SpikeFit: Towards Optimal Deployment of Spiking Networks on Neuromorphic Hardware

TL;DR

SpikeFit tackles the challenge of deploying spiking neural networks on neuromorphic hardware under strict constraints such as limited neuron budgets, a fixed set of synaptic states, and low-precision weights. The authors introduce Clusterization Aware Training (CAT) to learn per-layer codebooks of a small number of discrete weight values and Fisher Spike Contribution (FSC) pruning to select hardware-relevant channels guided by an approximate Fisher information objective. Together, CAT and FSC enable end-to-end training that yields highly deployable SNNs with strong accuracy and low energy, demonstrated on CIFAR-10 with VGG-16 where a four-value weight codebook achieves state-of-the-art performance and best DeployRatio. A novelDeployRatio metric combines accuracy, latency, energy, and model size to quantify hardware deployability, and SpikeFit outperforms the strongest baselines (quantization-plus-clustering and ternary) while achieving broader hardware compatibility. The work provides a practical framework for hardware-aware SNN compression and deployment across a range of neuromorphic processors, potentially broadening the applicability of energy-efficient SNNs in real devices.

Abstract

This paper introduces SpikeFit, a novel training method for Spiking Neural Networks (SNNs) that enables efficient inference on neuromorphic hardware, considering all its stringent requirements: the number of neurons and synapses that can fit on a single device, and lower bit-width representations (e.g., 4-bit, 8-bit). Unlike conventional compressing approaches that address only a subset of these requirements (limited numerical precision and limited number of neurons in the network), SpikeFit treats the allowed weights' discrete values themselves as learnable parameters co-optimized with the model, allowing for optimal Clusterization-Aware Training (CAT) of the model's weights at low precision (2-, 4-, or 8-bit) which results in higher network compression efficiency, as well as limiting the number of unique synaptic connections to a value required by neuromorphic processor. This joint optimization allows SpikeFit to find a discrete weight set aligned with hardware constraints, enabling the most complete deployment across a broader range of neuromorphic processors than existing methods of SNN compression support. Moreover, SpikeFit introduces a new hardware-friendly Fisher Spike Contribution (FSC) pruning method showing the state-of-the-art performance. We demonstrate that for spiking neural networks constrained to only four unique synaptic weight values (M = 4), our SpikeFit method not only outperforms state-of-the-art SNNs compression methods and conventional baselines combining extreme quantization schemes and clustering algorithms, but also meets a wider range of neuromorphic hardware requirements and provides the lowest energy use in experiments.

Figures (2)

• Figure 1: Overview of the Clusterization Aware Training (CAT) pipeline.
• Figure 2: Overview of the Fisher Spike Contribution (FSC) pruning pipeline.