AI 2.0: Augmented Intelligence

Abstract

Computational intelligence (CI) models are often evaluated on the basis of predictive performance, lacking appropriate consideration of other aspects which might make a claim to the intelligence of the model and which can be critical for their use by a subject expert who is not a CI expert. Yet appearances can be deceiving, especially with summary performance measures e.g. AUROC. This is especially the case for non-linear models given their ability to exploit any weaknesses in the data, for instance structural artefacts which can add confounding effects. In addition, many applied CI 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. There is a drive for the use of CI to complement rather than automate decision making This is fundamental to make CI useful in practice and has been termed Augmented Intelligence, or AI 2.0. The talk will illustrate some of the pitfalls in the design and validation of databased models. It will then describe how rules can be efficiently derived from neural networks so openi ng the black-box. An alternative and popular way of presenting and using complex models e.g. to clinicians, is the use of nomograms. They will be derived from SVMs so extending this graphical approach to non-linear models. Finally, the concept of case-based reasoning will be explored using information geometry to calculate similarity metrics directly to identify patients-like-mine with reference to specific clinical queries e.g. diagnosis or prognosis. This comprises a statistically principled intelligent query system for case-based reasoning, enabling a subject expert to diagnose probabilistic classifiers with respect to patient cohorts where there are significantly more or fewer cases of interest, separating them from mixed groups for whom more information is certain to be required. This provides a direct route to interpretation and a way for subject experts to access generic non-linear models as a smart approach to data retrieval, complementing the numerical outputs.