Change search
ReferencesLink to record
Permanent link

Direct link
Modularity as a Solution to Spatial Interference in Neural Networks
Norwegian University of Science and Technology, Faculty of Information Technology, Mathematics and Electrical Engineering, Department of Computer and Information Science.
2012 (English)MasteroppgaveStudent thesis
Abstract [en]

Modularity is an architectural trait that is prominent in biological neural networks, but strangely absent in evolved artificial neural networks. This report contains the results of a theoretical study focusing on two questions about modularity in neural network systems. How does modularity emerge in biological neural networks, and when could modularity be useful in artificial neural networks? The theoretical study resulted in a hypothesis that modularity in biological neural networks is the result of physical constraints on their architectures. Because these physical constraints affect the digital environments in a different way, modularity does not emerge naturally during evolution of neural networks in a digital medium. Secondly, it is hypothesised that modularity in artificial neural networks can reduce the amount of spatial interference during learning. A phenomenon that is here shown to occur when two outputs that exhibit low correlation are solved using the same neural network structures. Experiments have been performed in order to verify if there are advantages to having modular topologies in order to limit the detrimental effects of spatial interference occurring during learning in artificial neural networks.

Place, publisher, year, edition, pages
Institutt for datateknikk og informasjonsvitenskap , 2012. , 110 p.
Keyword [no]
ntnudaim:5999, MIT informatikk, Kunstig intelligens og læring
URN: urn:nbn:no:ntnu:diva-18510Local ID: ntnudaim:5999OAI: diva2:565996
Available from: 2012-11-08 Created: 2012-11-08

Open Access in DiVA

fulltext(1459 kB)221 downloads
File information
File name FULLTEXT01.pdfFile size 1459 kBChecksum SHA-512
Type fulltextMimetype application/pdf
cover(220 kB)17 downloads
File information
File name COVER01.pdfFile size 220 kBChecksum SHA-512
Type coverMimetype application/pdf
attachment(21406 kB)10 downloads
File information
File name ATTACHMENT01.zipFile size 21406 kBChecksum SHA-512
Type attachmentMimetype application/zip

By organisation
Department of Computer and Information Science

Search outside of DiVA

GoogleGoogle Scholar
Total: 221 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

Total: 66 hits
ReferencesLink to record
Permanent link

Direct link