Chip-In-Loop SNN Proxy Learning: a new method for efficient training of spiking neural networks

Area of Science:

• Artificial Intelligence
• Neuroscience
• Computer Engineering

Background:

• Current spiking neural network (SNN) training methods, including artificial neural network (ANN) conversion and direct surrogate gradient training, rely on frame-based computations.
• This frame-based approach creates a mismatch with the inherent asynchronous, event-driven nature of SNNs, leading to significant performance degradation on simulators and hardware.
• This degradation hinders the efficient deployment of SNNs for real-time applications.

Purpose of the Study:

• To propose and validate a hardware-based SNN proxy learning method, Chip-In-Loop SNN Proxy Learning (CIL-SPL).
• To eliminate the performance gap caused by the discrepancy between synchronous training methodologies and asynchronous SNN hardware execution.
• To demonstrate the efficacy of CIL-SPL in preserving model performance during real-time asynchronous computation.

Main Methods:

• Developed a novel hardware-based approach named Chip-In-Loop SNN Proxy Learning (CIL-SPL).
• Trained SNN models using CIL-SPL on public datasets, including N-MNIST.
• Evaluated the trained models on both SNN simulators and actual hardware chips to assess performance.

Main Results:

• CIL-SPL effectively mitigated the performance degradation typically observed when deploying frame-based trained SNNs onto asynchronous hardware.
• Models trained with CIL-SPL demonstrated comparable or improved performance on SNN simulators and hardware compared to classical training methods.
• The proposed method successfully addressed the synchronous-asynchronous computation mismatch.

Conclusions:

• Chip-In-Loop SNN Proxy Learning (CIL-SPL) offers a viable solution to the performance degradation issue in SNN deployment.
• This hardware-aware training approach enables more efficient and accurate real-time asynchronous computation with SNNs.
• CIL-SPL paves the way for more effective utilization of SNNs in edge computing and neuromorphic hardware applications.