Towards a multipurpose neural network approach to novelty detection

Novelty detection, the identification of data that is unusual or different in some way, is relevant in a wide number of real-world scenarios, ranging from identifying unusual weather conditions to detecting evidence of damage in mechanical systems. However, utilising novelty detection approaches in a particular scenario presents significant challenges to the non-expert user. They must first select an appropriate approach from the novelty detection literature for their scenario. Then, suitable values must be determined for any parameters of the chosen approach. These challenges are at best time consuming and at worst prohibitively difficult for the user. Worse still, if no suitable approach can be found from the literature, then the user is left with the impossible task of designing a novelty detector themselves. In order to make novelty detection more accessible, an approach is required which does not pose the above challenges. This thesis presents such an approach, which aims to automatically construct novelty detectors for specific applications. The approach combines a neural network model, recently proposed to explain a phenomenon observed in the neural pathways of the retina, with an evolutionary algorithm that is capable of simultaneously evolving the structure and weights of a neural network in order to optimise its performance in a particular task. The proposed approach was evaluated over a number of very different novelty detection tasks. It was found that, in each task, the approach successfully evolved novelty detectors which outperformed a number of existing techniques from the literature. A number of drawbacks with the approach were also identified, and suggestions were given on ways in which these may potentially be overcome.