Project Ideas
I am interested in supervising projects at the intersection of explainable AI, human-centric modelling, and deep learning. Below are some potential research directions — these are meant as starting points and can be adapted to suit your interests and level (UG, PG, or PhD).
If you are interested, please get in touch with a brief proposal via the contact details in the page footer.
Integration of Deep Learning Concepts with Fuzzy Systems
Many cutting-edge deep learning architectures, such as transformers, leverage modularised structures for their effectiveness. Could the foundational concepts of these deep learning structures be adapted to design similarly modular fuzzy systems? This approach could potentially bridge the performance gap between deep learning models and fuzzy systems.
Key Questions
How can transformer-style attention mechanisms be incorporated into fuzzy inference?
Can modular deep learning designs inspire more scalable fuzzy architectures?
Optimisation Techniques in Fuzzy Systems
Optimisation strategies like bagging, boosting, and stacking have significantly enhanced the performance of machine learning models. An interesting research avenue would be exploring the applicability and potential benefits of these strategies within fuzzy system optimisation.
Key Questions
How might ensemble techniques be tailored to fuzzy systems?
Can these strategies inspire entirely new optimisation methods for fuzzy models?
Preserving Interpretability in Deep–Fuzzy Hybrid Systems
As we incorporate deep learning techniques to boost the performance of fuzzy systems, a critical question arises: how do we ensure that these enhanced systems retain the high level of interpretability and explainability that is intrinsic to fuzzy logic? This topic invites exploration into balancing performance improvements with the preservation of these essential characteristics.
Key Questions
What is the trade-off boundary between performance gains and interpretability loss?
Can we define formal constraints that guarantee interpretability during optimisation?
Standardising Metrics for Interpretability and Explainability
Despite the recognised advantages of fuzzy systems in terms of interpretability and explainability, there lacks a consensus on metrics to quantify these qualities. Research in this domain could focus on developing or promoting standardised metrics, encompassing sub-topics such as enhancing model explainability and generating interpretable results.
Key Questions
Can we establish benchmark datasets and protocols for interpretability evaluation?
How do existing XAI metrics compare when applied to fuzzy vs. neural models?
Adapting Fuzzy Systems for Diverse Applications
While fuzzy systems have been traditionally associated with decision-making problems, expanding their application to areas such as image processing represents a novel research direction. Investigating methodologies to adapt fuzzy logic to such domains could significantly broaden the scope and impact of fuzzy systems.
Key Questions
How can fuzzy reasoning be effectively applied to high-dimensional visual data?
What domain adaptations are needed to make fuzzy systems competitive in new areas?
Development of Toolkits for Fuzzy System Research
Conducting research in these areas necessitates accessible toolkits tailored for designing, optimising, and evaluating fuzzy systems. With deep learning benefiting from comprehensive tools, creating or adapting similar resources for fuzzy system research could greatly facilitate advancements — for example, integrating features like PyTorch's automatic differentiation with fuzzy logic platforms.
Key Questions
How can gradient-based optimisation tools be extended to support fuzzy set parameters?
What toolkit features would most accelerate reproducible fuzzy systems research?
Expected Skills
A solid understanding of machine learning, with a particular focus on deep learning techniques.
Programming skills (e.g. Python or R) and experience in data analysis.
Prior knowledge of fuzzy systems is not required but would be advantageous.