August 2014
Volume 14, Issue 10
Free
Vision Sciences Society Annual Meeting Abstract  |   August 2014
Learning of New Perceptual Groupings - A Biologically Plausible Recurrent Neural Network Model that Learns Contour Integration
Author Affiliations
  • Tobias Brosch
    Institute of Neural Information Processing, Ulm University, Ulm, Germany
  • Pieter Roelfsema
    Netherlands Institute for Neuroscience, Amsterdam, The Netherlands
  • Heiko Neumann
    Institute of Neural Information Processing, Ulm University, Ulm, Germany
Journal of Vision August 2014, Vol.14, 941. doi:10.1167/14.10.941
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to Subscribers Only
      Sign In or Create an Account ×
    • Get Citation

      Tobias Brosch, Pieter Roelfsema, Heiko Neumann; Learning of New Perceptual Groupings - A Biologically Plausible Recurrent Neural Network Model that Learns Contour Integration. Journal of Vision 2014;14(10):941. doi: 10.1167/14.10.941.

      Download citation file:


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

      ×
  • Supplements
Abstract

Problem. Mechanisms of perceptual organization can be subdivided into base-grouping, operating in parallel over the visual field, and incremental grouping that operates sequentially and requires selective attention (Roelfsema, Ann. Rev. Neurosci, 2006). The underlying neural mechanisms recruit circuits and cortical subsystems that interact in feedforward and feedback streams (Poort et al., Neuron, 2012). Evidence suggests that the neural computational mechanisms are not inert but are influenced by perceptual learning (Li et al., Neuron, 2008). It is currently unknown what the underlying mechanisms are to implement such perceptual learning. Method. We propose the biologically inspired REinforcement LEarning Algorithm for Recurrent Neural Networks (RELEARNN). Our model consists of mutually connected model areas that include intra-areal and inter-areal excitatory, inhibitory and modulating connections that influence the mean firing rates of model neurons. Learning alters these connections and utilizes a biologically plausible Hebbian plasticity mechanism that is gated by two factors, a localized attentional feedback and a global reinforcement learning signal. The model is shown to provide a biologically plausible link to the Almeida-Pineda backpropagation scheme (Almeida, IEEE 1987; Pineda, Physical Review Letters, 1987). Results and Conclusion. We demonstrate how RELEARNN can account for the performance of the visual system in two different grouping tasks. The first is a curve-tracing task, and we demonstrate that a model trained in this task qualitatively reproduces the activity profile of neurons in the visual cortex of monkeys. Activation of neurons that are driven by feedforward signals are enhanced by sustained laterally propagated modulations that serve as grouping label. The second task demands the detection of a "snake" of collinearly aligned contour elements. Here, the model reproduces psychometric performance curves as well as neuronal activity in monkey area V1. The new findings suggest that multi-stage grouping operations in the brain may be learned by one common learning mechanism.

Meeting abstract presented at VSS 2014

×
×

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.

×