User: Guest  Login
Less Searchfields
Simple search
If you experience problems opening the document, please try this link.
Original title:
Hardware Aware Robust Compression of Neural Networks
Translated title:
Hardware-bewusste robuste Komprimierung neuronaler Netze
Author:
Vemparala, Manoj Rohit
Year:
2023
Document type:
Dissertation
Faculty/School:
TUM School of Computation, Information and Technology
Advisor:
Stechele, Walter (Prof. Dr.)
Referee:
Stechele, Walter (Prof. Dr.); Martina, Maurizio (Prof., Ph.D.)
Language:
en
Subject group:
ELT Elektrotechnik
Keywords:
pruning,quantization,neural networks,compression,compute complexity
TUM classification:
DAT 200
Abstract:
Neural Networks achieve improved performance on various applications. The compute complexity of these algorithms is reduced using pruning and quantization. We use execution estimates to obtain efficient compression configurations. We simplify the compression phase by proposing in-train optimization method. We further improve the adversarial robustness of compressed networks. This work proposes compression techniques, which improve the trade-off between accuracy, robustness and hardware metrics.
Translated abstract:
Neuronale Netze erzielen bei verschiedenen Anwendungen eine bessere Leistung. Die Komplexität dieser Algorithmen wird durch Pruning und Quantisierung reduziert. Wir verwenden Schätzungen zur Ausführung, um effiziente Kompressionskonfigurationen zu erhalten. Wir vereinfachen die Kompressionsphase, indem wir eine In-Train-Methode entwickeln. Wir verbessern die Robustheit der komprimierten Netze. Diese Arbeit verbessert den Kompromiss zwischen Genauigkeit, Robustheit und Hardwaremetriken.
WWW:
https://mediatum.ub.tum.de/?id=1662981
Date of submission:
12.07.2022
Oral examination:
18.01.2023
File size:
19355209 bytes
Pages:
162
Urn (citeable URL):
https://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:91-diss-20230118-1662981-1-6
Last change:
30.03.2023
 BibTeX