Compressing network data with deep learning

This thesis delves into the problem of compressing data generated by mobile networks, which are formed by many antennas and IoT devices, that emit huge quantities of measurements. Storing this data is growing more difficult and expensive, a trend driven by continuously escalating data transmission s...

Full description

Bibliographic Details
Author: Gili Fernández De Romarategui, David
Format: master thesis
Publication Date:2024
Country:España
Institution:Universitat Politècnica de Catalunya (UPC)
Repository:UPCommons. Portal del coneixement obert de la UPC
Language:English
OAI Identifier:oai:upcommons.upc.edu:2117/406468
Online Access:https://hdl.handle.net/2117/406468
Access Level:Open access
Keyword:Deep learning
Data compression (Computer science)
compressió amb aprenentatge profund
codificació entròpica
transformadors
dades de xarxes mòbils
deep learning compression
entropy coding
transformers
mobile network data
Aprenentatge profund
Dades -- Compressió (Informàtica)
Àrees temàtiques de la UPC::Informàtica::Arquitectura de computadors
Description
Summary:This thesis delves into the problem of compressing data generated by mobile networks, which are formed by many antennas and IoT devices, that emit huge quantities of measurements. Storing this data is growing more difficult and expensive, a trend driven by continuously escalating data transmission speeds, the sheer volume of traffic, and the increasing number of devices. To solve this problem, we surveyed some compression algorithms, both classical and learning-based, with a particular focus on deep-learning predictors applied in conjunction with entropy encoders. The core of our research is centered around compression algorithms based on Transformers and Arithmetic encoders. We propose several improvements like pretraining and using different models for each column, which outperform current approaches for our particular data domain. The focus of this research will be trying to improve compression on a large-scale dataset provided by Telefonica, which contains heterogeneous measurements from a huge network of antennas. Additionally, we provide a comprehensive analysis of the outcomes obtained, giving an intuition about why deep learning approaches work better than the current baselines, albeit at a higher computational cost, and explaining why our domain-specific solution slightly outperforms the rest.