Heartbeat Classification with Spiking Neural Networks on the Loihi Neuromorphic Processor

Abstract

Neuromorphic processors are attractive for energy-constrained applications as they are designed to emulate the energy-efficient spiking neural networks (SNNs) of the human brain. This research aims to evaluate whether a state-of-the-art SNN design methodology, artificial-to-spiking neural network (ANN-to-SNN) conversion, and a novel neuromorphic processor, Loihi, together provide an accurate and energy-efficient approach for heartbeat classification with neural networks. To perform this evaluation, a 1D-convolutional neural network (1D-CNN) is first trained to classify arrhythmias in the artificial domain. The ANN is then converted to an architecturally identical SNN with the SNN-Toolbox framework. Finally, the performance of the SNN on Loihi is compared to the performance of the ANN on Intel Core i7 CPU, Intel Neural Compute Stick 2 (NCS2), and Google Coral Edge TPU (Edge TPU) devices. Over five classes, the SNN reaches an accuracy and macro-averaged F1 score of 97.8% and 87.9%, respectively, compared to 98.4% and 90.8% for the ANN. In terms of performance, Loihi is found to operate at the lowest dynamic power, but also at the highest latency. Overall, Loihi is estimated to result in a 1.5× and 110× higher energy-delay product versus the NCS2 and Edge TPU, respectively. These results demonstrate other edge neural network devices to be more dynamic energy-efficient for the model tested. Based on the insights gained, this study discusses future directions to enhance neuromorphic computing for energy-constrained applications like heartbeat classification.