Advanced Search
DSpace@MIT

On the Robustness of Convolutional Neural Networks to Internal Architecture and Weight Perturbations

Research and Teaching Output of the MIT Community

Show simple item record

dc.contributor.author Cheney, Nicholas
dc.contributor.author Schrimpf, Martin
dc.contributor.author Kreiman, Gabriel
dc.date.accessioned 2017-04-07T15:25:52Z
dc.date.available 2017-04-07T15:25:52Z
dc.date.issued 2017-04-03
dc.identifier.uri http://hdl.handle.net/1721.1/107935
dc.description.abstract Deep convolutional neural networks are generally regarded as robust function approximators. So far, this intuition is based on perturbations to external stimuli such as the images to be classified. Here we explore the robustness of convolutional neural networks to perturbations to the internal weights and architecture of the network itself. We show that convolutional networks are surprisingly robust to a number of internal perturbations in the higher convolutional layers but the bottom convolutional layers are much more fragile. For instance, Alexnet shows less than a 30% decrease in classification performance when randomly removing over 70% of weight connections in the top convolutional or dense layers but performance is almost at chance with the same perturbation in the first convolutional layer. Finally, we suggest further investigations which could continue to inform the robustness of convolutional networks to internal perturbations. en_US
dc.description.sponsorship This work was supported by the Center for Brains, Minds and Machines (CBMM), funded by NSF STC award CCF-1231216. en_US
dc.language.iso en_US en_US
dc.publisher Center for Brains, Minds and Machines (CBMM), arXiv en_US
dc.relation.ispartofseries CBMM Memo Series;065
dc.rights Attribution-NonCommercial-ShareAlike 3.0 United States *
dc.rights.uri http://creativecommons.org/licenses/by-nc-sa/3.0/us/ *
dc.title On the Robustness of Convolutional Neural Networks to Internal Architecture and Weight Perturbations en_US
dc.type Technical Report en_US
dc.type Working Paper en_US
dc.type Other en_US
dc.identifier.citation arXiv:1703.08245 en_US


Files in this item

Name Size Format Description
CBMM-Memo-065.pdf 687.7Kb PDF

The following license files are associated with this item:

This item appears in the following Collection(s)

Show simple item record

Attribution-NonCommercial-ShareAlike 3.0 United States Except where otherwise noted, this item's license is described as Attribution-NonCommercial-ShareAlike 3.0 United States
MIT-Mirage