Optimizacija sažimajućeg očitavanja neuronskim mrežama

master's thesis

• Title (english): Optimization of Compressive Sensing Using Neural Networks
• Author: Leonarda Gajdić
• Mentor: Ana Sović Kržić (mentor)
• Committee member: Ana Sović Kržić (predsjednik povjerenstva)
• Committee member: Damir Seršić (član povjerenstva)
• Committee member: Davor Petrinović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-13, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering

Abstract

Sažimajuće očitavanje vrlo uspješno rekonstruira signal iz smanjenog broja mjerenja. Rekonstrukcija se najčešće provodi uz pretpostavku rijetkosti u nekoj domeni, izvođenjem optimizacijskog postupka.
Kombiniranjem sažimajućeg očitavanja i neuronskih mreža mogu se postići još bolji rezultati, koji se ne oslanjaju isključivo na svojstvo rijetkosti. Vrlo efikasne u tom smislu pokazuju se generativne suparničke mreže (eng. Generative Adversarial Networks, GAN).
Generativna mreža proizvodi bolje rješenje na način da poboljšava performanse i brzinu rekonstrukcije signala. Koristeći pretpostavke korištene kod metode sažimajućeg očitavanja razvija se novi način optimizacije GAN mreža koristeći gradijentne informacije diskriminatora.

Abstract (english)

The compression sensing very successfully reconstructs the signal from the reduced number of measurements. Reconstruction is most often carried out under the assumption of rarity in some of the domain, by performing an optimization procedure.
Combining compression sensing and neural networks, one can achieve even better results, which do not rely solely on the rarity property. Generative Adversarial Networks (GAN) are proving to be very effective in this regard.
The generative network produces a better solution in a way that it improves the performance and speed of signal reconstruction. Using the assumptions used in the compression sensing method, a new way of optimizing GAN networks using discriminant gradient information is developed.

• Keywords: sažimajuće očitavanje; rijetkost; optimizacijski postupci; minimizacija; neuronske mreže; generativne suparničke mreže; diskriminator; gradijentni spust; latentne reprezentacije
• Keywords (english): compressive sensing; rarity; optimization procedure; minimization; neural networks; generative adversarial networks; discriminator; gradient descent; latent representations
• Language: croatian
• URN:NBN: urn:nbn:hr:168:652326
• Study programme: Title: Electrical Engineering and Information Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije (magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 12:16:44