Analiza učinkovitosti učenja s prenosom znanja pri detekciji objektov : magistrsko delo

Žalik, Mitja

| | SLO | ENG | Cookies and privacy

Bigger font | Smaller font

First page > Show document

Show document

Title:	Analiza učinkovitosti učenja s prenosom znanja pri detekciji objektov : magistrsko delo
Authors:	ID Žalik, Mitja (Author) ID Mongus, Domen (Mentor) More about this mentor... ID Strnad, Damjan (Comentor)
Files:	MAG_Zalik_Mitja_2022.pdf (16,36 MB) MD5: 00E04D030FC4EF33823039378BFA7329 PID: 20.500.12556/dkum/298a13eb-b517-485b-9298-ee42ac40bd49
Language:	Slovenian
Work type:	Master's thesis/paper
Typology:	2.09 - Master's Thesis
Organization:	FERI - Faculty of Electrical Engineering and Computer Science
Abstract:	Zaradi nedefiniranosti procesov odločanja globokih nevronskih mrež in njihovega dolgotrajnega učenja predstavlja določitev prenesenega znanja ključen izziv pri implementaciji učinkovite detekcije objektov na novih domenah. Preneseno znanje opredeljuje struktura plasti nevronske mreže, nad katerimi izgradimo nov model, ter izbira plasti, ki jim med učenjem zamrznemo vrednosti uteži. V magistrskem delu analiziramo vpliv števila zamrznjenih plasti na uspešnost učenja s prenosom znanja. V prvem delu opišemo tehnike prenosa znanja ter podamo formalno definicijo detekcije objektov, pri čemer opredelimo poznane metode in izpostavimo ključne izzive, povezane z njimi. Nato predstavimo izveden eksperiment, v katerem primerjamo uspešnost štirih konfiguracij pri prenosu znanja na modelu YOLOv4 na štiri različne ciljne domene. Ugotovimo, da so pri različnih ciljnih domenah uspešne različne konfiguracije, ki so odvisne od stopnje podobnosti izvorne in ciljne domene ter plasti izvornega modela, na kateri je določena značilka izluščena. Čeprav predstavljeni rezultati kažejo nemožnost predvidevanja optimalne konfiguracije prenosa znanja, izveden eksperiment nakazuje, da je učenje tudi v primeru neoptimalnega prenosa znanja uspešnejše od učenja brez prenosa znanja.
Keywords:	učenje s prenosom znanja, prenos znanja, detekcija objektov, obdelava videoposnetkov, globoko učenje
Place of publishing:	Maribor
Place of performance:	Maribor
Publisher:	[M. Žalik]
Year of publishing:	2022
Number of pages:	1 spletni vir (1 datoteka PDF (XVII, 49 f.))
PID:	20.500.12556/DKUM-81794
UDC:	004.853:[004.81:159.953.5](043.2)
COBISS.SI-ID:	113505027
Publication date in DKUM:	07.06.2022
Views:	1223
Downloads:	267
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:	05.06.2022

Secondary language

Language:	English
Title:	Analysis of the efficiency of transfer learning for object detection
Abstract:	Due to the unexplainability of the decision-making processes within deep neural networks and their long training times, determining the suitable knowledge to transfer represents a key challenge in their usage in new domains. The transferred knowledge is determined by the structure of the layers, over which a new model is built, and the choice of the layers with frozen weights during training. In the master's thesis, the influence of the number of frozen layers on the success of learning through knowledge transfer is analyzed. For this purpose, we first examine knowledge transfer techniques in the context of object detection, together with key methods and challenges associated with them. Afterwards, the performed experiment is presented, in which the performance of four transfer learning configurations to four different target domains is compared on the YOLOv4 model. Based on the conducted experiment, we conclude that different configurations are successful for different target domains, depending on the degree of similarity between the source and target domains and the layer of the source model where a particular feature is extracted. Although the presented results show the impossibility of predicting the optimal configuration of knowledge transfer in advance, the performed experiment suggests that transfer learning is more successful than learning without knowledge transfer even in the case of suboptimal transfer learning configuration.
Keywords:	transfer learning, knowledge transfer, object detection, video processing, deep learning

Comments

Leave comment

You must log in to leave a comment.

Comments (0)

0 - 0 / 0

There are no comments!

Back