Image Classification with Evolved Convolutional Neural Networks

Abstract

Convolutional neural networks (CNNs) are a well-established technique for image classification problems. While the topology of a CNN strongly affects the performance of that CNN, designing a CNN’s topology remains a difficult task, often with nothing better than some empirical rules-of-thumb for guidance. Evolutionary algorithms are a family of metaheuristics that can be applied to optimization problems where good solutions are hard to create from first principles, but the quality of a given solution is easy to measure. In this research, we develop and evaluate several variations on an algorithm, SDAG, which applies evolutionary methods to finding performant topologies for CNN-based image classifiers.