Reproduce “CaRe-BN: Precise Moving Statistics for Stabilizing Spiking Neural Networks in Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zijie Xu, Xinyu Shi, Yiting Dong, Zihan Huang, Zhaofei Yu
CaRe-BN improves normalization for spiking neural network agents in online reinforcement learning by adapting batch statistics to changing signals and periodically recalibrating them from replay-buffer experiences. In a CLIF-neuron, TD3 comparison across five MuJoCo tasks, it achieved a 5.90% average performance gain over an ANN baseline. It did not outperform the ANN on every task, including HalfCheetah-v4 and Walker2d-v4.
Key insights
- Precise Batch Normalization statistics are critical for effective exploration and learning in online SNN-RL, directly influencing an agent's ability to select optimal actions.
- Combining adaptive weighting of current batch statistics (Ca-BN) with periodic global corrections from a replay buffer (Re-BN) effectively addresses the non-stationarity challenges in SNN-RL.
- Robust, SNN-specific normalization can enable SNNs to match or exceed the performance of Artificial Neural Networks (ANNs) in dynamic RL settings while retaining their inherent energy efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.