We use cookies to ensure that we give you the best experience on our website. By continuing to browse this repository, you give consent for essential cookies to be used. You can read more about our Privacy and Cookie Policy.

Durham e-Theses
You are in:

Control and Analysis for Sequential Information based on Machine Learning

ZHANG, PENG (2023) Control and Analysis for Sequential Information based on Machine Learning. Doctoral thesis, Durham University.



Sequential information is crucial for real-world applications that are related to time, which is same with time-series being described by sequence data followed by temporal order and regular intervals. In this thesis, we consider four major tasks of sequential information that include sequential trend prediction, control strategy optimisation, visual-temporal interpolation and visual-semantic sequential alignment. We develop machine learning theories and provide state-of-the-art models for various real-world applications that involve sequential processes, including the industrial batch process, sequential video inpainting, and sequential visual-semantic image captioning. The ultimate goal is about designing a hybrid framework that can unify diverse sequential information analysis and control systems

For industrial process, control algorithms rely on simulations to find the optimal control strategy. However, few machine learning techniques can control the process using raw data, although some works use ML to predict trends. Most control methods rely on amounts of previous experiences, and cannot execute future information to optimize the control strategy. To improve the effectiveness of the industrial process, we propose improved reinforcement learning approaches that can modify the control strategy. We also propose a hybrid reinforcement virtual learning approach to optimise the long-term control strategy. This approach creates a virtual space that interacts with reinforcement learning to predict a virtual strategy without conducting any real experiments, thereby improving and optimising control efficiency.

For sequential visual information analysis, we propose a dual-fusion transformer model to tackle the sequential visual-temporal encoding in video inpainting tasks. Our framework includes a flow-guided transformer with dual attention fusion, and we observe that the sequential information is effectively processed, resulting in promising inpainting videos.
Finally, we propose a cycle-based captioning model for the analysis of sequential visual-semantic information. This model augments data from two views to optimise caption generation from an image, overcoming new few-shot and zero-shot settings. The proposed model can generate more accurate and informative captions by leveraging sequential visual-semantic information.

Overall, the thesis contributes to analysing and manipulating sequential information in multi-modal real-world applications. Our flexible framework design provides a unified theoretical foundation to deploy sequential information systems in distinctive application domains. Considering the diversity of challenges addressed in this thesis, we believe our technique paves the pathway towards versatile AI in the new era.

Item Type:Thesis (Doctoral)
Award:Doctor of Philosophy
Faculty and Department:Faculty of Science > Computer Science, Department of
Thesis Date:2023
Copyright:Copyright of this thesis is held by the author
Deposited On:20 Oct 2023 15:54

Social bookmarking: del.icio.usConnoteaBibSonomyCiteULikeFacebookTwitter