Eventprop training for efficient neuromorphic applications

TL;DR

The paper presents an end-to-end pipeline for training recurrent spiking neural networks (RSNNs) on GPUs using the Eventprop algorithm in mlGeNN and deploying them to Intel's Loihi 2 neuromorphic chip for keyword spotting tasks. By quantizing pretrained weights to 8-bit and converting models via NetX/NxKernel, the approach delivers near-equivalent accuracy to GPU-based training while achieving up to 10× faster inference and ~200× lower energy on Loihi 2 compared to a Jetson Orin Nano. The work demonstrates substantial memory and compute efficiency in training, and shows Loihi 2's suitability for energy-efficient, low-latency edge inference on SHD and SSC datasets. This pipeline enables practical neuromorphic edge deployment and paves the way for expanding SNN deployments to more complex architectures and hardware ecosystems.

Abstract

Neuromorphic computing can reduce the energy requirements of neural networks and holds the promise to `repatriate' AI workloads back from the cloud to the edge. However, training neural networks on neuromorphic hardware has remained elusive. Here, we instead present a pipeline for training spiking neural networks on GPUs, using the efficient event-driven Eventprop algorithm implemented in mlGeNN, and deploying them on Intel's Loihi 2 neuromorphic chip. Our benchmarking on keyword spotting tasks indicates that there is almost no loss in accuracy between GPU and Loihi 2 implementations and that classifying a sample on Loihi 2 is up to 10X faster and uses 200X less energy than on an NVIDIA Jetson Orin Nano.

Figures (4)

• Figure 1: Example Loihi 2 systems, ranging from a 31 mm$^2$ single chip to High Performance Computing (HPC)-scale. Loihi 2 chips can be connected through six asynchronous parallel interfaces, enabling the efficient extension of the Loihi 2 neuromorphic mesh in three dimensions.
• Figure 2: Examples from datasets (A) "zwei" from SHD (B) "left" from SSC.
• Figure 3: Comparing SHD training cost of Eventprop using mlGeNN against BPTT in Spyx. (A) Peak GPU memory usage, (B) Time to train 100 epochs. All experiments were performed on a workstation with NVIDIA RTX A5000 GPU. All models use batch size 32.
• Figure 4: Accuracy of models trained on SHD and SSC and evaluated using mlGeNN using 32 floating point as well as using Lava's with the "fixed_pt" tag and on Loihi 2 using the NxKernel API. For mlGeNN and Lava results, bar heights represent mean times and error bars standard deviations, all calculated over 5 models trained with a different random seed. On Loihi 2, bar heights represent the accuracy of a single model.