画像認識のための階層型ニューラルネットワークの進化的構造最適化

概要

論文の詳細を見る
The purpose of this paper is to optimize the structure of hierarchical neural network. In this paper, structure optimization is to represent neural network by minimum number of nodes and connections, and is performed by eliminating unnecessary connections from trained neural network by using genetic algorithm. We focus on the neural network which specialized for image recognition problems. The flow of the proposed method is as follows. Firstly, walsh-hadamard transform is applied to images for feature extraction. Secondly, neural network is trained with extracted features based on back-propagation algorithm. After neural network training, unnecessary connections are eliminated from trained neural network by utilizing genetic algorithm. Finally, neural network is retrained to recover the degradation caused by connection elimination. In order to validate the usefulness of the proposed method, face recognition and texture classification examples are used. From the experimental results, it was shown that compact neural network was generated, keeping generalization performance by proposed method.