Deep Learning Papers
This is a list of deep learning papers distilled by YSDA students @ Yandex Research seminar and anyone else who wishes to contribute. Each paper is described with a simple digest highlighting what this paper actually does.
(Template) Deep Reinforcement Learning that Matters [arxiv] [code(n/a)]
Peter Henderson et al., digest by Template Author Name, YYYY/MM/DD
TL;DR many deep RL papers are not reproducible: even a slightest implementation detail leads to drastic changes in model performance. Figure above: two runs of exactly the same code with different random seeds.
Main insight: do not trust a deep RL paper unless
- it agregates training over several random seeds;
- it reports all hyperparameters, evaluation methods;
- releases the implementation or has a public implementation that matches the reported numbers;
Furthermore, many RL papers are prone to dataset/environment selection to conceal the situations in which the algorithm performs poorly. Authors propose that instead you explicitly state the conditions where the specific algorithm is likely to result in a success and evaluate in all environments where such conditions are met.