Multi-task Learning by Leveraging the Semantic Information

One crucial objective of multi-task learning is to align distributions across tasks so that the information between them can be transferred and shared. However, existing approaches only focused on matching marginal feature distribution while ignoring semantic information, which may hinder performance. To address this issue, we propose leverage label in by exploring conditional relations among tasks. We first theoretically analyze generalization bound based notion Jensen-Shannon divergence, provides new insights into value learning. Our analysis also leads a concrete algorithm jointly matches controls divergence. confirm effectiveness proposed method, compare with several baselines some benchmarks then test algorithms under space shift conditions. Empirical results demonstrate method could outperform most achieve state-of-the-art performance, particularly showing benefits