Distribution-specific hardness of learning neural networks

Abstract

Although neural networks are routinely and successfully trained in practice using simple gradient-based methods, most existing theoretical results are negative, showing that learning such networks ...