Learning multiple online tasks with a global objective

Ofer Dekel

The Hebrew University of Jerusalem

Abstract

The simplicity and elegance of online learning make it a practical tool with many useful applications. In practice, we are often faced with multiple online prediction tasks in parallel. The naive approach to dealing with such situations is to learn each task separately and independently of the others. Can we do any better than this?

We present an online multitask learning framework where the multiple tasks all contribute to a common goal and share the consequences of their prediction mistakes. We present several new learning algorithms that take advantage of this framework and learn the multiple tasks jointly. We prove cumulative loss bounds that provide a guarantee on the worst-case performance of our algorithms and elucidate the advantages of our approach over the naive approach of learning each task separately.

Many real-world online prediction applications naturally fit within our framework. We present several concrete examples, including a multiclass-multilabel text categorization algorithm capable of processing millions of examples in minutes, a surprising application of our technique to online ordinal regression, and a randomized algorithm that dynamically allocates memory to multiple kernel-based classifiers.

Joint work with Yoram Singer and Phil Long

Bio:

Ofer Dekel is a Ph.D student at The Hebrew University of Jerusalem, School of Computer Science and Engineering.
Maintained by: Fei Sha