Fast and Sample-Efficient Relevance-Based Multi-Task Representation Learning

Jiabin Lin, Shana Moothedath

Abstract: This letter explores an approach for task-relevant multi-task representation learning when the amount of data is limited for both source tasks and target tasks. Specifically, we consider a low-dimensional setting where the goal is to sample source task data based on their relevance so as to utilize task-relevant information effectively. We present a novel learning algorithm based on an alternating projected gradient descent (GD) and minimization estimator. We present the convergence guarantee of our algorithm, excess risk, and the sample complexity of our approach. We evaluated the effectiveness of our algorithm via numerical experiments and compared it empirically against three benchmark approaches.