E-Thesis 292 views 349 downloads
Human Interfaces with Machine Learning Recognition Systems / CONNOR CLARKSON
Swansea University Author: CONNOR CLARKSON
-
PDF | E-Thesis – open access
Copyright: The author, Connor Clarkson, 2025 Distributed under the terms of a Creative Commons Attribution 4.0 License (CC BY 4.0).
Download (6.95MB)
DOI (Published version): 10.23889/SUThesis.69942
Abstract
Building large pools of data has become a relatively straightforward task, with many automated ways of obtaining different sources of data. Labelling such data has resulted in becoming an exponential problem, both in terms of time and in the form of an interaction-heavy task. This task only becomes...
| Published: |
Swansea University, Wales, UK
2025
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Xianghua, X., and Edwards, M. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa69942 |
| Abstract: |
Building large pools of data has become a relatively straightforward task, with many automated ways of obtaining different sources of data. Labelling such data has resulted in becoming an exponential problem, both in terms of time and in the form of an interaction-heavy task. This task only becomes exponential with feature-rich structures of data and labelling systems, as well as requiring more advanced expertise for many different domains of a task to model. A prominent set of techniques utilising this data, and large networks have reformed machine learning into what we call deep learning today. Within this field, we can form levels of supervision that allow for stronger signals of inductive bias for both deep network architectures and in the training scheme. In this work, we explore both types with the target application and domain being the manufacturing of steel. Firstly, we present an exploratory approach to assist in decision-making for the task of clustering by utilising the feature-rich representations provided by generative models. By forming it as a semi-supervised problem we can provide varying degrees of supervision to enhance performance as a form of inductive bias into the training scheme. Supervision can be formalised into labels from data or in an active learning setting where we request help from an expert. If we are required to make a request, then we must provide information and visualisations so that an accurate decision can be made. Following this, in our second body of work we extend on an active learning setting by introducing a new acquisition function based on the distance from different representations. We apply it to a data refinement strategy where we fix mistakes in bounding-box labelled datasets to form a dense segmentation. Different forms of user interaction provide different levels of information to the training scheme, we explore the effects of these user interactions on the performance of this refinement task. Lastly, we apply stronger forms of inductive bias into the network architecture by modelling hierarchical labelling systems, where such relationships between labels form an abstraction and fine-grained level of the data. Inspired by the structure of human cognition and perception where we recognise patterns of various levels of abstraction to define an object. By invoking an explicit form of deep learning with feature-rich structures like graphs we can model these interconnected labels. We define two types of hierarchical relationships: the first is a break-up of the physical or geometric structure of the object, referred to as an encapsulation relationship. The second is sub-classification relationships which are semantic relations of labels provided by domain knowledge of what we are trying to capture in the dataset. We utilise both to solve classification and segmentation tasks. |
|---|---|
| Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information. |
| College: |
Faculty of Science and Engineering |
| Funders: |
EPSRC Centre For Doctoral Training in Enhancing Human Interactions and Collaborations with Data and Intelligence Driven Systems |