Généralisations de la théorie PAC-bayésienne pour l'apprentissage inductif, l'apprentissage transductif et l'adaptation de domaine
|Advisor:||Laviolette, François; Marchand, Mario|
|Abstract:||In machine learning, the PAC-Bayesian approach provides statistical guarantees on the risk of a weighted majority vote of many classifiers (named voters). The “classical” PAC-Bayesian theory, initiated by McAllester (1999), studies the inductive learning framework under the assumption that the learning examples are independently generated and are identically distributed (i.i.d.) according to an unknown but fixed probability distribution. The thesis contributions are divided in two major parts. First, we present an analysis of majority votes based on the study of the margin as a random variable. It follows a new conceptualization of the PAC-Bayesian theory. Our very general approach allows us to recover several existing results for the inductive PAC-Bayesian framework, and link them in a whole. Among other things, we highlight the notion of expected disagreement between the voters. Building upon an improved understanding of the PAC-Bayesian theory, gained by studying the inductive framework, we then extend it to two other learning frameworks. On the one hand, we study the transductive framework, where the learning algorithm knows the description of the examples to be classified. In this context, we state risk bounds on majority votes that improve those from the current literature. On the other hand, we study the domain adaptation framework, where the generating distribution of the labelled learning examples differs from the generating distribution of the examples to be classified. Our theoretical analysis is the first PAC-Bayesian approach of this learning framework, and allows us to conceive a new machine learning algorithm for domain adaptation. Our empirical experiments show that our algorithm is competitive with other state-of-the-art algorithms.|
|Document Type:||Thèse de doctorat|
|Open Access Date:||23 April 2018|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.