%0 Journal Article %1 xu2017informationtheoretic %A Xu, Aolin %A Raginsky, Maxim %D 2017 %K bounds generalization information theory %T Information-theoretic analysis of generalization capability of learning algorithms %U http://arxiv.org/abs/1705.07809 %X We derive upper bounds on the generalization error of a learning algorithm in terms of the mutual information between its input and output. The bounds provide an information-theoretic understanding of generalization in learning problems, and give theoretical guidelines for striking the right balance between data fit and generalization by controlling the input-output mutual information. We propose a number of methods for this purpose, among which are algorithms that regularize the ERM algorithm with relative entropy or with random noise. Our work extends and leads to nontrivial improvements on the recent results of Russo and Zou.

@article{xu2017informationtheoretic, abstract = {We derive upper bounds on the generalization error of a learning algorithm in terms of the mutual information between its input and output. The bounds provide an information-theoretic understanding of generalization in learning problems, and give theoretical guidelines for striking the right balance between data fit and generalization by controlling the input-output mutual information. We propose a number of methods for this purpose, among which are algorithms that regularize the ERM algorithm with relative entropy or with random noise. Our work extends and leads to nontrivial improvements on the recent results of Russo and Zou.}, added-at = {2019-09-19T13:38:51.000+0200}, author = {Xu, Aolin and Raginsky, Maxim}, biburl = {https://www.bibsonomy.org/bibtex/2586e90e3ad5f07e0f485cbd56a17d182/kirk86}, description = {[1705.07809] Information-theoretic analysis of generalization capability of learning algorithms}, interhash = {e3c98e9c1ea2e29fd7af1990be99c63f}, intrahash = {586e90e3ad5f07e0f485cbd56a17d182}, keywords = {bounds generalization information theory}, note = {cite arxiv:1705.07809Comment: Final version, accepted to NIPS 2017}, timestamp = {2019-09-19T13:38:51.000+0200}, title = {Information-theoretic analysis of generalization capability of learning algorithms}, url = {http://arxiv.org/abs/1705.07809}, year = 2017 }