Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised State Representation Learning in Atari
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use spatiotemporal patch-level InfoNCE objectives | Atari state representation learning · AtariARI, random-agent data | mean probe F1: 0.61 -> 0.68 | Helped |
| use spatiotemporal patch-level InfoNCE objectives | Atari state representation learning · AtariARI, PPO-agent data | mean probe F1: 0.56 -> 0.62 | Helped |
| add spatial contrastive constraints to temporal contrastive learning (ablation) | ST-DIM · AtariARI, random-agent data | mean probe F1: 0.55 -> 0.68 | Helped |
| use InfoNCE instead of JSD loss (ablation) | ST-DIM · AtariARI, random-agent data | mean probe F1: 0.58 -> 0.68 | Helped |
| retain temporal context rather than sample negatives from a different game (ablation) | ST-DIM · AtariARI, random-agent data | mean probe F1: 0.26 -> 0.68 | Helped |
| add spatial contrastive constraints to temporal contrastive learning (ablation) | ST-DIM · AtariARI, random-agent data, Small Loc. | probe F1: 0.37 -> 0.51 | Helped |
| use InfoNCE instead of JSD loss (ablation) | ST-DIM · AtariARI, random-agent data, Small Loc. | probe F1: 0.44 -> 0.51 | Helped |
| retain temporal context rather than sample negatives from a different game (ablation) | ST-DIM · AtariARI, random-agent data, Small Loc. | probe F1: 0.18 -> 0.51 | Helped |
| use spatiotemporal patch-level contrastive learning rather than global temporal contrast (ablation) | Atari state representations · Boxing | player_score probe F1: 0.45 -> 0.88 | Helped |
| use spatiotemporal patch-level contrastive learning rather than global temporal contrast (ablation) | Atari state representations · Boxing | enemy_x probe F1: 0.17 -> 0.51 | Helped |
| use ST-DIM rather than global temporal contrast (ablation) | Atari state representations · Boxing | enemy_score probe F1: 0.74 -> 0.70 | Hurt |
| use ST-DIM rather than global temporal contrast (ablation) | Atari state representations · Boxing | clock probe F1: 0.81 -> 0.92 | Helped |
Closest papers by meaning
- Learning to Play Atari in a World of Tokens (ICML 2024 · similarity 0.87)
- Pretraining Representations for Data-Efficient Reinforcement Learning (NeurIPS 2021 · similarity 0.87)
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.87)
- Provable Representation Learning for Imitation with Contrastive Fourier Features (NeurIPS 2021 · similarity 0.86)
- On the Importance of Feature Decorrelation for Unsupervised Representation Learning in Reinforcement Learning (ICML 2023 · similarity 0.86)
- Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.