Negative transfer

RotoGrad: Gradient Homogenization in Multitask Learning

We propose an algorithm to simultaneously homogenize gradient magnitudes and directions across tasks in multitask learning.

Adrián Javaloy, Isabel Valera