Research
Explainable, human-centric AI — building and optimising fuzzy inference systems to match deep-learning performance without sacrificing transparency.
Explainable, Human-Centric AI via Fuzzy Systems
My primary research line advances optimisation methods for fuzzy inference systems — inherently interpretable and uncertainty-aware — combining gradient-based, differentiable, and hybrid approaches to narrow the performance gap with modern deep learning while preserving the transparent reasoning that distinguishes them from black-box alternatives.
Key contributions include:
- Towards more flexible membership functions via B-splines —
- CASPs: Constraints Always Satisfied Parameters for gradient-based fuzzy set optimisation —
- Gradient-based optimisation via automatic differentiation —
- A comprehensive study of Type-Reduction algorithms for interval Type-2 fuzzy systems —
- FuzzyR: open-source Fuzzy Logic Toolkit for R, on CRAN —
- Novel switch-point determination methods for the Karnik–Mendel algorithm —
Community building. Co-organising the Special Session on Advances in Optimisation for Fuzzy Systems at FUZZ-IEEE / IEEE WCCI 2026 (Maastricht, 21–26 June 2026), with Dr Luca Ferranti (Aalto) and Prof. Jonathan Garibaldi. Earlier invited talk on this theme at the 1st CIHI Workshop @ IEEE WCCI 2024, Yokohama, and the FuzzyR tutorial at FUZZ-IEEE / IEEE WCCI 2022, Padua.
Deep Learning + Explainable AI Fusion
Deep learning achieves excellent accuracy but explainability and uncertainty quantification remain critical challenges. Fuzzy inference systems offer inherent interpretability (transparent rules, not post-hoc explanations) and built-in uncertainty handling, yet have historically been difficult to train end-to-end alongside neural networks.
Key contributions include:
- Graph neural networks for neurodegenerative disease diagnosis —
- Explainable AI for clinical decision support —
- Incorporating fuzzy uncertainty into deep neural networks for medical image segmentation —
This work aims to deliver AI models that are both high-performing and inherently interpretable — not black boxes wrapped in post-hoc explanations.
Evaluation & Quality Measures
A key question in AI research is how improvements can be more accurately measured. This line of work develops novel evaluation metrics, loss functions, and quality control frameworks to ensure AI systems are assessed rigorously and fairly.
Key contributions include:
- Fuzzy-based quality control for clinical AI —
- Boundary-aware loss functions using fuzzy rough sets —
- A new accuracy measure based on bounded relative error for time series forecasting —