Federativno učenje z nevronskimi mrežami : magistrsko delo

Čugalj, Jaka

| | SLO | ENG | Cookies and privacy

Bigger font | Smaller font

First page > Show document

Show document

Title:	Federativno učenje z nevronskimi mrežami : magistrsko delo
Authors:	ID Čugalj, Jaka (Author) ID Strnad, Damjan (Mentor) More about this mentor... ID Žalik, Mitja (Comentor)
Files:	MAG_Cugalj_Jaka_2024.pdf (6,46 MB) MD5: A5200269D2B0B4D5A21C689B601E4B07
Language:	Slovenian
Work type:	Master's thesis/paper
Typology:	2.09 - Master's Thesis
Organization:	FERI - Faculty of Electrical Engineering and Computer Science
Abstract:	V magistrskem delu raziščemo postopek učenja nevronskih mrež in predstavimo idejo federativnega učenja, ki omogoči sodelovanje več naprav pri učenju enega modela nevronske mreže brez izmenjave učnih primerov. Glavna prednost federativnega učenja je, da naprava svojih lokalnih podatkov ne deli z ostalimi napravami, zato ostanejo zasebni. Preučili smo algoritme federativnega učenja FedSGD, FedAvg, FedProx, SCAFFOLD, FedVARP, ClusterFedVARP in FedRolex, ki na različne načine rešujejo izzive takšnega načina učenja. Prav tako predstavimo novo rešitev, ki komplementarno združi nekatere naštete algoritme tako, da se lahko v nekaterih primerih ob istih pogojih učenje izvaja učinkoviteje. Učinkovitost učenja smo testirali na klasifikacijskem problemu razpoznave ročno napisanih števil podatkovne zbirke MNIST ter problemu napovedovanja naslednje črke v stavku, kjer smo učne primere generirali s pomočjo literarnih del Williama Shakespeara. Izvedli smo analizo vpliva različnih parametrov algoritmov na učenje nevronskih mrež in primerjali vpliv neenakomerne porazdelitve podatkov na hitrost konvergence posameznih algoritmov na različnih podatkovnih zbirkah. Implementirali smo simulator federativnega učenja z uporabniškim vmesnikom, preko katerega lahko uporabnik ureja parametre učnih algoritmov in odjemalcev ter izvaja učenje in testiranje različnih modelov v ločenih nitih.
Keywords:	nevronske mreže, federativno učenje, klasifikacija, MNIST, stohastični gradientni spust
Place of publishing:	Maribor
Place of performance:	Maribor
Publisher:	[J. Čugalj]
Year of publishing:	2024
Number of pages:	1 spletni vir (1 datoteka PDF (XI, 90 f.))
PID:	20.500.12556/DKUM-88491
UDC:	004.85:004.032.26(043.2)
COBISS.SI-ID:	205879811
Publication date in DKUM:	01.07.2024
Views:	349
Downloads:	98
Metadata:
Categories:	KTFMB - FERI
:	Copy citation

Average score:	(0 votes)
Your score:	Voting is allowed only for logged in users.
Share:

Hover the mouse pointer over a document title to show the abstract or click on the title to get all document metadata.

Licences

License:	CC BY-NC-ND 4.0, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International

Link:	http://creativecommons.org/licenses/by-nc-nd/4.0/
Description:	The most restrictive Creative Commons license. This only allows people to download and share the work for no commercial gain and for no other purposes.
Licensing start date:	06.05.2024

Secondary language

Language:	English
Title:	Federated learning with neural networks
Abstract:	In this master thesis we explore the process of training neural networks and present the concept of federated learning. Federated learning allows multiple edge devices to collaborate on training a single neural network without the need to exchange training examples. Each device retains its own data, keeping it secure and private. We examine several federated learning algorithms, including FedSGD, FedAvg, FedProx, SCAFFOLD, FedVARP, ClusterFedVARP and FedRolex, each addressing different challenges of this learning approach. Additionally, we present a new solution that combines some of these algorithms in a complementary manner, improving learning efficiency under certain conditions. To assess the effectiveness of this learning approach, we conducted experiments on two tasks: classifying handwritten digits from the MNIST dataset and predicting the next character in a sentence, for which training examples were generated using the literary works of William Shakespeare. We analyzed how different algorithm parameters affect neural network learning and compared how uneven data distribution impacts the convergence speed of individual algorithms. We implemented a federated learning simulator with a graphical user interface, through which the user can adjust the parameters of learning algorithms and clients, and conduct training and testing of various models in separate threads.
Keywords:	neural networks, federated learning, classification, MNIST, stochastic gradient descent

Comments

Leave comment

You must log in to leave a comment.

Comments (0)

0 - 0 / 0

There are no comments!

Back