Model-Centric Data Manifold: The Data Through the Eyes of the Model

We show that deep ReLU neural network classifiers can see a low-dimensional Riemannian manifold structure on data. Such comes via the \sl local data matrix, variation of Fisher information where role model parameters is taken by variables. obtain foliation domain, and we dataset which trained lies leaf, whose dimension bounded number classification labels. validate our results with some experiments MNIST dataset: paths leaf connect valid images, while other leaves cover noisy images.