CERAMATHS - DMATHS Seminar: presentation by Faicel Chamroukhi

The weekly seminar of the mathematics department will host, on Thursday, December 8, Faicel Chamroukhi (University of Caen and IRT SystemX), for the following talk:

On a new family of mixture-of-experts models for learning from heterogeneous data at scale
. On a new family of mixture-of-experts models for learning from heterogeneous data at scale

Abstract: Efficient machine learning of knowledge from massive data is strongly mobilizing academic and industrial actors, in particular in the framework of the national strategy in artificial intelligence. This question poses fundamental challenges related to the design of adapted models and the study of their approximation and inference capacities, and is at the heart of the mathematical sciences and their interactions with the digital domain. It also poses applied challenges regarding the efficient implementation of the algorithms that are derived from it and their appropriation by industry. In this seminar, I will present a family of expert mixture models adapted to data that are both heterogeneous, high-dimensional and potentially massively available, and the associated estimation and model selection tools. In particular, I will consider the issue of high dimensionality, especially when these models are built from a functional covariate for prediction, and for the sake of parsimony and interpretability, as well as the issue of clustering at scale when the data are distributed, for operational purposes.