Tag
This paper introduces SURGE, a novel learnable gradient compensation framework for training Binary Neural Networks that addresses gradient mismatch and information loss issues found in traditional methods like the Straight-Through Estimator.