A collections of must-read AI-related papers.
I read a lot of papers that now I think I should keep track of them. In this list, I list papers that I think are important in AI/ML/DL (and even other topics).
Disclaimer: the list is totally opinion-based and does not mean "you should read X and should not read Y". Mostly, the papers are rated based on the idea and impact rather than the SOTA leaderboard and the conference/journal they are published in. Yes, number of citations is not rated.
Note: the year is noted as the lastest version on arXiv/OpenReview.
SUGGESTIONS ARE WELCOME. Create a pull request, issue, or drop me a message on LinkedIn or email (link on my main profile). Note that I do not promote papers, so I will not response to messages like "Hey my method is novel, can you add it?". I am open to discussions and collaborations.
- Deep Reinforcement Learning that Matters (Henderson et al., 2019): an important paper in RL. It compares algorithms using multiple seeds (I mean, multiple), and gives advices how to conduct RL research and experiments.
- Implementation Matters in Deep RL: A Case Study on PPO and TRPO (Engstorm et al., 2020): a study how code-level optimization in deep RL can significantly change results.
- How Many Random Seeds? Statistical Power Analysis in Deep Reinforcement Learning Experiments (Colas et al., 2018): how many seeds you use to express statistical significance in RL? 5? 10? 20? This paper explains why, from a statistical view.
- RoFormer: Enhanced Transformer with Rotary Position Embedding (Su et al., 2021): a rotary positional encoding technique that researchers at EleutherAI found a significant improvement in performance.
- On the Origin of Deep Learning (Wang and Raj, 2017): ever wonder when, why and how DL emerges? This one got you covered!
- Co-Mixup: Saliency Guided Joint Mixup with Supermodular Diversity (Kim et al., 2021): a novel mix-up technique that creates new data in batches instead of a pair of images.
- Combining Ensembles and Data Augmentation Can Harm Your Calibration (Wen et al., 2021): experiments show that combining ensembles and augmentations can decrease your performance.
- What Matters In On-Policy Reinforcement Learning? A Large-Scale Empirical Study (Andrychowicz et al., 2020): advices about how to choose hyperparameters for PPO on continuous control task. Although useful, but I think this paper is limited to the tasks it evaluates.
- Action Space Shaping in Deep Reinforcement Learning (Kanervisto et al., 2020): how to shape your action space? Should it
(1, 6)
or(2, 3)
or(3, 2)
? - A Closer Look at Invalid Action Masking in Policy Gradient Algorithms (Huang and Ontañón, 2020): usually we give a big negative reward if the agent makes an invalid action (for example, a chess piece moves left and gets out of the board), but that is totally a naive workaround: say, doing so may make the agent thinks that action is very bad (moving left) instead of invalid (out of the board).
- An Empirical Model of Large-Batch Training (also blog) (McCandlish et al., 2018): a paper that surveys and explains large batch size training in AI. Batch size speeds up training and reduces variance, so this is a must-read one!
- Writing Great Reward Functions - Bonsai (Bonsai, 2017): reward function plays an imporant role on how the agent learns. Be careful what you give them!
- Critizing Neural Architecture Search (NAS) and explaining why Randomness just performs equally good.
- Random Search and Reproducibility for Neural Architecture Search (Li and Talwalkar, 2019),
- Pitfalls in Machine Learning Research: Reexamining the Development Cycle (Biderman and Scheirer, 2020),
- Local Search is a Remarkably Strong Baseline for Neural Architecture Search (Ottelander et al., 2020)
- The Risk of Invariant Risk Minimization: why IRM fails to deliver what it promises.
- In Search of Lost Domain Generalization: a comprehensive benchmark of domain generalization algorithm. Overall, ERM works good in practice.
- Descending through a Crowded Valley: a comprehensive benchmark of optimizers and LR schedulers to support that Adam is still very competitive.
- Simple random search provides a competitive approach to reinforcement learning: you don't need complicated algorithm. Randomness is enough!
- Competing in the Obstacle Tower Challenge: A very interesting blog that tells why entropy should not be included in the loss calculation, why RNN-based network fails in RL, and more!
- Deep Reinforcement Learning Doesn't Work Yet: why RL is so damn hard? This blog explains possible reasons why RL does not work yet, despite breaking news about RL in recent years.
- Approximating KL Divergence: a short blog explains why in RL, approx KL divergence is enough, instead of calculating true KL.
- Machine Learning in Physics resources (ChrisRackauckas): an extremely detailed collection of resources about ML in Physics. I am overwhelmed by his effort typing this answer!
- Lessons Learned Reproducing a Deep Reinforcement Learning Paper (Amid Fish): a blog post told how the author tried to reproduce a paper's result, and turned out it was (A LOT) harder than he expected.
- A collection of failed RL algorithms
- Visualizing the Attention mechanism
- BOLA: Near-Optimal Bitrate Adaptation for Online Videos: determine the near-optimal video quality based on your network speed!
- How to Design for Color Blindness