Learning objectives

Machine Teaching studies how efficiently a teacher can guide a learner to acquire a target hypothesis.
The classic works date back to the 1990’s [Shinohara91,Goldman95] consider the setting where the teacher sends in one shot a set of labeled examples to the learner, who then has to output the correct target hypothesis. In the more recent studies, the focus has been on the interactive setting, where the Teacher and Leaner interact over multiple rounds. In each round, the teacher sends examples to the learner, who returns some feedback; this process continues until the learner reaches the target hypothesis (or a good approximation of it). Machine teaching models have proved useful in several contexts, e.g., crowd sourcing, intelligent tutoring systems, analysis of training set attacks. Moreover, commercial tools are under development by the Microsoft Machine Teaching Group, as detailed on their web page, which are based on, or employ, the paradigm of machine teaching, e.g., PICL, which leverages the selection of examples that maximize the training value of the interaction with the teacher; LUIS for natural language understanding; and other projects on building models for autonomous systems, and tools enabling non-experts of machine learning to build their models.

Prerequisites and basic notions

Basic knowledge of algorithm analysis and discrete probability

Program

Foundations: From PAC learning to Active learning, to Machine Teaching; Teaching dimension concepts (batch, sequential, recursive, VC-dimension and sample compression); Interactive Machine Teaching and Black Box machine teaching; Application: human/robot/computer interaction, training-set attacks, crowdsourcing.

Bibliography

Didactic methods

Lectures (blackboard and slides)

Learning assessment procedures

Reading assignments and oral discussion

Assessment

Understanding of the basic concepts and ability to apply them in new contexts

Criteria for the composition of the final grade

The result will be Pass/Fail