A Hybrid Evolutionary Algorithm, Utilizing Novelty Search and Local Optimization, Used to Design Convolutional Neural Networks for Handwritten Digit Recognition

Tabish Ashfaq, Nivedha Ramesh, Nawwaf Kharma

2021

Abstract

Convolutional neural networks (CNNs) are deep learning models that have been successfully applied to various computer vision tasks. The design of CNN topologies often requires extensive domain knowledge and a high degree of trial and error. In recent years, numerous Evolutionary Algorithms (EAs) have been proposed to automate the design of CNNs. The search space of these EAs is very large and often deceptive, which entails great computational cost. In this work, we investigate the design of CNNs using Cartesian Genetic Programming (CGP), an EA variant. We then augment the basic CGP with methods for identifying potential/actual local optima within the solution space (via Novelty Search), followed by further local optimization of each of the optima (via Simulated Annealing). This hybrid EA methodology is evaluated using the MNIST data-set for handwritten digit recognition. We demonstrate that the use of the proposed method results in considerable reduction of computational effort, when compared to the basic CGP approach, while still returning competitive results. Also, the CNNs designed by our method achieve competitive recognition results compared to other neuroevolutionary methods.

Paper Citation

in Harvard Style

Ashfaq T., Ramesh N. and Kharma N. (2021). A Hybrid Evolutionary Algorithm, Utilizing Novelty Search and Local Optimization, Used to Design Convolutional Neural Networks for Handwritten Digit Recognition. In Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021) - Volume 1: IJCCI; ISBN 978-989-758-534-0, SciTePress, pages 123-133. DOI: 10.5220/0010648300003063

in Bibtex Style

@conference{ijcci21,
author={Tabish Ashfaq and Nivedha Ramesh and Nawwaf Kharma},
title={A Hybrid Evolutionary Algorithm, Utilizing Novelty Search and Local Optimization, Used to Design Convolutional Neural Networks for Handwritten Digit Recognition},
booktitle={Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021) - Volume 1: IJCCI},
year={2021},
pages={123-133},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010648300003063},
isbn={978-989-758-534-0},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 13th International Joint Conference on Computational Intelligence (IJCCI 2021) - Volume 1: IJCCI
TI - A Hybrid Evolutionary Algorithm, Utilizing Novelty Search and Local Optimization, Used to Design Convolutional Neural Networks for Handwritten Digit Recognition
SN - 978-989-758-534-0
AU - Ashfaq T.
AU - Ramesh N.
AU - Kharma N.
PY - 2021
SP - 123
EP - 133
DO - 10.5220/0010648300003063
PB - SciTePress