[C80] Efficient and Robust SNN Decoder Training for Closed-Loop Brain-Machine Interfaces

Abstract

To enable low-power, closed-loop control in implantable brain-machine interfaces (BMIs), we explore efficient training strategies for spiking neural network (SNN) decoders under realistic conditions. Using the Online Prosthesis Simulator (OPS), we implement and evaluate three learning paradigms—Reinforcement Learning (RL), Random-Step (RS), and Episode-Based (EP)—each reflecting trade-offs in temporal supervision, training cost, and robustness. Based on these insights, we adopt a two-phase training approach that combines RS-based initialization with EP-based fine-tuning, achieving fast convergence and strong closed-loop performance. To improve robustness against neural signal degradation, we further introduce perturbation-injected training (PIT), which exposes the model to randomized channel dropouts and distribution shifts during training. Our optimized SNN decoder achieves an average time-to-target of sub-0.84s, requires zero multiply-accumulate (MAC) operations, and completes under 34k accumulate operations, highlighting its suitability for resource-constrained, implantable BMI systems.