Pascal Germain, Researcher in Machine Learning

At first, this master thesis presents a general PAC-Bayes theorem, from which we can easily obtain some well-known PAC-Bayes bounds. Those bounds allow us to compute a guarantee on the risk of a classifier from its achievements on the training set. We analyze the behavior of two PAC-Bayes bounds and we determine peculiar characteristics of classifiers favoured by those bounds. Then, we present a specialization of those bounds to the linear classifiers family. Secondly, we conceive three new machine learning algorithms based on the minimization, by conjugate gradient descent, of various mathematical expressions of the PAC-Bayes bounds. The last algorithm uses a part of the training set to capture a priori knowledges. One can use those algorithms to construct majority vote classifiers as well as linear classifiers implicitly represented by the kernel trick. Finally, an elaborated empirical study compares the three algorithms and shows that some versions of those algorithms are competitive with both AdaBoost and SVM.