Two-Stage Evolutionary Neural Architecture Search for Transfer Learning

Convolutional neural networks (CNNs) have achieved state-of-the-art performance in many image classification tasks. However, training a deep CNN requires a massive amount of training data, which can be expensive or unobtainable in practical applications, such as defect inspection and medical diagnosis. Transfer learning has been developed to address this issue by transferring knowledge learned from source domains to target domains. A common approach is fine-tuning, which adapts the parameters of a trained neural network for the new target task. Nevertheless, the network architecture remains designed for the source task rather than the target task. To optimize the network architecture in transfer learning, we propose a two-stage evolutionary neural architecture search for transfer learning (EvoNAS-TL), which searches for an efficient subnetwork of the source model for the target task. EvoNAS-TL features two search stages: 1) structure search and 2) local enhancement. The former conducts a coarse-grained global search for suitable neural architectures, while the latter acts as a fine-grained local search to refine the models obtained. In this study, neural architecture search (NAS) is formulated as a multiobjective optimization problem that concurrently minimizes the prediction error and model size. The knee-guided multiobjective evolutionary algorithm, a modern multiobjective optimization approach, is employed to solve the NAS problem. In this study, several experiments are conducted to examine the effectiveness of EvoNAS-TL. The results show that applying EvoNAS-TL on VGG-16 can reduce the model size by 52%–85% and simultaneously improve the testing accuracy by 0.7%–6.9% in transferring from ImageNet to CIFAR-10 and NEU surface detection datasets. In addition, EvoNAS-TL performs comparably to or better than state-of-the-art methods on the CIFAR-10, NEU, and Office-31 datasets.