Mitigating Negative Transfer in Multi-Task Learning with Exponential Moving Average Loss Weighting Strategies (Student Abstract)

Multi-Task Learning (MTL) is a growing subject of interest in deep learning, due to its ability train models more efficiently on multiple tasks compared using group conventional single-task models. However, MTL can be impractical as certain dominate training and hurt performance others, thus making some perform better model multi-task one. Such problems are broadly classified negative transfer, many prior approaches the literature have been made mitigate these issues. One such current approach alleviate transfer weight each losses so that they same scale. Whereas loss balancing rely either optimization or complex numerical analysis, none directly scale based their observed magnitudes. We propose techniques for scaling by exponential moving average benchmark them against best-performing methods three established datasets. On datasets, achieve comparable, if not higher, methods.