[Paper Review] Supervised training of spiking neural networks for robust deployment on mixed-signal neuromorphic processors

Muir, Dylan|arXiv (Cornell University)|Jan 1, 2021

Advanced Memory and Neural Computing64 references24 citations

TL;DR

This paper proposes a supervised training method for spiking neural networks (SNNs) that achieves robust performance on mixed-signal neuromorphic hardware without per-chip calibration. By mimicking a pre-trained recurrent neural network (RNN) using a local learning rule derived from control theory, the method produces SNNs resilient to device mismatch, thermal noise, and quantization, enabling reliable deployment on energy-efficient neuromorphic chips with minimal hardware overhead.

ABSTRACT

Mixed-signal analog/digital circuits emulate spiking neurons and synapses with extremely high energy efficiency, an approach known as “neuromorphic engineering”. However, analog circuits are sensitive to process-induced variation among transistors in a chip (“device mismatch”). For neuromorphic implementation of Spiking Neural Networks (SNNs), mismatch causes parameter variation between identically-configured neurons and synapses. Each chip exhibits a different distribution of neural parameters, causing deployed networks to respond differently between chips. Current solutions to mitigate mismatch based on per-chip calibration or on-chip learning entail increased design complexity, area and cost, making deployment of neuromorphic devices expensive and difficult. Here we present a supervised learning approach that produces SNNs with high robustness to mismatch and other common sources of noise. Our method trains SNNs to perform temporal classification tasks by mimicking a pre-trained dynamical system, using a local learning rule from non-linear control theory. We demonstrate our method on two tasks requiring temporal memory, and measure the robustness of our approach to several forms of noise and mismatch. We show that our approach is more robust than common alternatives for training SNNs. Our method provides robust deployment of pre-trained networks on mixed-signal neuromorphic hardware, without requiring per-device training or calibration.

Motivation & Objective

• To address the challenge of device mismatch in mixed-signal neuromorphic processors, which degrades SNN performance across chips.
• To eliminate the need for costly per-device calibration or on-chip learning by training SNNs to be inherently robust to parameter variation.
• To enable reliable deployment of pre-trained SNNs on neuromorphic hardware without post-deployment adaptation.
• To develop a training framework that combines knowledge distillation with local learning rules for temporal classification tasks.

Proposed method

• Train a non-spiking rate RNN using Back-Propagation Through Time (BPTT) to solve a target temporal classification or regression task.
• Use the RNN's internal dynamics as a teacher system to train a spiking neural network (SNN) via a local learning rule inspired by non-linear control theory.
• Implement an SNN with leaky integrate-and-fire (LIF) neurons and balanced fast/slow recurrent weights to ensure stability and robustness.
• Apply feedback error correction using a decaying feedback rate to minimize the difference between the SNN's output and the teacher RNN's hidden state.
• Use knowledge distillation by training the SNN to replicate the RNN’s hidden state dynamics, ensuring task performance is preserved.
• Apply noise injection during training to simulate device mismatch, thermal noise, and quantization, enhancing generalization to hardware imperfections.

Experimental results

Research questions

• RQ1 Can a supervised training method produce SNNs that are robust to device mismatch without requiring per-chip calibration or on-chip learning?
• RQ2 How does the proposed method compare to standard SNN training approaches in terms of robustness to parameter variation and noise?
• RQ3 To what extent can knowledge distillation from a rate RNN enable SNNs to maintain high performance on temporal tasks under hardware-imposed noise?
• RQ4 Does the use of a local learning rule derived from control theory enable effective training of SNNs for arbitrary supervised tasks?
• RQ5 Can the resulting SNNs maintain accuracy when deployed on real mixed-signal neuromorphic hardware with known parameter mismatch?

Key findings

• The proposed method achieves significantly higher robustness to device mismatch (up to 20%) compared to standard SNN training approaches, with minimal performance degradation.
• The SNNs trained with the proposed method maintain high accuracy even when synaptic weights are quantized to as low as 2 bits, demonstrating resilience to low-precision implementations.
• The method is robust to thermal noise, with performance degradation remaining low even at high noise levels (σ = 0.1), indicating suitability for real hardware.
• The SNNs remain stable and accurate when 40% of neurons are silenced, showing resilience to neuron dropout and hardware faults.
• The method enables reliable deployment on DYNAP™-SE1 neuromorphic hardware with only 7.6 µW dynamic power and 30 µW static power, achieving ultra-low energy consumption.
• The approach outperforms baseline methods in all tested noise conditions, including mismatch, quantization, and thermal noise, confirming its generalization capability.