DUAN, HAORAN (2024) Attention Mechanism for Adaptive Feature Modelling. Doctoral thesis, Durham University.
| PDF - Accepted Version 21Mb |
Abstract
This thesis presents groundbreaking contributions in machine learning by exploring and advancing attention mechanisms within deep learning frameworks. We introduce innovative models and techniques that significantly enhance feature recognition and analysis in two key application areas: computer vision recognition and time series modeling. Our primary contributions include the development of a dual attention mechanism for crowd counting and the integration of supervised and unsupervised learning techniques for semi-supervised learning. Furthermore, we propose a novel Dynamic Unary Convolution in Transformer (DUCT) model for generalized visual recognition tasks, and investigate the efficacy of attention mechanisms in human activity recognition using time series data from wearable sensors based on the semi-supervised setting.
The capacity of humans to selectively focus on specific elements within complex scenes has long inspired machine learning research. Attention mechanisms, which dynamically modify weights to emphasize different input elements, are central to replicating this human perceptual ability in deep learning. These mechanisms have proven crucial in achieving significant advancements across various tasks.
In this thesis, we first provide a comprehensive review of the existing literature on attention mechanisms. We then introduce a dual attention mechanism for crowd counting, which employs both second-order and first-order attention to enhance spatial information processing and feature distinction. Additionally, we explore the convergence of supervised and unsupervised learning, focusing on a novel semi-supervised method that synergizes labeled and unlabeled data through an attention-driven recurrent unit and dual loss functions. This method aims to refine crowd counting in practical transportation scenarios.
Moreover, our research extends to a hybrid attention model for broader visual recognition challenges. By merging convolutional and transformer layers, this model adeptly handles multi-level features, where the DUCT modules play a pivotal role. We rigorously evaluate DUCT's performance across critical computer vision tasks. Finally, recognizing the significance of time series data in domains like health surveillance, we apply our proposed attention mechanism to human activity recognition, analyzing correlations between various daily activities to enhance the adaptability of deep learning frameworks to temporal dynamics.
Item Type: | Thesis (Doctoral) |
---|---|
Award: | Doctor of Philosophy |
Faculty and Department: | Faculty of Science > Computer Science, Department of |
Thesis Date: | 2024 |
Copyright: | Copyright of this thesis is held by the author |
Deposited On: | 22 Feb 2024 10:33 |