eduzhai > Applied Sciences > Engineering >

CognitiveCNN Mimicking Human Cognitive Models to resolve Texture-Shape Bias

  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: Recent works demonstrate the texture bias in Convolutional Neural Networks(CNNs), conflicting with early works claiming that networks identify objectsusing shape. It is commonly believed that the cost function forces the networkto take a greedy route to increase accuracy using texture, failing to exploreany global statistics. We propose a novel intuitive architecture, namelyCognitiveCNN, inspired from feature integration theory in psychology to utilisehuman-interpretable feature like shape, texture, edges etc. to reconstruct, andclassify the image. We define two metrics, namely TIC and RIC to quantify theimportance of each stream using attention maps. We introduce a regulariserwhich ensures that the contribution of each feature is same for any task, as itis for reconstruction; and perform experiments to show the resulting boost inaccuracy and robustness besides imparting explainability. Lastly, we adaptthese ideas to conventional CNNs and propose Augmented Cognitive CNN to achievesuperior performance in object recognition.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×