Kvantizacija globokih opisnikov za kompresijo v robnem računalništvu

ŠUŠTAR, GREGA (author), Čehovin Zajc, Luka (mentor)

V diplomskem delu obravnavamo idejo kvantizacije globokih opisnikov, ki jih vračajo skriti sloji konvolucijskih nevronskih mrež. Cilj pristopa je zmanjšati število bitov in posledično količino informacij, da bi se te hitreje in bolj učinkovito poslali prek mreže na oblak, kjer bi se procesiranje nadaljevalo. To nam bi omogočilo nadzorovano stiskanje vhodnih podatkov, porazdeljeno procesiranje med dvema sistemoma in morda celo delno anonimizacijo surovih podatkov. V okviru dela smo izvedli eksperiment z modelom za kategorizacijo, ki smo ga učili na dveh standardnih zbirkah slik. Rezultati eksperimenta kažejo, da je tak razcep obdelave slike mogoč. Količina podatkov za prenos kvantiziranih opisnikov je nižja kot v primeru uporabe brezizgubnega slikovnega kodeka. V diskusiji opišemo tudi pomanjkljivosti zasnove našega eksperimenta in podamo predloge za nadaljnje raziskave na tem področju.

• Language: Slovenian
• Keywords: nevronske mreže, globoko učenje, kvantizacija, robno računalništvo
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-140541
• COBISS.SI-ID: 123688963
• Publication date in RUL: 15.09.2022

Secondary language

• Language: English
• Title: Quantization of deep descriptors for compression in edge computing

Abstract

In this thesis, we discuss the idea of quantizing deep descriptors returned by the hidden layers of convolutional neural networks. The aim of this approach is to reduce the number of bits and, consequently the amount of information in order to send this information more quickly and efficiently through the network to the cloud, where the processing would continue. This would allow us to compress the input data in a controlled way, distribute the processing between the two systems and perhaps even to partially anonymise the raw data. In the scope of our thesis, we ran an experiment with a categorisation model trained on two standard image collections. The results of the experiment show that such a split of image processing is possible. The amount of data for transfering the quantized descriptors is lower than in if we were to use a lossless image codec. In the discussion, we also describe the shortcomings of our experimental design and make suggestions for further research in this area.

• Keywords: neural networks, deep learning, quantization, edge computing