Abstract
dc:description.abstractWith the rising use of Artificial Intelligence (AI) and Machine Learning (ML) methods, there comes an increasing need to understand how automated systems make decisions. Interpretable ML provides insight into the underlying reasoning behind AI and ML models while not stifling their predictive performance. Doing so is important for many reasons, such as facilitating trust, increasing transparency, and providing improved collaboration and control through a better understanding of automated decision-making. Interpretability is very relevant across many ML paradigms and application domains. Multiple Instance Learning (MIL) is an ML paradigm where data are grouped into bags of instances, and only the bags are labelled (rather than each instance). This is beneficial in alleviating expensive labelling procedures and can be used to exploit the underlying structure of data. This thesis investigates how interpretability can be achieved within MIL. It begins with a formalisation of interpretable MIL, and then proposes a suite of model-agnostic post-hoc methods. This work is then extended to the specific application domain of high-resolution satellite imagery, using novel inherently interpretable MIL approaches that operate at multiple resolutions. Following on from work in the vision domain, new methods for interpretable MIL are developed for sequential data. First, it is explored in the domain of Reward Modelling (RM) for Reinforcement Learning (RL), demonstrating that interpretable MIL can be used to not only understand a model but also improve its predictive performance. This is mirrored in the application of interpretable MIL to Time Series Classification (TSC), where it is integrated into state-of-the-art methods and is able to improve both their interpretability and predictive performance. The integration into existing models to provide inherent interpretability means these benefits are delivered with little additional computational cost. <br/>
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Early, Joseph Arthur
- Advisors dc:contributor.advisor
-
- Ramchurn, Gopal
- Evers, Christine
- ORCID person:identifier.orcid
Rights
- Language dc:language
- English, eng