Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Deployment of Spiking Neural Networks on SpiNNaker2 for DVS Gesture Recognition Using Neuromorphic Intermediate Representation

Sirine Arfa, Bernhard Vogginger, Chen Liu, Johannes Partzsch, Mark Schone, Christian Mayr

TL;DR

This work tackles efficient deployment of Spiking Neural Networks (SNNs) for DVS gesture recognition on the SpiNNaker2 neuromorphic platform under tight memory constraints. It presents an end-to-end deployment pipeline that uses the neuromorphic intermediate representation (NIR) and compares two quantization strategies—Post-Training Quantization (PTQ) and Quantization Aware Training (QAT)—to achieve 8-bit on-chip inference with accuracy close to full-precision. The results show that QAT yields smaller accuracy losses (about 0.56% on-chip for the Q-SNN) than PTQ (up to ~1.07%), with on-chip accuracies around 94.13% for the quantized model and 94.0% for the FP32 baseline, while delivering energy-efficient inference (≈0.459 J per gesture). The study demonstrates SpiNNaker2’s viability for low-energy edge recognition and provides an end-to-end, open-source deployment workflow that can be extended to multi-chip systems and real-time spiking sensing tasks.

Abstract

Spiking Neural Networks (SNNs) are highly energy-efficient during inference, making them particularly suitable for deployment on neuromorphic hardware. Their ability to process event-driven inputs, such as data from dynamic vision sensors (DVS), further enhances their applicability to edge computing tasks. However, the resource constraints of edge hardware necessitate techniques like weight quantization, which reduce the memory footprint of SNNs while preserving accuracy. Despite its importance, existing quantization methods typically focus on synaptic weights quantization without taking account of other critical parameters, such as scaling neuron firing thresholds. To address this limitation, we present the first benchmark for the DVS gesture recognition task using SNNs optimized for the many-core neuromorphic chip SpiNNaker2. Our study evaluates two quantization pipelines for fixed-point computations. The first approach employs post training quantization (PTQ) with percentile-based threshold scaling, while the second uses quantization aware training (QAT) with adaptive threshold scaling. Both methods achieve accurate 8-bit on-chip inference, closely approximating 32-bit floating-point performance. Additionally, our baseline SNNs perform competitively against previously reported results without specialized techniques. These models are deployed on SpiNNaker2 using the neuromorphic intermediate representation (NIR). Ultimately, we achieve 94.13% classification accuracy on-chip, demonstrating the SpiNNaker2's potential for efficient, low-energy neuromorphic computing.

Efficient Deployment of Spiking Neural Networks on SpiNNaker2 for DVS Gesture Recognition Using Neuromorphic Intermediate Representation

TL;DR

This work tackles efficient deployment of Spiking Neural Networks (SNNs) for DVS gesture recognition on the SpiNNaker2 neuromorphic platform under tight memory constraints. It presents an end-to-end deployment pipeline that uses the neuromorphic intermediate representation (NIR) and compares two quantization strategies—Post-Training Quantization (PTQ) and Quantization Aware Training (QAT)—to achieve 8-bit on-chip inference with accuracy close to full-precision. The results show that QAT yields smaller accuracy losses (about 0.56% on-chip for the Q-SNN) than PTQ (up to ~1.07%), with on-chip accuracies around 94.13% for the quantized model and 94.0% for the FP32 baseline, while delivering energy-efficient inference (≈0.459 J per gesture). The study demonstrates SpiNNaker2’s viability for low-energy edge recognition and provides an end-to-end, open-source deployment workflow that can be extended to multi-chip systems and real-time spiking sensing tasks.

Abstract

Spiking Neural Networks (SNNs) are highly energy-efficient during inference, making them particularly suitable for deployment on neuromorphic hardware. Their ability to process event-driven inputs, such as data from dynamic vision sensors (DVS), further enhances their applicability to edge computing tasks. However, the resource constraints of edge hardware necessitate techniques like weight quantization, which reduce the memory footprint of SNNs while preserving accuracy. Despite its importance, existing quantization methods typically focus on synaptic weights quantization without taking account of other critical parameters, such as scaling neuron firing thresholds. To address this limitation, we present the first benchmark for the DVS gesture recognition task using SNNs optimized for the many-core neuromorphic chip SpiNNaker2. Our study evaluates two quantization pipelines for fixed-point computations. The first approach employs post training quantization (PTQ) with percentile-based threshold scaling, while the second uses quantization aware training (QAT) with adaptive threshold scaling. Both methods achieve accurate 8-bit on-chip inference, closely approximating 32-bit floating-point performance. Additionally, our baseline SNNs perform competitively against previously reported results without specialized techniques. These models are deployed on SpiNNaker2 using the neuromorphic intermediate representation (NIR). Ultimately, we achieve 94.13% classification accuracy on-chip, demonstrating the SpiNNaker2's potential for efficient, low-energy neuromorphic computing.

Paper Structure

This paper contains 20 sections, 6 equations, 3 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Spiking Neuron Model and QSNNs
  4. Neuromorphic Intermidiate Representation
  5. The SpiNNaker2 System
  6. DVS Gesture Recognition
  7. Methods
  8. Training pipeline and architectures
  9. Spiking Neural Networks
  10. Slicing Method
  11. Setup
  12. Quantization Pipelines
  13. PTQ
  14. QAT
  15. SpiNNaker2 Implementation
  16. ...and 5 more sections

Figures (3)

  • Figure 1: A: Event stream from the DVS Gestures dataset amir2017lowB: Full precision and quantization aware training (QAT) is conducted in snnTorch eshraghian2023training. C: The trained models are parsed with NIR pedersen2024neuromorphic and mapped to SpiNNaker2 with the py-spinnaker2 library vogginger2023pyspinnaker2. Post training quantization (PTQ) is applied to the full precision model. D: Quantization aware training stores as full precision representation of the weights. The forward pass converts the weights to 8-bit integers, and the backward pass updates the full precision parameters directly.
  • Figure 2: SpiNNaker2 Network Architecture: Visualization of LIF neuron populations, indicating the number of neurons in each population and the synaptic projection connecting them.
  • Figure 3: Profiling the on-chip classification accuracy for different percentile values. The highest accuracy is achieved at 100th percentile of weights.