Quantum-Inspired Evolutionary Algorithm for Convolutional Neural Networks Architecture Search

Convolutional neural networks (CNN) are widely used and effective deep learning methods for image classification tasks. But the architecture of CNN such as LetNet and AlexNet were designed elaborately by experts because designing the neural networks is time-consuming and requires expert knowledge. This paper proposed a quantum-inspired evolutionary algorithm to search the neural architectures. First, we encode CNNs into quantum chromosomes and distinguish these chromosomes from the Convolutional Layer, Pooling Layer, Fully-connected Layer and Disabled Layer with its range. Second, quantum chromosomes are updated by applying quantum gates and find the best individual with quantum genetic algorithm. Third, we can predict the network performance after a few steps of stochastic gradient descent by means of evaluation estimate strategy so that we can stop training the bad networks early, which can speed up evolutionary process. The proposed algorithm is examined and compared with some state-of-art methods for image classification in three benchmark datasets. The experimental results prove the proposed algorithm can search a strong classifier robustly. In addition, it performs better than the general evolutionary algorithm. More importantly, with the help of evaluation estimate strategy, it is substantially faster than the algorithms without evaluation estimate strategy which means we can take less time to search a good network for the given task.