Growing an architecture for a neural network

Abstract

We propose a new kind of automatic architecture search algorithm. The algorithm alternates pruning connections and adding neurons, and it is not restricted to layered architectures only. Here architecture is an arbitrary oriented graph with some weights (along with some biases and an activation function), so there may be no layered structure in such a network. The algorithm minimizes the complexity of staying within a given error. We demonstrate our algorithm on the brightness prediction problem of the next point through the previous points on an image. Our second test problem is the approximation of the bivariate function defining the brightness of a black and white image. Our optimized networks significantly outperform the standard solution for neural network architectures in both cases.