September 2018
Volume 18, Issue 10
Open Access
Vision Sciences Society Annual Meeting Abstract  |   September 2018
Pinwheel-like Iso-Orientation Domains in a Convolutional Neural Network Model
Author Affiliations
  • Eshed Margalit
    Department of Psychology, Stanford University
  • Hyodong Lee
    Department of Brain and Cognitive Sciences, MIT
  • James DiCarlo
    Department of Brain and Cognitive Sciences, MIT
  • Daniel Yamins
    Department of Psychology, Stanford University
Journal of Vision September 2018, Vol.18, 771. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Eshed Margalit, Hyodong Lee, James DiCarlo, Daniel Yamins; Pinwheel-like Iso-Orientation Domains in a Convolutional Neural Network Model. Journal of Vision 2018;18(10):771.

      Download citation file:

      © ARVO (1962-2015); The Authors (2016-present)

  • Supplements

Introduction A hallmark of the mammalian visual system is the orderly arrangement of neurons into functional maps, ranging from pinwheel-like representations of oriented stimuli in early visual areas (Bonhoeffer and Grinvald, 1991) to category-selective regions (e.g., face patches) in higher visual areas (Kanwisher et al., 1997). Despite the identification of such maps in many species, it remains unclear which characteristics may depend on visual experience and, critically, which kinds of visual experience are necessary and sufficient for the formation of such maps. Methods Here, we take a goal-driven modeling approach to understand the development of primary visual cortex. Under the hypothesis that biological maps emerge as a solution to the problem of understanding visual input, we trained a deep convolutional neural network to categorize natural images while respecting the retinotopic constraints present before the onset of visual experience. Formally, units in the model's first convolutional layer were assigned a 2-dimensional spatial position consistent with their receptive fields, and the network was penalized during training if adjacent units differed too strongly in their responses or if distant units had similar responses. Results We evaluated the trained model by presenting structured stimuli—gabor wavelets at specified spatial frequencies, orientations, and positions—and constructing tuning curves over the parameters of interest for units in the first convolutional layer. The map of orientation tuning is strikingly similar to biological findings, with clear iso-orientation domains and pinwheel-like patterns. The spatial frequency tuning map similarly recapitulates many of the features observed in macaques and cats. Conclusion Our results suggest that training a model to solve a challenging visual task—image classification—is sufficient to reproduce maps of orientation and spatial frequency preference in early visual areas. Future work will be needed to determine the developmental program and model architecture required to recapitulate functional maps along the extent of visual cortex.

Meeting abstract presented at VSS 2018


This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.