Multi Task Learning Package for Matlab


Multi Task Learning Package for Matlab. Code implementing the work in

Ciliberto, Carlo, Tomaso Poggio, and Lorenzo Rosasco. Convex Learning of Multiple Tasks and their Structure. International Conference on Machine Learning (ICML), 2015.

##Installation & Usage

Just addpath('learning-machine') to your MATLAB path. An example of how to use the package is in the file main.m.

##A couple of words on the package

  1. The goal/scope of our work was to present a general convex framework for multi-task learning, which would allow on one hand to capture several previous approaches proposed in Multi-task Learning (e.g. Argyriou '08, Jacob '09, Zhang '10, Dinuzzo '11, etc.) and on the other hand offer a general (meta) block coordinate strategy to solve problems of this form, with guarantees to converge to the global minimum. The code in this repository consists in the implementation of such a meta-strategy for some of these multi-task settings.

  2. The package has been designed with the intention of being be plug-and-play, however it is not yet ready for distribution and unfortunately still not documented. In particular the parameter selection routine is available but there's no documentation at all. TBD!

###References [1] Argyriou, Andreas, Theodoros Evgeniou, and Massimiliano Pontil. "Convex multi-task feature learning". Machine Learning, 2008.

[2] Jacob, Laurent, Jean-philippe Vert, and Francis R. Bach. "Clustered multi-task learning: A convex formulation." Advances in neural information processing systems, 2009.

[3] Zhang, Yu, and Dit-Yan Yeung. "A convex formulation for learning task relationships in multi-task learning." International Conference on Machine Learning (ICML), 2012.

[4] Dinuzzo, Francesco, Cheng S. Ong, Gianluigi Pillonetto, and Peter V. Gehler. "Learning output kernels with block coordinate descent." Proceedings of the 28th International Conference on Machine Learning (ICML), 2011.

Get A Weekly Email With Trending Projects For These Categories
No Spam. Unsubscribe easily at any time.