Self-adaptive network structure tuning method based on NSGA-III

Abstract

Neural networks are arising a wave in the various areas of artificial intelligence and they have adopted in daily life successfully. The structure tuning of neural networks is crucial when building the relative models. The structure of neural networks is usually designed and tuned with experience and plenty of attempts. To reduce the difficulty and cost of structure tuning meanwhile improving its rationality, we propose a new method to tune the structure of neural networks adaptively. In this method, the related structure parameters are optimised. A many-objective algorithm is employed as the optimised tool to get a better structure. We design the experiments combining convolutional neural network (CNN) with Non-dominated Sorting Genetic Algorithm III (NSGA-III). The related experiments are conducted on the MNIST and Malware image datasets. Results show that the method has promising performance on neural networks tuning and can improve the robustness.