The Sample Compression Framework in Machine Learning

Опубликовано 6 сентября 2016, 18:37

This talk will focus on the Sample Compression learning framework emphasizing some of its advantages over more conventional frameworks such as the VC learning paradigm. Moreover, unlike traditional VC and Rademacher based learning paradigms, we will show how practical realizable guarantees on the generalization performance of the learning algorithms can be obtained in this framework. We will also study examples of learning algorithms where such risk bounds can practically guide the model selection process yielding competitive results with the state-of-the-art learning algorithms as well as conventional re-sampling techniques such as the k-fold cross validation. Finally, we will show how some of the well known algorithms such as decision trees and SVM can be characterized within this framework, followed with a discussion on some open questions.