Research map / Reinforcement learning / Visual and representation RL
On the Importance of Feature Decorrelation for Unsupervised Representation Learning in Reinforcement Learning
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 |
|---|---|---|---|
| add feature decorrelation loss to temporally predictive pretraining | SimTPR · Atari100k, pretrained from state | IQM HNS: 0.353 -> 0.451 | Helped |
| pretrain from demonstrations with action labels | SimTPR · Atari100k, pretrained from demonstration | IQM HNS: 0.451 -> 0.500 | Helped |
| add contrastive loss instead of using no collapse prevention (ablation) | SimTPR · Atari100k, collapse prevention comparison | IQM HNS: 0.189 -> 0.331 | Helped |
| add batch normalization for collapse prevention (ablation) | SimTPR · Atari100k, collapse prevention comparison | IQM HNS: 0.189 -> 0.068 | Hurt |
| add decorrelation loss (ablation) | SimTPR without collapse prevention · Atari100k, collapse prevention comparison | IQM HNS: 0.189 -> 0.451 | Helped |
| add contrastive loss alongside decorrelation loss (ablation) | SimTPR · Atari100k, collapse prevention comparison | IQM HNS: 0.451 -> 0.342 | Hurt |
| add batch normalization alongside decorrelation loss (ablation) | SimTPR · Atari100k, collapse prevention comparison | IQM HNS: 0.451 -> 0.264 | Hurt |
| use causal transformer rather than non-causal masked reconstruction at mask ratio 0.3 (ablation) | SimTPR · Atari100k, transition model | IQM HNS: 0.270 -> 0.451 | Helped |
| use causal transformer rather than non-causal masked reconstruction at mask ratio 0.5 (ablation) | SimTPR · Atari100k, transition model | IQM HNS: 0.385 -> 0.451 | Helped |
| use causal transformer rather than non-causal masked reconstruction at mask ratio 0.7 (ablation) | SimTPR · Atari100k, transition model | IQM HNS: 0.367 -> 0.451 | Helped |
| replace causal transformer with a 2-layer GRU (ablation) | SimTPR · Atari100k, transition model | IQM HNS: 0.451 -> 0.387 | Hurt |
| finetune the encoder instead of freezing it (ablation) | pretrained SimTPR · Atari100k, finetuning strategy | IQM HNS: 0.451 -> 0.366 | Hurt |
Closest papers by meaning
- Online Laplacian-Based Representation Learning in Reinforcement Learning (ICML 2025 · similarity 0.87)
- Representation Matters: Offline Pretraining for Sequential Decision Making (ICML 2021 · similarity 0.87)
- Improving Unsupervised Hierarchical Representation with Reinforcement Learning (CVPR 2024 · similarity 0.87)
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.87)
- Bridging State and History Representations: Understanding Self-Predictive RL (ICLR 2024 · similarity 0.86)
- Learning Successor Features the Simple Way (NeurIPS 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.