Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Gradient-descent hardware-aware training and deployment for mixed-signal Neuromorphic processors

Uğurcan Çakal, Maryada, Chenxi Wu, Ilkay Ulusoy, Dylan R. Muir

TL;DR

The paper addresses the challenge of deploying robust spiking neural networks on mixed-signal neuromorphic hardware by introducing a gradient-based hardware-aware training workflow that uses a differentiable device model (DynapSim) and a Rockpool extension. It demonstrates mismatch-aware offline training and an end-to-end deployment pipeline to the DYNAP-SE2, including unsupervised 4-bit weight quantization and hardware mapping. The key contributions are the DynapSim-enabled differentiable training framework, mismatch-robust optimization, and an automated Rockpool-based deployment path with reverse-mapping capability. The results show preserved temporal dynamics and discrimination after quantization and hardware deployment, highlighting a scalable path toward commercial-grade mixed-signal neuromorphic applications.

Abstract

Mixed-signal neuromorphic processors provide extremely low-power operation for edge inference workloads, taking advantage of sparse asynchronous computation within Spiking Neural Networks (SNNs). However, deploying robust applications to these devices is complicated by limited controllability over analog hardware parameters, as well as unintended parameter and dynamical variations of analog circuits due to fabrication non-idealities. Here we demonstrate a novel methodology for ofDine training and deployment of spiking neural networks (SNNs) to the mixed-signal neuromorphic processor DYNAP-SE2. The methodology utilizes gradient-based training using a differentiable simulation of the mixed-signal device, coupled with an unsupervised weight quantization method to optimize the network's parameters. Parameter noise injection during training provides robustness to the effects of quantization and device mismatch, making the method a promising candidate for real-world applications under hardware constraints and non-idealities. This work extends Rockpool, an open-source deep-learning library for SNNs, with support for accurate simulation of mixed-signal SNN dynamics. Our approach simplifies the development and deployment process for the neuromorphic community, making mixed-signal neuromorphic processors more accessible to researchers and developers.

Gradient-descent hardware-aware training and deployment for mixed-signal Neuromorphic processors

TL;DR

The paper addresses the challenge of deploying robust spiking neural networks on mixed-signal neuromorphic hardware by introducing a gradient-based hardware-aware training workflow that uses a differentiable device model (DynapSim) and a Rockpool extension. It demonstrates mismatch-aware offline training and an end-to-end deployment pipeline to the DYNAP-SE2, including unsupervised 4-bit weight quantization and hardware mapping. The key contributions are the DynapSim-enabled differentiable training framework, mismatch-robust optimization, and an automated Rockpool-based deployment path with reverse-mapping capability. The results show preserved temporal dynamics and discrimination after quantization and hardware deployment, highlighting a scalable path toward commercial-grade mixed-signal neuromorphic applications.

Abstract

Mixed-signal neuromorphic processors provide extremely low-power operation for edge inference workloads, taking advantage of sparse asynchronous computation within Spiking Neural Networks (SNNs). However, deploying robust applications to these devices is complicated by limited controllability over analog hardware parameters, as well as unintended parameter and dynamical variations of analog circuits due to fabrication non-idealities. Here we demonstrate a novel methodology for ofDine training and deployment of spiking neural networks (SNNs) to the mixed-signal neuromorphic processor DYNAP-SE2. The methodology utilizes gradient-based training using a differentiable simulation of the mixed-signal device, coupled with an unsupervised weight quantization method to optimize the network's parameters. Parameter noise injection during training provides robustness to the effects of quantization and device mismatch, making the method a promising candidate for real-world applications under hardware constraints and non-idealities. This work extends Rockpool, an open-source deep-learning library for SNNs, with support for accurate simulation of mixed-signal SNN dynamics. Our approach simplifies the development and deployment process for the neuromorphic community, making mixed-signal neuromorphic processors more accessible to researchers and developers.
Paper Structure (20 sections, 3 equations, 11 figures, 3 tables)

This paper contains 20 sections, 3 equations, 11 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Overview of the DYNAP-SE2 Hardware
  3. Methods
  4. Task and Training approach
  5. Toy task: Frozen Noise Classification
  6. Data
  7. Network
  8. Response Analysis
  9. Mismatch Simulation
  10. Optimization
  11. Deployment to DYNAP™-SE2
  12. Computational Graph
  13. Mapping
  14. Quantization
  15. Deployment
  16. ...and 5 more sections

Figures (11)

  • Figure 1: DYNAP-SE2 Architecture. See text for details. NC: Neural Core; R: Router; BG: Bias Generator. Other acronyms: see text.
  • Figure 2: Frozen noise classification task. 60 input channels provide input spiking patterns to the network (left; 4 shown here). The network provides two output channels (right), which should emit high spiking activity when presented with one of two target frozen noise inputs. "LinearJax" and "DynapSim" modules, provided by Rockpool, are used to simulate the weights, synapse and neuron dynamics of the network (see text for further details).
  • Figure 3: Frozen noise recordings used in training.
  • Figure 4: Mismatch Simulation. Nominal values for two parameters (orange) are applied to the network, following which DynapSim is used to simulate the parameter mismatch that would be experienced when deployed to a DYNAP-SE2 device (blue distributions).
  • Figure 5: Mean square error (MSE) loss over the course of training.
  • ...and 6 more figures