Institutional tag:

In this talk, I will present the main ideas underlying the PAC-Bayesian learning theory - which provides statistical guarantees on the expected loss of an averaging/aggregation/ensemble of multiple predictors - using a simplified approach. This approach leads to a general theorem that embraces several existing PAC-Bayesian results, and eases the “customization” of PAC-Bayesian theorems. In particular, I will show how this can be used to express generalization bounds and design new learning algorithms for semi-supervised learning and similar frameworks, like transductive learning and domain adaptation.

Dates:

Tuesday, January 24, 2017 - 14:00 to 16:00

Location:

Inria Lille - Nord Europe, bâtiment A, salle plénière

Speaker(s):

Pascal Germain

Affiliation(s):

Inria

Speaker's URL: