DNA computing inspired deep networks design

Abstract

Deep neural networks have gained state-of-the-art results in many applications, such as pattern recognition and computer vision. However, most of the deep neural networks are designed manually by researchers. This architecture design process is generally time consuming and needs much expertise. Hence, automatic neural network design becomes an important issue. In this paper, we propose a novel method, called DNA computing inspired networks design (DNAND), to automatically learn high performance deep networks. In DNAND, we use DNA strands to represent blocks of a model, and these DNA strands are reacted to construct the overall networks according to the base pairing principle. We also present the killing strategy, with which we stop training “bad” models if they fail to reach the specific accuracy threshold on the validation set, so as to reduce the computational cost and accelerate the learning process. Extensive experiments on image classification and detection data sets demonstrate the effectiveness of the proposed method.