Research map / Reinforcement learning / Visual and representation RL
Self-Supervised Discovering of Interpretable Features for Reinforcement Learning
Wenjie Shi, Gao Ming Huang, Shiji Song, Zhuoyuan Wang, Tingyu Lin, Cheng Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 SSINet masks instead of perturbation-based masks | policy trained on mask pixels · Enduro | average return: 1741.10 -> 2755.42 | Improved |
| use SSINet masks instead of gradient-based masks | policy trained on mask pixels · Enduro | average return: 819.92 -> 2755.42 | Improved |
| use SSINet masks instead of perturbation-based masks | policy trained on mask pixels · Seaquest | average return: 1830.00 -> 2356.67 | Improved |
| use SSINet masks instead of perturbation-based masks | policy trained on mask pixels · SpaceInvaders | average return: 562.92 -> 740.45 | Improved |
| use SSINet masks instead of gradient-based masks | policy trained on mask pixels · Seaquest | average return: 836.00 -> 2356.67 | Improved |
| use SSINet masks instead of gradient-based masks | policy trained on mask pixels · SpaceInvaders | average return: 297.73 -> 740.45 | Improved |
| overlay states with SSINet attention (mask policy vs expert policy) | Lane-following agent · Duckietown (ten maps) | average return | Improved |
| use product of FOR and BER instead of FOR alone as predictor of return | mask metrics · Duckietown empty | R^2 with average return: 0.525 -> 0.655 | Improved |
| use AUC instead of F1 as predictor of return | mask metrics · Duckietown U-turn | R^2 with average return: 0.148 -> 0.812 | Improved |
| use PPO instead of SAC/TD3 as training algorithm | Lane-following agent · Duckietown (all maps) | average return | Improved |
| use more complex actor architecture | Lane-following agent · Duckietown | average return | No clear effect |
| increase regularization scale alpha (ablation) | SSINet · Duckietown masks | attended region | Mixed |
Closest papers by meaning
- Self-Supervised Attention-Aware Reinforcement Learning (AAAI 2021 · similarity 0.91)
- Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data (ICLR 2024 · similarity 0.86)
- Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play (ICLR 2026 · similarity 0.86)
- Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.85)
- 1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities (NeurIPS 2025 · similarity 0.85)
- VLMs-Guided Representation Distillation for Efficient Vision-Based Reinforcement Learning (CVPR 2025 · 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.
