Transfer Learning Layer Selection Using Genetic Algorithm

Abstract

The performance of transfer learning in convolutional neural networks depends on the selection of which layers are to be learned again and which are not. Generally, layers selection is performed manually; however, as the number of layers increases, the layers selection process becomes increasingly difficult. Thus, we propose a method to select an effective layers in transfer learning automatically using a genetic algorithm. In the proposed method, a genotype representing which layers’ weights are updated or fixed in transfer learning is considered, and we achieve efficient layers selection in the way that a genotype with high validation accuracy is survived during genotype selection. Experiments are performed using the InceptionV3 network that pre-trained ImageNet as source images with transfer learning to CIFAR-100 as target images. Experimental results demonstrate that the test data accuracy in an ensemble of models whose layers are selected by the genetic algorithm is 15% and 12% greater than that of models trained by from-scratch and fine-tuning, respectively. In general transfer learning approach, layers on the output side are selected as adjustable layers; however, it is found that the distribution of the selected layers as an effective adjustable layers obtained by the genetic algorithm extends to the entire network. Transfer learning using a genetic algorithm may successfully capture the characteristics of a convolutional neural network’s structure.