AIAI Seminar - 7 November 2022 - Talks by Filip Smola, Thomas Fletcher and Giannis Papantonis

Speaker:      Filip Smola

Title:           Resources and Processes in Isabelle/HOL

Abstract:

In this talk I will discuss the current state of our work on a theory of linear resources and process compositions, and its mechanisation in the proof assistant Isabelle/HOL. The correctness of the compositions is argued by linking them to deduction in linear logic, a connection which we also mechanise. Not only does the mechanisation let us verify the compositions, it also allows us to generate verified executable code implementing our definitions. I will talk about some of the complexities we found in this mechanisation, and then conclude on some of the ways in which we are currently extending this theory.

Speaker:      Thomas Fletcher

Title:            Statistics Automation in a Query-Answering System

Abstract:        

FRANK is a multi-domain Query Answering (QA) system which performs inferential and simple statistical reasoning on data which it automatically identifies and retrieves (e.g. to provide predictions). SMART is a statistics advisor system designed to allow FRANK to answer queries of types which have not previously been within the purview of QA systems, including statistical significance, functional shape description and analysis of variance. The combined SMART FRANK system allows users to quickly obtain answers to data-oriented questions involving choices of analysis, modelling and visualisation which are tedious for an expert and beyond the knowledge of a novice.

Speaker:      Giannis Papantonis

Title:                Trust in AI: The role of users' epistemic state

Abstract:

As AI models get increasingly integrated in practical applications, it is important to develop practices that encourage an effective human-machine synergy. To this end, a series of recent approaches study the impact of different uncertainty measures on calibrating users' trust towards a model, however the findings so far have been inconclusive. In this work we build upon such approaches, while attempting to expand their scope by taking into account the effect of a user's own epistemic state. Moreover, instead of providing users with a single uncertainty measure, we advocate in favour of blending multiple sources of information to more accurately convey a model's decision-making process.