TASK GROUPING FOR BETTER MULTI-TASK LEARNING

Summary from the TLDR AI newsletter

A new approach to multi-task learning mitigates negative transfer by dynamically grouping tasks and updating them sequentially during training. The method, based on proximal inter-task affinity, significantly improves performance over existing multi-task optimization techniques.