[Paper Review] RMP-SNN: Residual Membrane Potential Neuron for Enabling Deeper High-Accuracy and Low-Latency Spiking Neural Network

Bing Han, Gopalakrishnan Srinivasan+1 more|arXiv (Cornell University)|Feb 25, 2020

Advanced Memory and Neural Computing46 references34 citations

TL;DR

The paper introduces Residual Membrane Potential (RMP) spiking neurons for ANN-SNN conversion, enabling near loss-less conversion and much lower latency than hard-reset spiking neurons across VGG-16, ResNet-20/34 on CIFAR-10/100 and ImageNet.

ABSTRACT

Spiking Neural Networks (SNNs) have recently attracted significant research interest as the third generation of artificial neural networks that can enable low-power event-driven data analytics. The best performing SNNs for image recognition tasks are obtained by converting a trained Analog Neural Network (ANN), consisting of Rectified Linear Units (ReLU), to SNN composed of integrate-and-fire neurons with "proper" firing thresholds. The converted SNNs typically incur loss in accuracy compared to that provided by the original ANN and require sizable number of inference time-steps to achieve the best accuracy. We find that performance degradation in the converted SNN stems from using "hard reset" spiking neuron that is driven to fixed reset potential once its membrane potential exceeds the firing threshold, leading to information loss during SNN inference. We propose ANN-SNN conversion using "soft reset" spiking neuron model, referred to as Residual Membrane Potential (RMP) spiking neuron, which retains the "residual" membrane potential above threshold at the firing instants. We demonstrate near loss-less ANN-SNN conversion using RMP neurons for VGG-16, ResNet-20, and ResNet-34 SNNs on challenging datasets including CIFAR-10 (93.63% top-1), CIFAR-100 (70.93% top-1), and ImageNet (73.09% top-1 accuracy). Our results also show that RMP-SNN surpasses the best inference accuracy provided by the converted SNN with "hard reset" spiking neurons using 2-8 times fewer inference time-steps across network architectures and datasets.

Motivation & Objective

• Address information loss in ANN-SNN conversion caused by hard-reset spiking neurons.
• Propose a soft-reset Residual Membrane Potential (RMP) neuron to preserve residual membrane potential after firing.
• Develop threshold initialization and training constraints to enable near loss-less conversion for deep SNNs.
• Demonstrate high-accuracy, low-latency SNNs on VGG-16, ResNet-20, and ResNet-34 across CIFAR-10, CIFAR-100, and ImageNet.

Proposed method

• Introduce Residual Membrane Potential (RMP) spiking neuron that performs soft reset by subtracting V_th at spike instants (V_m <- V_m - V_th).
• Show that RMP yields linear input-output behavior f_out ≈ η f_in over a wide range of η, enabling accurate ANN-SNN mapping.
• Derive threshold bounds V_in and V_th to keep f_out in the desirable range [f_in, 1) and prevent excessive spiking (Eq. 6 and Eq. 7).
• Propose threshold balancing with V_th upper bound using V_in^max and scale thresholds (α) for optimal accuracy-latency trade-off.
• Apply constrained ANN training (remove batch norm and bias) and layer-wise threshold initialization for deep networks.

Experimental results

Research questions

• RQ1 How does a soft-reset (RMP) neuron affect information preservation during ANN-SNN conversion compared to hard-reset IF neurons?
• RQ2 Can threshold initialization and constrained ANN training enable near loss-less conversion for deep networks (VGG-16, ResNet-20/34) on CIFAR-10/100 and ImageNet?
• RQ3 What is the impact of RMP on inference latency and spike activity relative to traditional SNNs?
• RQ4 What thresholds and operating regimes maximize accuracy while minimizing latency in RMP-SNNs?

Key findings

• RMP neurons retain residual membrane potential, enabling near-lossless ANN-SNN conversion with top-1 accuracies close to or equal to ANN baselines across tested networks and datasets.
• RMP-SNNs require 2–8× fewer inference time-steps than hard-reset SNNs to reach comparable or better accuracy on CIFAR-10, CIFAR-100, and ImageNet.
• Layer-wise threshold initialization using V_in^max and constrained ANN training yield significantly lower conversion loss than prior hard-reset based conversions (examples: CIFAR-10, CIFAR-100, ImageNet).
• VGG-16 RMP-SNN achieves 93.63% top-1 on CIFAR-10, matching the ANN, with 2048 time-steps, while reduced-threshold variants reach high accuracy in as few as 64–512 time-steps.
• ResNet-20/34 show substantial accuracy gains over baseline hard-reset SNNs, with reduced-threshold variants delivering up to 8× faster inference while keeping spike activity around 1–2%.
• Across datasets, RMP-SNNs exhibit the best reported accuracy and the lowest ANN-SNN conversion loss among the compared methods.