Science from the Computer Vision Zoo
November 19, 2018, 2PM
The tremendous progress in computer vision has led to new unresolved questions about their emergent properties.
Understanding the emergent behaviour of computer vision algorithms can fuel the engineering of computer vision and help understand biological intelligence.
In this talk, I will show three recent contributions that advance our understanding of the generalization capabilities of deep neural networks:
redundancy in the neural activity as a mechanisms that allows the network to perform remarkably well despite the fact that most such networks are vastly overparametrized
; ii) a shared failure mode with humans in which the network's accuracy sharply drops due to small changes of the visible region; and
iii) single units in deep neural networks functionally correspond with neurons in the brain