Technical Program

Paper Detail

Paper: PS-1B.51
Session: Poster Session 1B
Location: H Fl├Ąche 1.OG
Session Time: Saturday, September 14, 16:30 - 19:30
Presentation Time:Saturday, September 14, 16:30 - 19:30
Presentation: Poster
Publication: 2019 Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany
Paper Title: Optimizing a recurrent neural architecture for contour detection produces a tilt illusion
Manuscript:  Click here to view manuscript
License: Creative Commons License
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
DOI: https://doi.org/10.32470/CCN.2019.1244-0
Authors: Drew Linsley, Junkyung Kim, Thomas Serre, Brown University, United States
Abstract: Context influences our perception of visual scenes. While there is a consensus in vision science that processes like contour integration depend on recurrent contextual processing, leading computer vision architectures for contour detection rely solely on feedforward operations. One explanation for this inconsistency is that existing recurrent neural network models do not capture the contextual mechanisms that biological vision relies on for contour integration. Here we rectify this issue by extending a neural field model for contextual interactions in primate visual cortex into a trainable module that can learn the patterns of feedback connections via backpropagation. We next introduce the \gnet, which incorporates this module into a deep network for dense image prediction. We find that the \gnet performs on par or better than the state-of-the-art model for contour detection, demonstrating the effectiveness of recurrent contextual processing. We also find that training the \gnet for contour detection in natural images causes it to exhibit a similar ``tilt illusion'' in orientation estimation as humans; a non-trivial contextual bias which has mystified visual psychologists. The emergence of this visual illusion supports the theory that contextual illusions are a feature -- not a bug -- of robust visual strategies implemented by recurrent contextual processing.