A transfer learning based evolutionary deep learning framework to evolve convolutional neural networks

Bin Wang, Bing Xue, Mengjie Zhang

Abstract: The manual design of CNNs has become exceptionally complex due to the more sophisticated CNN architectures. Thankfully, more and more researchers endeavour to mitigate the difficulty of manual design by designing automated process, but the computational cost of the automatic methods is extremely high due to the huge search space. In this paper, an evolutionary deep learning framework based on transfer learning is proposed to reduce the computational cost, while maintaining the classification at a competitive level. The main idea is to evolve a CNN block from smaller datasets, and then increasing the capacities of the evolved block to handle larger datasets. The proposed method obtains good CNNs with less than 40 GPU-hours. It also achieves a promising error rate of 3.46% on the CIFAR-10 dataset.