published by Alexander Usvyatsov
in 2019
in Informatics Engineering
and research's language is
English
Download
Abstract in English
We consider functions defined by deep neural networks as definable objects in an o-miminal expansion of the real field, and derive an almost linear (in the number of weights) bound on sample complexity of such networks.