Concurrent Evolution of Neural Networks and Their Data Sets

摘要

The ultimate joal of designing and training a neural network is optimizing the ability to minimize the expectation of the generalization error. Because active learning techniques can be used to find optimal complexity of network, active learning has emerged as an efficient alternative to improve the generalization performance of neural networks. In this paper, we propose an evolutionary approach that can design networks automatically through active data selection, where networks and data sets are evolved at the same time. Empirical results on regression and classification show improved generalization accuracy of the proposed approach for two real-world problems.