E-prop on SpiNNaker 2: Exploring online learning in spiking RNNs on neuromorphic hardware

Abstract

IntroductionIn recent years, the application of deep learning models at the edge has gained attention. Typically, artificial neural networks (ANNs) are trained on graphics processing units (GPUs) and optimized for efficient execution on edge devices. Training ANNs directly at the edge is the next step with many applications such as the adaptation of models to specific situations like changes in environmental settings or optimization for individuals, e.g., optimization for speakers for speech processing. Also, local training can preserve privacy. Over the last few years, many algorithms have been developed to reduce memory footprint and computation.MethodsA specific challenge to train recurrent neural networks (RNNs) for processing sequential data is the need for the Back Propagation Through Time (BPTT) algorithm to store the network state of all time steps. This limitation is resolved by the biologically-inspired E-prop approach for training Spiking Recurrent Neural Networks (SRNNs). We implement the E-prop algorithm on a prototype of the SpiNNaker 2 neuromorphic system. A parallelization strategy is developed to split and train networks on the ARM cores of SpiNNaker 2 to make efficient use of both memory and compute resources. We trained an SRNN from scratch on SpiNNaker 2 in real-time on the Google Speech Command dataset for keyword spotting.ResultWe achieved an accuracy of 91.12% while requiring only 680 KB of memory for training the network with 25 K weights. Compared to other spiking neural networks with equal or better accuracy, our work is significantly more memory-efficient.DiscussionIn addition, we performed a memory and time profiling of the E-prop algorithm. This is used on the one hand to discuss whether E-prop or BPTT is better suited for training a model at the edge and on the other hand to explore architecture modifications to SpiNNaker 2 to speed up online learning. Finally, energy estimations predict that the SRNN can be trained on SpiNNaker2 with 12 times less energy than using a NVIDIA V100 GPU.

SENECA: building a fully digital neuromorphic processor, design trade-offs and challenges

“…Even though the deployed algorithm can be further optimized for SENECA (for example, by quantization, sparsification, and spike grouping), it demonstrates the capability of SENECA to execute such a complex pipeline efficiently. Due to the algorithm's popularity, e-prop and its close variants have been benchmarked on several other neuromorphic processors (Tang et al, 2021;Frenkel and Indiveri, 2022;Perrett et al, 2022;Rostami et al, 2022). Those implementations are either forced to be (1) less efficient due to hardware-algorithm mismatch (Tang et al, 2021;Perrett et al, 2022;Rostami et al, 2022) or (2) hard-wired only to execute a limited version of this algorithm (Frenkel and Indiveri, 2022) which cannot adapt to deploy the new and more efficient online learning algorithms (Yin et al, 2021;Bohnstingl et al, 2022).…”

Critically synchronized brain waves form an effective, robust and flexible basis for human memory and learning

“…4. For the traditional multi-layer ARCSe neural network the diagram (Left panel) includes the connection weights (27)…”

“…Spike based methods were also used for object tracking [23][24][25][26] . A research is booming in using LIF spiking networks for online learning 27 , braille letter reading 28 , different neuromorphic synaptic devices 29 for detection and classification of biological problems [30][31][32][33][34][35][36] . Significant research is focused on making human-level control 37 , optimizing back-propagation algorithms for spiking networks [38][39][40] , as well as penetrating much deeper into ARCSes core [41][42][43][44] with smaller number of time steps 41 , using an event-driven paradigm 36,40,45,46 , applying batch normalization 47 , scatter-and-gather optimizations 48 , supervised plasticity 49 , time-step binary maps 50 , and using transfer learning algorithms 51 .…”