Title: Interactive Optimal Teaching with Unknown Learners: An Experimental Overview

Date: 12/07/2018

Time: 10h30

Location: Sala 1.4 do IST, Taguspark

Advisors: Professor Francisco Melo / Professor Manuel Lopes

Abstract: In this work we empirically test a new approach for machine teaching that partly overcomes the common mismatch between the knowledge the teacher has about the students and the actual learning process of the students. We analyze a specific situation where the student learning algorithm is known but the corresponding parameters are not. We focus on the case of Bayesian Gaussian learners, where the lack of knowledge regarding the student parameters significantly deteriorates the performance of machine teaching. In this new approach we explore how interactivity can mitigate the impact of imperfect knowledge, leading to significantly faster convergence of the number of samples needed to teach. We performed three user studies where the teacher asks samples to the learner in order to better estimate the behaviour and model of the learner, validating that interactivity can be one answer for the strong assumptions required in machine teaching. In the first two experiments we considered the case of single learners, with weak and strong prior (respectively). In the third one we extended teaching to groups of students. Interactivity showed to increase the learning performance in all of them, when comparing to the classical machine teaching approach.