A Model For the Improved Representation of High Frequency Signals in Convolutional Neural Network Architectures

Abstract

This report details efforts to develop a novel framework to encode and generate data from a set of or single training example(s) such that the encoded or generated examples faithfully recreate the spatial frequency-domain information contained in the training data. Whereas there are several methods to encode data with neural frameworks, and several other, typically GAN-based, methods for generating novel examples of data based on a training set, these methods often exhibit poor performance in the representation of high-frequency components in the data This effect is particularly visible at early stages of training. Firstly, this report details the development of two novel architectures for the generation of novel examples data based on a training dataset. Both new architectures incorporate innovations which improve the generation of high-frequency components of the generated signals.

Secondly, this report applies one of these new architectures to a real-world problem. In this example, the framework is applied to generate a temporal progression of Vincent Van Gogh's painting `The Starry Night'. Considering the painting as a realistic turbulent fluid flow system, the painting is progressed in time in a realistic manner. This work of art is considered as a turbulent flow system since it has been shown to exhibit properties which mean that it can reasonably considered as a turbulent flow field. The methods introduced in the first part of this report mean that the high-frequency components of the temporally-progressed painting are accurately represented.