Adaptive and Robust Multi-Task Learning

TL;DR

This paper introduces adaptive multi-task learning methods that leverage task similarities, provide statistical guarantees, and demonstrate robustness and effectiveness through experiments on synthetic and real data.

Contribution

It proposes a new family of adaptive methods for multi-task learning that handle task differences and outliers with proven statistical guarantees.

Findings

Methods outperform baseline models on synthetic datasets.

Robustness against outlier tasks demonstrated.

Effective on real-world datasets.

Abstract

We study the multi-task learning problem that aims to simultaneously analyze multiple datasets collected from different sources and learn one model for each of them. We propose a family of adaptive methods that automatically utilize possible similarities among those tasks while carefully handling their differences. We derive sharp statistical guarantees for the methods and prove their robustness against outlier tasks. Numerical experiments on synthetic and real datasets demonstrate the efficacy of our new methods.