| Links' Seminars and Public Events |   |
| 2021 | |
|---|---|
Fri 19th Feb 10:00 am 11:00 am | Seminar: Bernardo Subercaseau Title: Foundations of Languages for Interpretability. Abstract: The area of interpretability in Machine Learning aims for the design of algorithms that we humans can understand and trust. One of the fundamental questions of interpretability is: given a classifier M, and an input vector x, why did M classify x as M(x)? In order to approximate an answer to this "why" question, many concrete queries, metrics and scores have emerged as proxies, and their complexity has been studied over different classes of models. Many of these analyses are ad-hoc, but they tend to agree on the fact that these queries and scores are hard to compute over Neural Networks, but easy to compute over Decision Trees. It is thus natural to think of a more general approach, like a query language in which users could write an arbitrary number of different queries, and that would allow for a generalized study of the complexity of interpreting different ML models. Our work proposes foundations for such a language, tying to First Order Logic, as a way to have a clear understanding of its expressiveness and complexity. We manage to define a minimalistic structure over FO that allows expressing many natural interpretability queries over models, and we show that evaluating such queries can be done efficiently for Decision Trees, in data-complexity. Zoom link: univ-lille-fr.zoom.us/j/95419000064 |
Fri 12th Feb 10:00 am 12:00 pm | Seminar: Florent Capelli Title: Regularizing the delay of enumeration algorithms Zoom link: univ-lille-fr.zoom.us/j/95419000064 Abstract: Enumeration algorithms are algorithms whose goal is to output the set of all solutions to a given problem. There exists different measures for the quality of such algorithm, whose relevance depends on what the user wants to do with the solutions set. If the goal of the user is to explore some solutions or to transform the solutions as they are outputted with a stream-like algorithm, a relevant measure of the complexity of an enumeration algorithm is the delay between the output of two distinct solutions. Following this line of thoughts, significant efforts have been made by the community to design polynomial delay algorithms, that is, algorithms whose delay between the output of two new solutions is polynomial in the size of the input. While this measure is interesting, it is not always completely necessary to have a bound on the delay and it is enough to ask for a guarantee that running the algorithm for O(t poly(n)) will result in the output of at least t solutions. Of course, by storing each solution seen and outputting them regularly, one can simulate a polynomial delay but if the number of solutions is large, it may result in a blow up in the space used by the enumerator. In this talk, we will present a new technique that allow to transform such algorithm into polynomial delay algorithm using polynomial space. This is joint work with Yann Strozecki. |
