Komprimace dat s využitím hlubokého učení

Abstract

The thesis deals with the problem of data compression using artificial neural networks. Within the thesis, I describe the use of artificial neural networks to calculate conditional probabilities of individual bytes in the file we want to compress. Based on these probabilities, the data can be compressed using an arithmetic coding algorithm. These probabilities can also be used to estimate the resulting compressed file size in advance using information entropy. The thesis also describes the work with the Tensorflow library, Keras and the jupyter lab environment. In several different series of experiments, it was found that data compression using neural network prediction gives us interesting results for different neural network architectures. The paper also includes the input data processing procedure, the different neural network models and the results of experiments.