | Alireza Makhzani | link |
| 45 | Neural Ordinary Differential Equations | Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, David Duvenaud | link |
| 46 | Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations | Vincent Sitzmann, Michael Zollhöfer, Gordon Wetzstein | link |
| 47 | PixelGAN Autoencoders | Alireza Makhzani, Brendan Frey | link |
| 48 | Density Estimation using Real NVP | Laurent Dinh, Jascha Sohl-Dickstein, Samy Bengio | link |
| 49 | NICE: Non-linear Independent Components Estimation | Laurent Dinh,David Krueger,Yoshua Bengio | link |
| 50 | Augmenting Physical Models with Deep Networks for Complex Dynamics Forecasting | Yuan Yin, Vincent Le Guen, Jérémie Dona, Emmanuel de Bezenac, Ibrahim Ayed, Nicolas Thome, Patrick Gallinari | link |
| 51 | Conservative Q-Learning for Offline Reinforcement Learning | Aviral Kumar, Aurick Zhou, George Tucker, Sergey Levine | link |
| 52 | Perceiver: General Perception with Iterative Attention | Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, Joao Carreira | link |
| 53 | On the mapping between Hopfield networks and Restricted Boltzmann Machines | Matthew Smart, Anton Zilman | link |
| 54 | Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients | Brenden K Petersen, Mikel Landajuela Larma, Terrell N. Mundhenk, Claudio Prata Santiago, Soo Kyung Kim, Joanne Taery Kim | link |
| 55 | Training Products of Experts by Minimizing Contrastive Divergence | Geoffrey E. Hinton | link |
| 56 | Linear Transformers Are Secretly Fast Weight Memory Systems | Imanol Schlag, Kazuki Irie, Jurgen Schmidhuber | link |
| 57 | Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images | Rewon Child | link |
| 58 | EigenGame: PCA as a Nash Equilibrium | Ian Gemp, Brian McWilliams, Claire Vernade, Thore Graepel | link |
| 59 | Disentangled Recurrent Wasserstein Autoencoder | Jun Han, Martin Renqiang Min, Ligong Han, Li Erran L, Xuan Zhang | link |
| 60 | VAEBM: A Symbiosis between Variational Autoencoders and Energy-based Models | Zhisheng Xiao, Karsten Kreis, Jan Kautz, Arash Vahdat | link |
| 61 | Learning Energy-Based Models by Diffusion Recovery Likelihood | Ruiqi Gao, Yang Song, Ben Poole, Ying Nian Wu, Diederik P Kingma | link |
| 62 | Improved Contrastive Divergence Training of Energy-Based Model | Yilun Du, Shuang Li, Joshua Tenenbaum, Igor Mordatch | link |
| 63 | Conjugate Energy-Based Models | Hao Wu, Babak Esmaeili, Michael L Wick, Jean-Baptiste Tristan, Jan-Willem van de Meent | link |
| 64 | Fully Unsupervised Diversity Denoising with Convolutional Variational Autoencoders | Mangal Prakash, Alexander Krull, Florian Jug | link |
| 65 | Bayesian Quadrature on Riemannian Data Manifolds | Christian Frohlich, Alexandra Gessner, Philipp Hennig, Bernhard Scholkopf, Georgios Arvanitidis | link |
| 66 | High-Dimensional Bayesian Optimization via Nested Riemannian Manifolds | Noémie Jaquier, Leonel Rozo | link |
| 67 | Risk-Averse Offline Reinforcement Learning | Núria Armengol Urpí, Sebastian Curi, Andreas Krause | link |
| 68 | Provably Good Batch Reinforcement Learning Without Great Exploration | Yao Liu, Adith Swaminathan, Alekh Agarwal, Emma Brunskill | link |
| 69 | LILA: Language-Informed Latent Actions | Siddharth Karamcheti, Megha Srivastava, Percy Liang, Dorsa Sadigh | link |
| 70 | Tesseract: Tensorised Actors for Multi-Agent Reinforcement Learning | Anuj Mahajan, Mikayel Samvelyan, Lei Mao, Viktor Makoviychuk, Animesh Garg, Jean Kossaifi, Shimon Whiteson, Yuke Zhu, Animashree Anandkumar | link |
This section outlines the guidlines which are used to write summaries for the repository. Note that these guidlines must be strictly followed for providing high quality summaries to the reader.
Since Deep Learning is a fast-moving field, papers can span a broad variety of topics. In order to shorten the range of literature, papers must be selected using the following rules-
- Any review paper, survey or a long article should not be considered since these are themselves a review of previous works and reviewing them would defeat the purpose of literature writing.
