Task Uncertainty Loss Reduce Negative Transfer in Asymmetric Multi-task\n Feature Learning

Multi-task learning (MTL) is frequently used in settings where a target task\nhas to be learnt based on limited training data, but knowledge can be leveraged\nfrom related auxiliary tasks. While MTL can improve task performance overall\nrelative to single-task learning (STL), these improvements can hide negative\ntransfer (NT), where STL may deliver better performance for many individual\ntasks. Asymmetric multitask feature learning (AMTFL) is an approach that tries\nto address this by allowing tasks with higher loss values to have smaller\ninfluence on feature representations for learning other tasks. Task loss values\ndo not necessarily indicate reliability of models for a specific task. We\npresent examples of NT in two orthogonal datasets (image recognition and\npharmacogenomics) and tackle this challenge by using aleatoric homoscedastic\nuncertainty to capture the relative confidence between tasks, and set weights\nfor task loss. Our results show that this approach reduces NT providing a new\napproach to enable robust MTL.\n