Lectures on Machine Learning Theory by Williamson and Kaufmann

Centrum Wiskunde & Informatica (CWI) kindly invites you to a lecture afternoon focused on theory of machine learning, with two distinguished speakers that will illuminate philosophical and technical aspects of machine learning.

The program on 14 February
14.00  Bob Williamson (University of Tübingen): Foundations of Machine Learning Systems
15.00  Break
15.30  Emilie Kaufmann (CNRS, Univ. Lille): A Tale of Two Non-parametric Bandit Problems
16.30  Reception

These research-level lectures are intended for a non-specialist audience and are freely accessible. Please register here. The event takes place in the Amsterdam Science Park Congress Centre.

Bob Williamson
Foundations of Machine Learning Systems

  • Abstract: I will present some new insights into some foundational assumptions about machine learning systems, including why we might want to replace the expectation in our definition of generalisation error, why independence is intrinsically relative and how it is intimately related to fairness, why the data we ingest might not even have a probability distribution, and what one might do in such cases, and how we have been (perhaps unwittingly) working with these more exotic notions for some time already.

Emilie Kaufmann
A Tale of Two Non-parametric Bandit Problems

  • Abstract: In a bandit model an agent sequentially collects samples (rewards) from different distributions, called arms, in order to achieve some objective related to learning or playing the best arms. Depending on the application, different assumptions can be made on these distributions, from Bernoulli (e.g., to model success/failure of a treatment) to complex multi-modal distributions (e.g. to model the yield of a crop in agriculture). In this talk, we will present non-parametric algorithms which adapt optimally to the actual distributions of the arms, assuming that they are bounded. We will first show the robustness of a Non Parametric Thompson Sampling strategy to a risk-averse performance metric. Then, we will discuss how the algorithm can be modified to tackle pure exploration objectives, bringing new insights on so-called Top Two algorithms.