Mingshu joined Rui to discuss spatially aware AI on The Science Show with Dr Dada
On 29 July 2026, Dr Mingshu Wang joined doctoral researcher Rui Deng for a live interview on The Science Show with Dr Dada at Jambo! Radio Scotland. The episode, titled “Can AI Understand Where It Is?”, explored their recent research on spatially aware artificial intelligence.
Artificial intelligence can generate text, write code, and identify patterns across very large datasets. However, it does not necessarily recognise why two neighbouring places may behave differently. This question provided the starting point for the discussion of TabPFN-GSA, a spatially explicit approach designed to give a tabular foundation model a practical mathematical “sense of place”.
The programme was hosted by Dr Adetunmise Dada, Lecturer in Optics and Head of the EQOQI Lab at the University of Glasgow. The discussion focused on research by Rui Deng, Dr Ziqi Li, and Dr Mingshu Wang, published in the International Journal of Geographical Information Science under the title “Do foundation models work for geospatial tabular data? An investigation of TabPFN and a proposed enhancement based on geospatial sparse attention”.
The study investigates whether general-purpose tabular foundation models can work effectively with geospatial tabular data. The results show that TabPFN performs strongly across many geospatial regression tasks, particularly on smaller datasets. However, its performance can deteriorate as datasets become larger and when spatial dependence is especially strong and localised.
To address these limitations, the researchers developed Geospatial Sparse Attention (GSA), an inference-time strategy that prioritises observations from nearby locations while retaining selected information from more distant areas. This introduces geographical knowledge into the model without retraining or redesigning the underlying foundation model, improving predictive robustness and scalability.
During the interview, Mingshu and Rui discussed how location influences prediction, why coordinates alone do not necessarily give an AI model an understanding of spatial relationships, and how geographical knowledge can support the development of modern foundation models. They also explained how Geospatial Sparse Attention concentrates computational attention on locally relevant information while preserving selected wider connections.
The conversation also considered broader questions surrounding intelligence, consciousness, and superintelligence. A key distinction is that a model can become more context-aware without becoming self-aware. In this research, a “sense of place” refers to a model’s ability to use spatial context more appropriately. It does not imply human understanding, lived experience, or consciousness.
“We are not teaching AI to experience place; we are teaching a prediction model to respect spatial structure.”
The interview further explored potential applications in environmental monitoring, housing, public health, and socioeconomic analysis, alongside limitations relating to bias, privacy, interpretation, and responsible use. More accurate prediction does not automatically provide causal explanations or guarantee fair outcomes. Spatially aware AI must therefore remain grounded in domain expertise, transparent evaluation, and human judgement.
Mingshu is grateful to Dr Dada and the Jambo! Radio Scotland team for the invitation and for an engaging and wide-ranging conversation. Jambo! Radio is a community radio station serving people of African and Caribbean heritage across Glasgow, Edinburgh, and Aberdeen through DAB, online broadcasting, and its mobile application.
The full episode, “Can AI Understand Where It Is?”, is available on YouTube.
