PhD opportunities in Computer Science

Generative AI for multimodal interaction, embodied AI, and safe collaborative robotics

Supervisor: Oliver Lemon

This project targets the intersection of Generative AI, Multimodal Interaction, and Embodied Collaborative Agents. The core challenge is to create safe, robust, and highly effective embodied AI capable of seamless multimodal collaboration with humans and other agents in dynamic, open-ended environments, with applications in domains such as healthcare and assisted living.The project will focus on fundamentally advancing Vision, Language, and Action (VLA) models for collaborative embodied AI, working on research problems such as:

• Robust Frameworks for Ad-Hoc Teamwork: Developing reliable, language-driven frameworks for multi-agent collaboration.-
• Safe Multimodal Conversational Collaboration: Integrating generative VLA models to enable safe, natural, and efficient conversational interaction with robots and AI systems, using speech, vision, and action.
• Embodied Communication and Action: Enabling agents to comprehend visual cues, interpret and generate contextually appropriate natural language, and execute coordinated physical actions.
• Advanced Dialogue Systems: Research in spoken and multimodal dialogue systems to enhance the fluidity and robustness of Conversational AI.

This project offers an opportunity to define the next generation of AI-driven robotics creating useful, safe, and collaborative multimodal agents.

Knowledge Graphs for Biomedical Image Data Management, Integration and Analysis

Supervisor: Albert Burger

Description: Biomedical images play a key role in research as well as clinical practice. A variety of imaging techniques are now available to study healthy as well as diseased tissue in human and model organisms, ranging from cell level to entire organs and organisms. In the clinical context, imaging diagnostics, such as MRI and CT scans, are now routinely used to better understand a patient's medical condition. To be able to take full advantage of the large image respositories that result from this, the information embedded in such images must be computationally represented. This is relevant for image retrieval, data integration and machine learning applications. Knowledge Graphs, a combination of graph-based data management and artificial intelligence, have emerged as a primary technique to model and capture such information, but many questions remain in terms of how to model the information in biomedical images and how to integrate and prepare them for machine learning based studies.  A variety of PhD topics are available in this area, including using knowledge graphs to model anatomical knowledge, integrating cinical image data sets across multiple patients, graph-based machine learning for biomedical images, and using Natural Language Processing techinques to integrate images with medical reports and publications.  If you are interested in discussing any of the above or other related topics for PhD study, please contact Prof Albert Burger.

Fair Graph Neural Architecture Search

Supervisor: Wei Pang

Description: This project will develop robust graph neural networks which can make fair and effective decisions. To achieve this, we will use multi-objective optimization techniques to make trade-off design solutions for graph neural networks. We will also explore the use of bio-inspired algorithms, such as evolutionary computing and immune-inspired computing algorithms to facilitate the search for suitable graph neural architectures.

Vision-based surgical robot for interventional vascular surgery

Supervisor: Chengjia Wang

Description: Introduction of robotic procedures has brought significant benefits to intervention vascular surgery with remotely conducted precise procedures and reduced exposure to radiation for clinicians. In the meantime, renaissance of AI in the past decade has fundamentally reshaped many fields with escalated performance in computer vision, natural language processing and other tasks. However, many challenges exists in robot aided vascular surgery, for example, lack of annotated data, reduced sensor feedback, requirement of real-time intra-operation prediction etc.. In this project, you will explore the development of autonomous surgical robot which combining multi-modal sensor data to assist interventional vascular surgery. This involves investigation of frontier designs of surgical robotics tool, development and exploration of advanced deep learning models, fast fusion and visualisation of multi-modality sensory data (e.g., haptics, digital subtraction angiography, and other vascular imaging data), and a full understand of operational procedures in the treatment of vascular disease.

Robot aid autonomous retrosynthesis for advanced drug discovery

Supervisor: Chengjia Wang

Description: As a fundamental problem in chemistry, Retrosynthesis is the process of decomposing a target molecule into readily available starting materials. It aims at designing reaction pathways and intermediates for a target compound. The goal of artificial intelligence (AI)-aided retrosynthesis is to automate this process by learning from the previous chemical reactions to make new predictions. Although several models have demonstrated their potentials for automated retrosynthesis, there is still a significant need to further enhance the prediction accuracy to a more practical level. This project aims to review, implement and review existing retrosynthesis methods and exploring the development of an autonomous verification robot for retrosynthecial process.

