CaRe-BN: Precise Moving Statistics for Stabilizing Spiking Neural Networks in Reinforcement Learning

A research team has introduced CaRe-BN (Confidence-adaptive and Re-calibration Batch Normalization), a novel technique designed to solve a critical instability problem in training Spiking Neural Networks (SNNs) for Reinforcement Learning (RL). SNNs are prized for their ultra-low power consumption on neuromorphic hardware, mimicking the event-driven nature of biological brains, but their discrete 'spikes' make gradient-based training notoriously unstable. Batch Normalization (BN) is essential for stabilizing SNNs, but in online RL, imprecise moving statistics within BN hinder the agent's ability to exploit learned knowledge, leading to slow convergence and poor policies. CaRe-BN directly addresses this by providing more precise normalization, enabling stable and efficient optimization specifically for RL tasks.

The technical innovation lies in two core mechanisms: a confidence-guided adaptive strategy for updating BN statistics and a re-calibration process to align feature distributions. This allows the SNN to learn more effectively without disrupting the RL training loop. Crucially, CaRe-BN is only active during training, preserving the SNN's inherent energy efficiency during deployment. In extensive testing on discrete and continuous control benchmarks, SNNs equipped with CaRe-BN showed performance gains of up to 22.6% across different neuron models and RL algorithms. Remarkably, these optimized SNNs even outperformed standard Artificial Neural Networks (ANNs) by 5.9%, challenging the assumption that ANNs are inherently superior for complex control. This work, accepted at ICLR 2026, paves the way for creating high-performing, energy-efficient neuromorphic agents for real-world robotics and embedded AI applications.