CONSTRUCTION OF NEURAL ARCHITECTURES WITH DESIRED BEHAVIOUR UNDER GEOMETRIC TRANSFORMATIONS OF THE INPUT

Abstract

We present a general method for analysis of convolutional layers under geometric transformations of the input that are linear with respect to pixel values. We also describe the algorithm for finding all possible types of behaviours of the output of convolutional layers under geometric transformations of the input. We also present a general method for construction of convolutional architectures with desired behaviour under geometric transformations of the input.