eduzhai > Applied Sciences > Engineering >

Seeing eye-to-eye? A comparison of object recognition performance in humans and deep convolutional neural networks under image manipulation

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 19 pages

Abstract: For a considerable time, deep convolutional neural networks (DCNNs) havereached human benchmark performance in object recognition. On that account,computational neuroscience and the field of machine learning have started toattribute numerous similarities and differences to artificial and biologicalvision. This study aims towards a behavioral comparison of visual core objectrecognition performance between humans and feedforward neural networks in aclassification learning paradigm on an ImageNet data set. For this purpose,human participants (n = 65) competed in an online experiment against differentfeedforward DCNNs. The designed approach based on a typical learning process ofseven different monkey categories included a training and validation phase withnatural examples, as well as a testing phase with novel, unexperienced shapeand color manipulations. Analyses of accuracy revealed that humans not onlyoutperform DCNNs on all conditions, but also display significantly greaterrobustness towards shape and most notably color alterations. Furthermore, aprecise examination of behavioral patterns highlights these findings byrevealing independent classification errors between the groups. The obtainedresults show that humans contrast strongly with artificial feedforwardarchitectures when it comes to visual core object recognition of manipulatedimages. In general, these findings are in line with a growing body ofliterature, that hints towards recurrence as a crucial factor for adequategeneralization abilities.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×