Why Smooth Stability Assumptions Fail for ReLU Learning

This article likely analyzes the limitations of using smooth stability assumptions in the context of training neural networks with ReLU activation functions. It probably delves into the mathematical reasons why these assumptions, often used in theoretical analysis, don't hold true in practice, potentially leading to inaccurate predictions or instability in the learning process. The focus would be on the specific properties of ReLU and how they violate the smoothness conditions required for the assumptions to be valid.