Design space exploration of Convolutional Neural Networks based on Evolutionary Algorithms

Abstract

This paper proposes a framework for design space exploration of
Convolutional Neural Networks (CNNs) using Genetic Algorithms
(GAs). CNNs have many hyperparameters that need to be tuned
carefully in order to achieve favorable results when used for image
classification tasks or similar vision applications. Genetic Algorithms
are adopted to efficiently traverse the huge search space
of CNNs hyperparameters, and generate the best architecture that
fits the given task. Some of the hyperparameters that were tested
include the number of convolutional and fully connected layers, the
number of filters for each convolutional layer, and the number of
nodes in the fully connected layers. The proposed approach was
tested using MNIST dataset for handwritten digit classification and
results obtained indicate that the proposed approach is able to generate
a CNN architecture with validation accuracy up to 96.66% on
average.