Recipient organisationKing's College LondonSource-published name: King's College London
Funding£335K
PeriodSept 2025 — Mar 2028
In plain English
AI plain-English summary
A new family of generative AI models, built from decision trees rather than neural networks, aims to make machine learning transparent enough for high-stakes fields like healthcare. Current deep neural networks excel at generating images and text, but they are black boxes: no one can fully explain why they produce a given output. They also struggle with mixed tabular data—the kind found in medical records—and require vast computing power. PRECISE-AI replaces neural networks with classification and regression trees, which are inherently interpretable and can handle messy, real-world datasets. These trees can be compiled into probabilistic circuits that answer exact queries—such as “what is the chance a patient has disease X given these symptoms?”—without approximation. If successful, the project could give clinicians, regulators, and patients a clear view of how an AI reaches its conclusions. The same models could generate synthetic patient data that preserves privacy, allowing researchers to share sensitive health records without compromising individual rights. They could also test whether apparent racial disparities in health outcomes stem from discrimination or from confounding factors. The project initially focuses on immune-mediated inflammatory diseases, but the approach could extend to any domain where trust, fairness, and explainability are non-negotiable.
View original technical description
The proposed project, Probabilistic REasoning with CIrcuits for Safe and Explainable Artificial Intelligence (PRECISE-AI), addresses the limitations of deep neural networks (DNNs) in generative modelling. While DNNs have excelled with structured data like images and text, they fall short in providing transparency, adaptability to unstructured data, and accessibility to practitioners due to their resource-intensive nature. PRECISE-AI aims to overcome these challenges by introducing a family of nonparametric generative models based on classification and regression trees. Trees offer several advantages, including interpretable explanations, adaptability to mixed tabular data, and ease of use. These benefits need not come at the cost of expressive power. Trees are universal approximators that often attain state of the art results on tabular data tasks. They can be efficiently compiled into circuits, providing exact, tractable inference for a range of common and important probabilistic queries. This project emphasises three key principles: privacy, explainability, and fairness, which are crucial in responsible AI but often neglected in unsupervised learning. For instance, in healthcare, where data is sensitive, a synthetic dataset generated with privacy guarantees could be shared with researchers without compromising patient rights. An explainable generative model could shed light on disease diagnosis by highlighting relevant biomarkers. A fair model could test whether an apparent disparity in health outcomes across racial groups is attributable to discrimination or confounding. By ensuring that generative models are private, explainable, and fair, PRECISE-AI will promote greater trust in machine learning, paving the way for safe adoption in high-risk domains. Beneficiaries of this research include practitioners, who will benefit from user-friendly, well-documented software; data subjects, whose rights to privacy, explanation, and fairness will be protected; and policymakers, who can make informed decisions regarding AI use in sensitive domains. Initially, the project will focus on healthcare applications, utilising biomolecular and clinical data from large studies on immune-mediated inflammatory diseases. In summary, PRECISE-AI proposes a novel, unified approach to generative modelling and probabilistic reasoning that prioritises issues of transparency, safety, and trust. Its potential applications extend to various high-risk domains, where the current state-of-the-art DNNs fall short. The project aims to provide accessible and responsible AI solutions with broad societal impact.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know