Adaptive STDP-based on-chip spike pattern detection.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) are vital for modeling brain microcircuits and are key neuromorphic computational models.
• Spike-timing-dependent plasticity (STDP) is a common unsupervised learning rule in SNNs, but its performance degrades on neuromorphic hardware due to limited synaptic efficacy resolution (typically ≤6 bits).
• Simulations often use 64-bit floating-point precision, creating a performance gap with hardware implementations.

Purpose of the Study:

• To experimentally validate the performance of adaptive STDP, a bio-inspired learning rule, on a neuromorphic chip.
• To demonstrate that adaptive STDP can achieve simulation-level performance with low-resolution synaptic efficacy (4-bit fixed-point).
• To showcase the first successful implementation of unsupervised noisy spatiotemporal spike pattern detection on a mixed-signal CMOS neuromorphic chip that maintains simulation performance.

Main Methods:

• Developed and simulated adaptive STDP, a bio-inspired learning rule using 4-bit fixed-point synaptic efficacy.
• Implemented adaptive STDP learning on a custom mixed-signal CMOS neuromorphic chip fabricated using TSMC 250nm technology.
• The chip integrates soma and 256 synapse circuits with learning circuitry.

Main Results:

• Experimental results confirm that adaptive STDP learning performs comparably to traditional STDP learning with high-precision simulations.
• The neuromorphic chip successfully demonstrated unsupervised noisy spatiotemporal spike pattern detection with low-resolution synaptic efficacy.
• The study achieved a significant reduction in synaptic efficacy resolution without compromising learning performance.

Conclusions:

• Adaptive STDP effectively bridges the performance gap between SNN simulations and neuromorphic hardware implementations.
• This work presents the first experimental demonstration of high-performance unsupervised noisy spike pattern detection on a neuromorphic chip with limited precision.
• The findings pave the way for more efficient and scalable neuromorphic systems utilizing bio-inspired learning rules.