The Effect of Shape Selective Kernels in Convolutional Neural Networks

Vries, Stijn de (2024) The Effect of Shape Selective Kernels in Convolutional Neural Networks. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI2024deVriesS.pdf Download (6MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (463kB)

Abstract

State of the art Convolutional Neural Networks (CNNs) can outperform humans in classification tasks, but are still not as robust and reliable as humans. Moreover, recent research suggests that CNNs have a strong texture bias, whereas humans develop a strong shape bias for object recognition. If CNNs were made to process visual data more like humans, an improvement in robustness and reliability of the visual models might be observed. The goal of this research is firstly to develop kernels that respond strongly to shapes in images, using biological visual cells for inspiration. Secondly, the effect that the shape selective kernels have on the performance of CNNs is investigated. Shape selective kernels were employed in CNNs which were trained in three different settings. The performances of the models were measured and the models were analyzed to reveal how the data was processed. Results show small improvements in some models, but a large decrease in both clean accuracy and robustness for most models using shape selective kernels. However, improvements in model certainty can be observed, and further analysis indicates that models with shape selective kernels process data differently than conventional CNNs.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Valdenegro Toro, M.A. and Azzopardi, G. and Bennabhaktula, G.S.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	19 Jul 2024 13:34
Last Modified:	19 Jul 2024 13:34
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/33558

Actions (login required)

View Item