The Taub Faculty of Computer Science Events and Talks

We present a novel algorithm based on a cascade of regularization terms designed to induce implicit hierarchical sharing of information among related learning tasks. Our approach can be viewed as training and combining a set of diverse classifiers. Such a combination is known to improve accuracy. The diversity is achieved by inducing different levels of sharing among tasks. Our approach is designed for multi-task and multi-class learning scenarios. Enabling different levels of shared information is particularly important in large scale problems such as multi-class classification with many classes. In such scenarios it is assumed that no single grouping of classes can capture all the shared information. We extend our batch approach to an online setting and provide regret analysis of the algorithm. We tested our approach extensively on synthetic and real datasets, showing significant improvement over baseline and state-of-the-art methods.