An Efficient Evolutionary Neural Architecture Search Algorithm Without Training

Neural Architecture Search (NAS) has made significant advancements in autonomously constructing high-performance network architectures, capturing extensive attention. However, a key challenge of existing NAS approaches is the intensive performance evaluation, leading to significant time and computational resource consumption. In this paper, we propose an efficient Evolutionary Neural Architecture Search (ENAS) method to address this issue. Specifically, in order to accelerate the convergence speed of the algorithm and shorten the search time, thereby avoiding blind searching in the early stages of the algorithm, we drew on the principles of biometrics to redesign the interaction between individuals in the evolutionary algorithm. By making full use of the information carried by individuals, we promoted information exchange and optimization between individuals and their neighbors, thereby improving local search capabilities while maintaining global search capabilities. Furthermore, to accelerate the evaluation process and minimize computational resource consumption, a multi-metric training-free evaluator is introduced to assess network performance, bypassing the resource-intensive training phase, and the adopted multi-metric combination method further solves the ranking offset problem. To evaluate the performance of the proposed method, we conduct experiments on two widely adopted benchmarks, NAS-Bench-101 and NAS-Bench-201. Comparative analysis with state-of-the-art algorithms shows that our proposed method identifies network architectures with comparable or better performance while requiring significantly less time.