Two Novel Performance Improvements for Evolving CNN Topologies

TL;DR

This paper introduces two innovative methods to improve the efficiency of evolving CNN topologies using genetic algorithms, reducing training complexity and time by nearly 20% while maintaining accuracy on CIFAR10.

Contribution

The paper presents two novel techniques—regularisation on training time and partial training for early ranking—that significantly reduce computational costs in CNN evolution.

Findings

Training time reduced by nearly 20%

Maintained accuracy on CIFAR10

Effective early ranking of architectures

Abstract

Convolutional Neural Networks (CNNs) are the state-of-the-art algorithms for the processing of images. However the configuration and training of these networks is a complex task requiring deep domain knowledge, experience and much trial and error. Using genetic algorithms, competitive CNN topologies for image recognition can be produced for any specific purpose, however in previous work this has come at high computational cost. In this work two novel approaches are presented to the utilisation of these algorithms, effective in reducing complexity and training time by nearly 20%. This is accomplished via regularisation directly on training time, and the use of partial training to enable early ranking of individual architectures. Both approaches are validated on the benchmark CIFAR10 data set, and maintain accuracy.