Automating Autoencoder Architecture Configuration: An Evolutionary Approach

Abstract

Learning from existing data allows building models able to classify patterns, infer association rules, predict future values in time series and much more. Choosing the right features is a vital step of the learning process, specially while dealing with high-dimensional spaces. Autoencoders (AEs) have shown ability to conduct manifold learning, compressing the original feature space without losing useful information. However, there is no optimal AE architecture for all datasets. In this paper we show how to use evolutionary approaches to automate AE architecture configuration. First, a coding to embed the AE configuration in a chromosome is proposed. Then, two evolutionary alternatives are compared against exhaustive search. The results show the great superiority of the evolutionary way.