깊은 스파이킹 신경망을 위한 뉴럴 아키텍처 탐색 알고리즘 성능 분석

Abstract

This paper presents analyses of the performance of three different search algorithms, including random, greedy, and Bayesian, in the neural architecture search (NAS). To conduct this study, we used Autokeras, a keras-based AutoML framework, to search architectures of deep neural networks (DNNs) and deep spiking neural networks (SNNs). We evaluated the performance of NAS algorithms on searching deep SNNs and DNNs on CIFAR-10 datasets. Our experimental results showed that the Bayesian algorithm outperformed the other two in terms of accuracy, while the greedy algorithm achieved the best accuracy on DNNs. Our findings suggest that the Bayesian algorithm is promising in NAS for both DNNs and SNNs.