"Inverse stability of the realization map for ReLU networks"Elbrächter, DennisThe realization map maps the parameters of a neural network (considered as a vector in R^n) to its realization, i.e. the function it computes. We consider how badly this map fails to be injective and present an approach of how to fix this, i.e. achieve some sort of inverse stability. In addition, we will discuss the implications of such an inverse stability property for the optimization problem on which neural network training is usually based. |
| https://arxiv.org/abs/1905.09803 |
« back