Allocating different neural network weights to different tasks or modalities to prevent conflicting learning signals.