Title: Tell me something I don’t know: the need for interpretation in computational intelligence
Abstract: Computational intelligence models are often evaluated on the basis of predictive performance, lacking appropriate consideration of other aspects than association which might make a claim to the intelligence of the model. Yet appearances can be deceiving, especially with single item performance measures e.g. AUROC. This is the case for non-linear models given their ability to exploit any weaknesses in the data, for instance structural artefacts which add a confounding effect over and above the presence of noise. In addition, models work well for well classified cases but cannot explain predictions for borderline cases. In other words, they confirm to expert users what they already know but do not add insights to the data in the difficult cases for which CI is most needed. The talk will illustrate some of the pitfalls in the design and validation of databased models. It will then explore principled approaches to interpreting neural networks using theoretical methodologies applied to the often opaque maximal separation models driven by computational learning theory and also probabilistic non-linear models from which the geometry of data spaces can be derived. Some important general questions will be explored including the derivation of nomograms for non-linear models, efficiency and interpretability of rule induction, but also a radically different approach to user interfaces for probabilistic classifiers by deriving statistically principled intelligent query systems for case-based reasoning. These models find particular application in clinical medicine where examples will illustrate tumour delineation and detection of response treatment from brain spectroscopy.
Biographical note: Paulo Lisboa is Professor and Head of Mathematics at Liverpool John Moores University. He chairs the Medical Data Analysis Task Force in the Data Mining Technical Committee of the IEEE-CIS and is vice-chair of the Advisory Group for Societal Challenge 1: Health, Demographics and Wellbeing in Horizon 2020 which is among the largest coordinated funding programme of health-related research worldwide. His research focus is machine learning explanation for computer-based decision support in healthcare, public health, sports analytics and computational marketing. He has over 250 refereed publications with awards for citations, serves on the editorial boards of a number of journal and evaluates extensively for RCUK and the European Commission.