Decentralised, privacy-preserving and explainable federated learning for annotation effective medical image analysis

Supervisors: Chengjia Wang and Wei Pang

Description: Several infectious diseases have attacked human society since the beginning of Since the beginning of this millennium, in the meantime advanced AI models have been widely deployed to assist a variety of tasks for the fight against these global pandemics. Modern clinical studies, especially for popular disease that affect a large cohort of patients, are often requires a collaboration among multiple research centres with large-scale machine learning models applied to analyse distributedly stored clinical data acquired at different sites. Protection of privacy, effective model supervision and interpretability of prediction results are critical in clinical practices while a trade-off exists between any two of them. Furthermore, the required computing power for training the machine learning models can easily exceed the capacity of modern computing devices in clinical centres. This project aims to investigate training and deployment of decentralised federated learning scheme for large scale deep learning models with good adhoc or posthoc interpretability and minimal data exchange between sites. Specifically, the student will explore and development frontier research of image understanding, federated learning, decentralised semi- or weakly- supervise, model compression, heterogeneous data/model distillation, domain adaptation, XAI and possibly advanced blockchain technology.

Swarm bots and swarm intelligence in healthcare internet of things

Supervisors: Chengjia Wang and Wei Pang

Description: Internet of things offer a number of new opportunities to improve patient outcomes through real-time patient monitoring, advanced diagnostics, robotic surgery, and much more. However, the computing power of the associated IoT devices can limit the performance and scale and inference speed of adoptable machine learning models in practice. This is common in complicated robotic surgery and monitoring of high-risk patients, e.g., in ICU. In this project, we aim at development of swarm intelligence methods, empowered by modern deep learning models, that can coordinate a group of intelligent devices for event recognition and decision making. This involves deliberate management of computing capacity, communication, collaborated training and abnormality detection among a heterogeneous group of devices, and further developing the theoretical model into a group of well coordinated swarm bots.

Reference: C.Wang et al., Industrial Cyber-Physical Systems-based Cloud IoT Edge for Knowledge Distillation, IEEE-TII 2021

Enhancing Confidence in Topic Models using Ensembles

Supervisor: Pierre Le Bras

Description: Topic modelling techniques typically rely on stochastic methods to explore large corpora and solve the multi-dimensional problem of identifying the themes discussed in documents in an unsupervised fashion. These approaches then propose one sample solution, out of many equally likely other solutions. While this is acceptable in most scenarios, confirmed users can still challenge the solution presented to them.

This project aims to address such limitation by investigating the use of ensemble methods in topic modelling. While some techniques have been described, such as term co-association and topic alignment, ensemble modelling offers other methods that could be applied or adapted to topic modelling. The initial stages of this project will mostly involve the development of data mining algorithms; however, the final goal should be to also investigate means of intuitive visualisations of ensemble topic models and evaluate their impact on user acceptance and confidence.

Machine/Deep Learning Applied to Neurodegenerative Diseases

Supervisor: Marta Vallejo

Description: Are you interested in the intersection between neurodegenerative diseases and machine/deep learning?  At the ML-Health group, we are offering PhD projects that contribute to understanding and combating neurodegenerative diseases.

Project 1: Advancing Diagnosis, Prognosis and Disease Progression in Parkinson’s and Huntington’s Disease. Join us in exploring the potential of wearable data for improving the diagnosis, prognosis, and disease progression of Parkinson's and Huntington's diseases. Collaborative opportunities exist with the University of York and partners in the Netherlands, Germany, and Australia.

Project 2: Unravelling ALS/Motor Neuron Disease and Protein Aggregation. Work closely with experts from the University of Edinburgh and Aberdeen University to investigate protein aggregation using brain imaging techniques in ALS/Motor Neuron Disease.

Relevant Techniques:

• Classification (including CNNs, LSTM..)
• Segmentation (such as U-Net, Yolo style networks)
• Data augmentation (utilising GANs)
• Disease representation (Graph neural networks)
• Network optimisation (Neuroevolution)
• Interpretability

We also encourage exploration of other deep learning-related areas.

Fair and Ethical Human-Centred NLP

Supervisor: Gavin Abercrombie

Description: From supervised classification and reinforcement-learning with human feedback (RLHF) to the data used to train large language models (LLMs), natural language processing (NLP) relies on data created by humans, but often produces inequitable results for different people. This project will combine computer science, linguistics, cognitive science, and human-centred design practices to investigate annotation and data practices and language modelling algorithms in order to create fair and ethical NLP systems.