Efficient Evolutionary Neural Architecture Search (NAS) by Modular Inheritable Crossover

Abstract

The convolution neural network is prominent in image processing, and a large number of excellent deep neural networks have been proposed in recent years. However, the hand-actuated design of a neural network is time-consuming, laborious, and challenging. Thus many neural architecture search (NAS) methods have been proposed, among which the evolutionary NAS methods have achieved encouraging results due to the global search capability of evolutionary algorithms. Nevertheless, most evolutionary NAS methods use only mutation operators for offspring generation, and the generated offspring networks could be quite different from their parent networks. To address this deficiency, we propose an efficient evolutionary NAS method using a tailored crossover operator. Different from existing mutation operators, the proposed crossover operator enables the offspring network to inherit promising modular from their parent networks. Experimental results indicate that our proposed evolutionary NAS method has achieved competitive results in comparison with some state-of-the-art NAS methods. Moreover, the effectiveness of our proposed modular inheritable crossover operator for offspring generation is validated.