Neural networks can be viewed through a statistical lens as (non-linear) regression models, where the inputs variables predict some output variable through a series of weighted sums of non-linear functions. They are a highly flexible class of models that have been successful in the prediction of various complex problems. However, interpretation of these models can be difficult, primarily due to their complexity. In this talk, we look into what neural networks actually are; how they relate to (simpler) statistical models; how to adapt their complexity using traditional stepwise statistical procedures; and how to incorporate uncertainty quantification (i.e., p-values) and covariate-effect visualisations to shift the focus away from black-box prediction. Overall, the methods presented help to align neural networks more with other statistical modelling approaches. Note: it is not expected that the audience will have knowledge/experience with neural networks.
Kevin Burke: Understanding neural networks as statistical models
Dr Kevin Burke is an Associate Professor of Statistics at the University of Limerick (UL) with research interests in flexible statistical modelling, penalised regression, neural networks, random effects, survival/reliability analysis, and industrial modelling. See https://kevinburke.ie/ for more details. Previously he was Director of UL's Industrial Mathematics Unit (from 2016 - 2020), and a Science Foundation Ireland Public Service Fellow with the Food Safety Authority of Ireland (2020 - 2021). He is the Vice President of the Irish Statistical Association, and the Chair of the International Workshop on Statistical Modelling (IWSM) 2025. IWSM 2025 will be hosted in Limerick City, Ireland, from Sunday 13th July to Friday 18th July 2025; the call for papers is currently open (https://iwsm2025.ie/call-for-papers).
