Research map / Reinforcement learning / Visual and representation RL
Data-Efficient Reinforcement Learning with Self-Predictive Representations
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 SPR future-prediction and augmentation objectives | Data-Efficient Rainbow · Atari 100k, 26 games | Median Human-Norm'd: 0.268 -> 0.415 | Helped |
| Disable data augmentation (ablation) | SPR · Atari 100k, 26 games | Median Human-Norm'd: 0.415 -> 0.307 | Hurt |
| Reduce future prediction depth to one step (ablation) | SPR · Atari 100k, 26 games | Mean Human-Norm'd: 0.704 -> 0.570 | Hurt |
| Remove temporal prediction (ablation) | SPR · Atari 100k, 26 games | Mean Human-Norm'd: 0.704 -> 0.507 | Hurt |
| Replace cosine similarity with unnormalized L2 loss (ablation) | SPR · Atari 100k, 26 games | Mean Human-Norm'd: 0.704 -> 0.047 | Hurt |
| Remove projection and prediction networks (ablation) | SPR · Atari 100k, 26 games | Mean Human-Norm'd: 0.704 -> 0.437 | Hurt |
| Allow gradients through online-encoder prediction targets (ablation) | SPR with augmentation · Atari 100k, 26 games | Median Human-Norm'd: 0.415 -> 0.278 | Hurt |
| Add SPR to the controlled Rainbow implementation with augmentation | Rainbow (controlled, w/ aug.) · Atari 100k, 26 games | Mean Human-Norm'd: 0.480 -> 0.704 | Helped |
| Use multi-step latent prediction rather than a non-temporal objective (ablation) | SPR · Atari 100k, 26 games | Median Human-Norm'd: 0.271 -> 0.415 | Helped |
Closest papers by meaning
- Understanding Self-Predictive Learning for Reinforcement Learning (ICML 2023 · similarity 0.91)
- Pretraining Representations for Data-Efficient Reinforcement Learning (NeurIPS 2021 · similarity 0.90)
- Bridging State and History Representations: Understanding Self-Predictive RL (ICLR 2024 · similarity 0.90)
- Accelerating Goal-Conditioned Reinforcement Learning Algorithms and Research (ICLR 2025 · similarity 0.89)
- Efficient Reinforcement Learning Through Adaptively Pretrained Visual Encoder (AAAI 2025 · similarity 0.88)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.87)
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.