Oral: On-chip spike pattern classification for neuromorphic systems

Spiking Neural Networks (SNNs) serve as a bottom-up tool for exploring information processing in brain microcircuits. Neuromorphic chips with on-chip learning capabilities facilitate this exploration. Spike timing-dependent plasticity (STDP) is a brain-like unsupervised learning rule, but its multi-bit implementation in a mixed-signal chip demands significant silicon area and power. Most chips restrict synapse resolution to under 6 bits, negatively impacting the network's performance. We previously introduced a hardware-friendly adaptive STDP rule. Through simulations, we showed that the adaptive STDP rule with only 4-bit fixed-point synapses performs comparably to the STDP rule with 64-bit floating-point synapses in a noisy spike pattern detection model. Here, we present on-chip learning results for the adaptive STDP rule. Its on-chip performance was observed to be slightly better than its ideal model performance, which we attribute to inherent thermal noise within the chip. To demonstrate its scalability, we also extended the adaptive STDP rule with lateral inhibition, a common motif in the brain, and applied it to a competitive spike pattern detection model featuring multiple neurons that compete to detect various patterns.