Date : May 2, 2024, 2:30 p.m. - Room :Amphi 1 - Pôle commun

Modern Algorithmic for Modern Data Efficiency

Vincent COHEN-ADDAD, Researcher - Google

Machine learning and Data Mining have advanced at a breakneck pace, achieving remarkable
performance in a large variety of tasks.
However, these cutting-edge models introduce novel algorithmic and computational questions.

One key challenge is to create robust models from small or biased datasets. I'll illustrate how
new sampling mechanisms with theoretical guarantees can lead to new data selection methods
for speeding up the training of foundation models.

Another key challenge is to understand the success of practical heuristics, highlighting the
algorithmic mechanisms that make them successful to further improve them: making them more
scalable, or succeed under some other constraints. I'll illustrate this with some recent result on the success
of the so called Louvain graph clustering algorithm.