Seminar Details
2019-11-05 (14:00) : Two 20-minute talks
At Euler building (room A.002)
Organized by Mathematical Engineering
Section 1: Convex optimization using tunable first-order inexact oracles (TFOIO)
Speaker :
Van Dessel, Guillaume
Abstract :
In this talk will be introduced/ motivated the concept of tunable first-order inexact oracle (TFOIO) in convex optimization.
More specifically: It will be first discussed convergence results for
smooth-convex optimization methods such as (P)GD and FGD taking
into account for inexactness both in the information used
(tunable first-order inexact oracle built upon surrogates for
the objective function and its (sub)gradient) and the accuracy
on the update steps. Then it will be shown how and when it is possible to derive, based on a cost (either computational or economical) associated with oracle calls, optimal schedules for the above methods. Worst-case cost gains will be presented for a proper set of chosen experiments.
Finally, if the time allows it, practical examples using a personal
toolbox will be displayed in live to confirm the empirical
behavior of the cited methods under the use of inexact
information.
Section 2: New developments in Machine Learning: a Kolmogorov Complexity approach
Speaker :
Pinon, Brieuc
Abstract :
We first identify some intrinsic limitations in currently used learning algorithms in Machine Learning. These algorithms lack of learning automatization and expressiveness in order to efficiently exploit some structures in some complex tasks.
A solution can be found within the Kolmogorov Complexity framework, "perfect" learning that exploits all computable patterns in the data exists.
However, it is non-computable.
We point to potential practical solutions for this problem, and theoretical questions that we investigate.
An understanding of introductory notions in Machine Learning and Theoretical Computer Science is advised.