Exploiting learned saliency to optimize machine learning training

Abstract

Embodiments herein describe techniques for setting and using saliency values to modify how a ML model is trained. In one embodiment, different blocks of data (referred to herein as tiles) are assigned respective saliency values. After performing one or more iterations, a training application can modify the default saliency values assigned to the tiles to reflect the importance of the tile. In one embodiment, the training application evaluates a weight gradient that indicates how a weight (or weights) corresponding to each tile are modified and modifies the saliency values accordingly. The ML training system can then use the saliency values to affect future training iterations to reduce the time required to train the ML model or save power.