Closing the Loop on the Edge: Local Learning in Spiking Neural Networks

Description

The goal of Brain-Computer Symbiosis is constrained by the need for ultra-low power, adaptive intelligence directly at the edge. Spiking Neural Networks (SNNs) are essential for bio-signal processing, but their training suffers from non-local memory and credit assignment bottlenecks, limiting implantable applications. We propose to resolve these limitations through two complementary, locality-focused SNN learning rules and efficient neuromorphic hardware. In this talk I'll showcase how Forward Propagation Through Time (FPTT) minimizes temporal dependencies via dynamic regularization, solving memory scaling. I further explain ho Trace Propagation (TP) introduces a fully local, forward-only learning rule for efficient, local updates. These algorithms achieve competitive accuracy on dynamic benchmarks and enable on-device continual adaptation for SNNs, critical for patient-specific BCI tuning. This algorithmic breakthrough supports the co-design of ultra-low power hardware, demonstrated by the microwatt μBrain ASIC, paving the way for scalable and energy-efficient neuromorphic systems at the edge.

Period	13 Dec 2025
Event title	From Brain-Computer Interface to Brain-Computer Symbiosis : AI & BCI AzureFlow Forum (人工智能与脑机接口碧海论坛)
Event type	Conference
Location	Shanghai, ChinaShow on map
Degree of Recognition	International