Investigating modularity and transparency within bioinspired connectionist architectures using genetic and epigenetic models

Lacey, George Alexander

Computer science
June 2022

Thesis or dissertation


Rights
© 2022 George Alexander Lacey. All rights reserved. No part of this publication may be reproduced without the written permission of the copyright holder.
Abstract

Machine learning algorithms allow computers to deal with incomplete data in tasks such as speech recognition and object detection. Some machine learning algorithms take inspiration from biological systems due to useful properties such as robustness, allowing algorithms to be flexible and domain agnostic. This comes at a cost, resulting in difficulty when one attempts to understand the reasoning behind decisions. This is problematic when such models are applied in realworld situations where accountability, legality, and maintenance are of concern. Artificial gene regulatory networks (AGRNs) are a type of connectionist architecture inspired by gene regulatory mechanisms. AGRNs are of interest within this thesis due to their ability to solve tasks in chaotic dynamical systems despite their relatively small size.
The overarching aim of this work was to investigate the properties of connectionist architectures to improve the transparency of their execution. Initially, the evolutionary process and internal structure of AGRNs were investigated. Following this, the creation of an external control layer used to improve the transparency of execution of an external connectionist architecture was attempted.
When investigating the evolutionary process of AGRNs, pathways were found that when followed, produced more performant networks in a shorter time frame. Evidence that AGRNs are capable of performing well despite internal interference was found when investigating their modularity, where
it was also discovered that they do not develop strict modularity consistently. A control layer inspired by epigenetics that selectively deactivates nodes in trained artificial neural networks (ANNs) was developed; the analysis of its behaviour provided an insight into the internal workings of the ANN.

Publisher
Department of Computer Science and Technology, The University of Hull
Qualification level
Doctoral
Qualification name
PhD
Language
English
Extent
659 KB
Identifier
hull:18561
QR Code