Research map / Continual learning / Online and replay-based
Rewiring Neurons in Non-Stationary Environments
Zhicheng Sun, Yadong MU
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 |
|---|---|---|---|
| add rewiring by neuron permutation (with caching) to FT-L2 baseline (ablation) | FT-L2 · HalfCheetah/forgetting | average performance: 0.67 -> 0.95 | Improved |
| add caching of each learned wiring (versus rewiring without caching) (ablation) | Rewire · HalfCheetah/forgetting | average performance: 0.76 -> 0.95 | Improved |
| add multi-mode rewiring strategy (ablation) | Rewire · HalfCheetah/forgetting | average performance: 0.95 -> 1.14 | Improved |
| add alignment mechanism for cached wirings (ablation) | Rewire + multi-mode · HalfCheetah/forgetting | average performance: 1.14 -> 1.31 | Improved |
| add alignment mechanism without multi-mode (ablation) | Rewire · HalfCheetah/forgetting | average performance: 0.95 -> 1.15 | Improved |
| add multi-mode on top of rewiring plus alignment (ablation) | Rewire + alignment · HalfCheetah/forgetting | average performance: 1.15 -> 1.31 | Improved |
| rewiring (without multi-mode) versus vanilla finetuning, speed to reward 2000 (ablation) | Rewire vs FT · HalfCheetah/forgetting first stage | steps to average reward 2000 | Improved |
| rewiring versus vanilla finetuning, final reward (ablation) | Rewire vs FT · HalfCheetah/forgetting first stage | final reward: over 3000 | Improved |
| add multi-mode rewiring, final reward (ablation) | Rewire · HalfCheetah/forgetting first stage | final reward: nearly 4000 | Improved |
| add multi-mode rewiring, early learning speed (ablation) | Rewire · HalfCheetah/forgetting first stage | reward at step 200k | Hurt |
| add KL distillation loss across wirings (ablation) | multi-mode rewiring · HalfCheetah/forgetting first stage | mean and median reward | Improved |
| increase number of modes K beyond 3 (ablation) | multi-mode rewiring · HalfCheetah/forgetting | performance | Hurt |
Closest papers by meaning
- Neuro-evolutionary Continual Reinforcement Learning (ICML 2026 · similarity 0.89)
- Stay Hungry, Keep Learning: Sustainable Plasticity for Deep Reinforcement Learning (ICML 2025 · similarity 0.88)
- Self-Normalized Resets for Plasticity in Continual Learning (ICLR 2025 · similarity 0.86)
- Neuroplastic Expansion in Deep Reinforcement Learning (ICLR 2025 · similarity 0.86)
- The Dormant Neuron Phenomenon in Multi-Agent Reinforcement Learning Value Factorization (NeurIPS 2024 · similarity 0.85)
- Feedback control guides credit assignment in recurrent neural networks (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.
