Abstract: The rise in deep neural networks (DNNs) has led to increased interest in explaining their predictions. While many methods for this exist, there is currently no consensus on how to evaluate ...

Abstract: Deep neural networks often suffer from poor performance or even training failure due to the ill-conditioned problem, the vanishing/exploding gradient problem, and the saddle point problem.