Synaptic bundle theory for spike-driven sensor-motor system: More than eight independent synaptic bundles collapse reward-STDP learning

A team from the University of Tokyo has published a groundbreaking paper titled 'Synaptic bundle theory for spike-driven sensor-motor system,' revealing why brain-inspired AI systems often fail to learn. The research, led by Takeshi Kobayashi, Shogo Yonekura, and Yasuo Kuniyoshi, demonstrates that when artificial sensorimotor systems using spiking neurons (which mimic biological nerve signals) exceed a critical limit of approximately eight independent synaptic bundles—the connections between sensory and motor neurons—their learning process completely collapses. This finding solves a long-standing mystery in neuromorphic engineering, where applying spike-based control signals to physical actuators like robot muscles has proven notoriously unstable. The team developed a novel experimental system that can precisely vary the number of these independent bundles to isolate the cause of failure.

The paper presents four key quantitative findings: learning failure becomes inevitable once motor neurons or synaptic bundles exceed the critical limit; paradoxically, systems with fewer motor neurons have a higher probability of initial learning failure, but if they do succeed, they learn significantly faster. The researchers identified the root cause: an excess of synaptic bundles leads to a high number of weight updates that move in the opposite direction of the optimal value, derailing the reward-modulated Spike-Timing-Dependent Plasticity (reward-STDP) learning rule. By mapping this 'parameter range of viability,' the work provides a crucial engineering blueprint for building stable, spike-driven robots and opens new avenues to study the true computational functions of biological spikes, which have remained elusive due to previous learning difficulties.