Structured Methods for Multi-Task Learning
Developing algorithms and open-source software for learning from related tasks simultaneously, addressing problems beyond the reach of standard multi-task learning.
Principal Investigators
Jiayu Zhou — Leading PI, Michigan State University
Shuiwang Ji — PI, Washington State University
Project Overview
The human ability to learn from and transfer knowledge across related tasks enables us to grasp complex concepts from only a few examples. A three-year-old child can discriminate chairs from tables without exposure to hundreds of different examples. In contrast, computer learning programs typically require large volumes of labeled data for each new task.
This NSF CRII collaborative project focuses on structured methods for multi-task learning — algorithms that learn from multiple related tasks simultaneously, transferring knowledge between them to improve performance on each. The research addresses problem structures not traditionally handled by standard multi-task learning frameworks.
A key outcome of this work is the development of open-source software tools that make these structured multi-task learning methods accessible to the broader research community.
Software & Resources
The project has produced the MALSAR (Multi-task Learning via Structural Regularization) toolkit, an open-source package for multi-task learning. Visit MALSAR.org for documentation, downloads, and examples.
Related Links
Contact
For inquiries about this project, contact Jiayu Zhou at jiayuz@0xmachine.com.