- Short papers, Journal papers and theoretical works are suitable as these present a single idea which may be of interest to the reader.
- Papers containing experimental results balanced with theory are encouraged as these validate the practical applications of the proposed methods.
- There is no constraint on the publication date and time of papers. However, papers which date back to the 'AI Winter' are highly encouraged since these provide insights which were never looked at for a very long time.
- Papers containing simply applications of pre-existing methods to various regimes are not preferred since they do not provide any new insights into the algorithm itself and only deal with its applicability.
- Workshop papers and incomplete works are also welcome as researchers can always build on these ideas.
- While there is no specific restriction on the content of papers, writer must consider fields in AI which are growing and require a new outlook, eg. Reinforcement Learning, Gradient-Free methods, Meta-Learning, Natural Language at Scale, Explainability, etc.
This section provides points on writing a good introduction for the summary. Note that these points are not strict and only serve as a guiding principle for drafting a good introduction.
- The introduction should be a high-level idea of the paper and what the work deals with.
- The writer must refrain from going into any technical detail and highlight the broad idea of the work and its scope.
- No mathematical terms, definitions, technical explanations or algorithmic details should be provided to the reader.
- The main focus should be on the problem statement and how the method aims to solve this.
- Key takeaway- Detail is your enemy!
This section deals with the proposed method and its essential aspects in solving the problem highlighted in the introduction section. Following points serve as a guide to writing this section-
- The content must provide a brief overview of the method. This informs the reader about what he is getting into.
- Once an overview has been provided, the draft can start diving into the detail which should be highlighted intuitively.
- Mathematical details must be followed by words, complicated terminology must be explained intuitively using examples or instances from work.
- Reasons related to technical details and there usage must be provided to the reader. The whole point of reading a summary is to crisply go over the details of the paper without wanting to read the entire text.
- While the draft should highlight the method and its details, it should also provide the reader with insights from the writer's point of view. These could consist of specific reasons for selecting a set of parameter values, usage of a specific technique existing in literature, novel contributions and the reason behind its usage and any improvements/changes from previous works.
- Key takeway- Intuition is king!
The critical analysis section deals with the writer's analysis and understanding of the method. Specifically, this section should highlight what the writer thinks about the proposed approach, its strengths, weaknesses and potential areas of improvement and applicability. This is the most important section of the summary since the writer needs to critically evaluate and comment on the technical aspects of the work. The following points should be kept in mind while writing a good critical analysis-
- One should evaluate all aspects of the work before diving into this section. A comprehensive analysis of the work is essential as it builds on the writer's understanding of the method.
- Once the complete details have been established, a proper structure for evaluation should be constructed. Typically, this structure should consist of comments on motivational aspects, strengths, weaknesses, novel contributions, applicability of the method, extensions and improvements of the method's components and their shortcomings.
- The draft should be written in a critical yet formal manner. For instance, the writer must highlight the intriguing aspects of the algorithm and at the same time throw light on its critical parts in a decorous manner.
- Writer's comments should be their own and not focus on explaining the work. The main idea behind this section is to present your understanding of the text to the reader, not the text itself.
- Key takeway- Your own contributions matter!
This section follows and builds upon the critical analysis section. The main idea behind this section is to highlight and examine the novel contributions of the proposed method and their contributions towards improving prior techniques. Another reason to study new ideas proposed by a paper is to identify potential directions of research and answer open questions from different perspectives. Following points may come in handy while writing this section-
- One should concisely summarize the novel contributions of the work (not more than 1-2 sentences).
- The summary should be followed by a critical analysis or a set of possible questions which maybe asked while using the novel aspects of the method.
- Since the focus is on novelty, writer may also be interested in asking questions related to other potential techniques and their applicability in place of the novel method.
- The section should conclude with brief comments on substitutes/extensions to the novel components and questions left unanswered (if any) in the work.
- Key takeway- Questions, Questions, Questions!
Lastly, the conclusions section should clearly and concisely sum up the summary of the work. This section should consist of a crisp summary of all the previous sections and only highlight their most important aspects. The content of this section should be directed to someone who does not have enough time to read the summary and only wishes to grab the essential points. A good conlusion could be written using the following points-
- The summary should start with a high level introductory note of the method, its usage and contributions. (not more 1-2 sentences)
- The summary should explore the novelty of the method, its implications and the resulting outcome achieved from the writer's point of view.
- The summary should walk over the key components of writer's evaluation and his comments on the method.
- The summary should conclude with a brief note of new ideas/questions introduced by the work and potential directions for future work.
- Key takeway- Just be done already!