Research map / Architectures and efficiency / RNNs and normalization
Fast And Slow Learning Of Recurrent Independent Mechanisms
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 Meta-RIMs instead of Vanilla LSTM for zero-shot transfer | DoorKey policy trained on the easiest environment · DoorKey More Difficult | mean reward: 0.18 -> 0.45 | Helped |
| Use Meta-RIMs instead of RIMs for zero-shot transfer | DoorKey policy trained on the easiest environment · DoorKey More Difficult | success rate: 0.02 -> 0.47 | Helped |
| Train a vanilla LSTM using the meta-learning update setup (ablation) | Vanilla LSTM · Curriculum learning experiments | performance | Helped |
| Update only a sparse subset of modules in the fast loop rather than all modules (ablation) | Meta-RIMs · MiniGrid and BabyAI tasks | learning speed and performance | Helped |
| Increase the interaction span in the outer loop (ablation) | Meta-RIMs · MiniGrid and BabyAI tasks | learning speed | Helped |
| Make attention parameters fast-updating and module parameters slow-updating (ablation) | Meta-RIMs · MiniGrid and BabyAI tasks | mean reward and success rate | Hurt |
| Reduce the attention learning rate to one-fourth while updating all parameters together (ablation) | Meta-RIMs · MiniGrid and BabyAI tasks | learning speed and performance | Hurt |
| Randomly deactivate one of the three active modules (ablation) | Meta-RIMs · MemoryS13Random | frames to complete 20 episodes: 192 -> 243 | Hurt |
| Randomly deactivate two of the three active modules (ablation) | Meta-RIMs · MemoryS13Random | frames to complete 20 episodes: 192 -> 2179 | Hurt |
Closest papers by meaning
- Been There, Done That: Meta-Learning with Episodic Recall (ICML 2018 · similarity 0.88)
- Learning to Learn How to Learn: Self-Adaptive Visual Navigation Using Meta-Learning (CVPR 2019 · similarity 0.88)
- Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments (ICLR 2018 · similarity 0.88)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.87)
- Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement (NeurIPS 2024 · similarity 0.86)
- MetaCARD: Meta-Reinforcement Learning with Task Uncertainty Feedback via Decoupled Context-Aware Reward and Dynamics Components (AAAI 2024 · similarity 0.86)
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.