A correlational model for the development of disparity selectivity in visual cortex that depends on prenatal and postnatal phases.

Gregory S. Berns, Peter Dayan, Terrence J. SejnowskiProceedings of the National Academy of Sciences 1993 · DOI

TL;DR

This paper proposes a computational model for how neurons in the visual cortex develop the ability to perceive depth (disparity selectivity). It suggests that a two-phase developmental process, starting with same-eye correlations and then incorporating between-eye correlations, is crucial for generating the diverse range of disparity-tuned neurons observed in animal brains.

What Problem Does This Paper Address?

The visual cortex, particularly the primary visual cortex (V1), is responsible for processing visual information, including depth perception. A key aspect of this is disparity selectivity, where neurons respond preferentially to stimuli at a specific depth relative to the viewer. This selectivity is thought to arise from the way neural connections are shaped by visual experience during a critical developmental period. While it's known that correlated neural activity plays a role, the precise mechanisms and the influence of different developmental stages (prenatal vs. postnatal) on the emergence of both monocular (responding to one eye) and binocular (responding to both eyes) cells with varying disparity preferences remained unclear. This paper aims to model how these different types of neurons, with their specific disparity tuning, come to exist through activity-dependent synaptic plasticity.

Methods

The authors developed a computational model of visual cortex development. This model consists of:

  • Input Layers: Two one-dimensional layers representing the retinae of the left and right eyes.
  • Cortical Layer: A one-dimensional layer representing a simplified cortical area.
  • Geniculocortical Connections: Synapses connecting the retinal layers to the cortical layer. These connections are modifiable using a linear Hebbian learning rule, meaning synaptic strength increases with correlated activity between pre- and post-synaptic neurons.
  • Intracortical Connections: Fixed lateral connections within the cortical layer, modeled as a "Mexican hat" function (excitatory at short range, inhibitory at longer range), to represent local cortical processing.

The core of the model is the Hebbian learning rule, which dictates how synaptic weights change based on the correlation of activity. The authors explored three developmental paradigms by manipulating the presence and strength of same-eye correlations (activity within one eye's retina) and between-eye correlations (activity between corresponding points in the left and right retinae):

  1. No between-eye correlations: Simulating a purely prenatal development scenario.
  2. Constant between-eye correlations: Simulating a purely postnatal development scenario.
  3. Two-phase development: An initial phase with only same-eye correlations, followed by a second phase with both same-eye and between-eye correlations.

A normalization scheme was also employed to prevent synaptic weights from growing unboundedly, ensuring a stable network. This normalization transitioned from subtractive to multiplicative as weights approached zero.

Datasets

Not discussed in the paper. The model uses simulated input patterns based on correlation matrices.

Key Insights and Ideas

  • The critical role of timing: The paper strongly suggests that the timing of between-eye correlations is crucial. Early, consistent between-eye correlations lead to a cortex dominated by binocular cells tuned to zero disparity.
  • Two-phase development is key for diversity: A two-phase model, starting with same-eye correlations (prenatal-like) and then introducing between-eye correlations (postnatal-like), is necessary to generate the observed mix of monocular and binocular cells with varying disparity preferences.
  • Monocular cells for non-zero disparity: The model predicts that more monocular cells tend to be selective for non-zero disparities, while binocular cells are more likely to be tuned to zero disparity. This aligns with experimental observations.
  • Correlation strength matters: The relative strength of between-eye correlations compared to same-eye correlations influences the outcome. Higher between-eye correlations lead to more binocular cells.
  • Topographical organization: The model can reproduce the topographical organization of receptive fields and disparity maps observed in biological visual systems.

Key Results

  • No between-eye correlations: Resulted in a cortex populated solely by monocular cells, lacking well-defined disparity tuning.
  • Constant between-eye correlations: Led to a cortex entirely composed of binocular cells, all tuned to zero disparity.
  • Two-phase development: Produced a mixture of monocular and binocular cells. Crucially, binocular cells were found to have a best disparity close to zero, while more monocular cells exhibited non-zero disparity tuning. This relationship was statistically significant.
  • The model also showed that sharp transitions or "fractures" in the disparity map tended to occur at the boundaries of ocular dominance columns, mirroring findings in orientation maps.

Limitations and Future Work

  • One-dimensional model: The current model is one-dimensional, focusing on horizontal disparity. The authors acknowledge that real visual cortex processes both horizontal and vertical disparities and that their preliminary two-dimensional simulations suggest similar relationships hold.
  • Simplified Hebbian rule: The model uses a simplified linear Hebbian rule. While evidence for Hebbian plasticity exists, the exact mechanisms are complex and not fully captured.
  • Normalization scheme: The normalization procedure is acknowledged as being "unsatisfactory" as it's not based on known biological mechanisms.
  • Lack of subunit structure: The model doesn't account for the detailed subunit structure of cortical cells that contribute to disparity tuning.

Future work could involve extending the model to two dimensions to fully capture horizontal and vertical disparity, incorporating more biologically realistic learning rules, and exploring how differences in developmental timing across species might explain variations in visual cortex architecture. The authors emphasize that despite these simplifications, the core conclusion about the importance of distinct developmental phases and correlational mechanisms for generating complex neural properties likely holds.

Summary updated

Abstract

Neurons in the visual cortex require correlated binocular activity during a critical period early in life to develop normal response properties. We present a model for how the disparity selectivity of cortical neurons might arise during development. The model is based on Hebbian mechanisms for plasticity at synapses between geniculocortical neurons and cortical cells. The model is driven by correlated activity in retinal ganglion cells within each eye before birth and additionally between eyes after birth. With no correlations present between the eyes, the cortical model developed only monocular cells. Adding a small amount of correlation between eyes at the beginning of development produced cortical neurons that were entirely binocular and tuned to zero disparity. However, if an initial phase of purely same-eye correlations was followed by a second phase of development that included correlations between eyes, the cortical model became populated with both monocular and binocular cells. Moreover, in the two-phase model, binocular cells tended to be selective for zero disparity, whereas the more monocular cells tended to have nonzero disparity. This relationship between ocular dominance and disparity has been observed in the visual cortex of the cat by other workers. Differences in the relative timing of the two developmental phases could account for the higher proportion of monocular cells found in the visual cortices of other animals.
BibTeX
@inproceedings{berns1993correlational,
  author = {Gregory S. Berns and Peter Dayan and Terrence J. Sejnowski},
  title = {{A correlational model for the development of disparity selectivity in visual cortex that depends on prenatal and postnatal phases.}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {1993},
  doi = {10.1073/pnas.90.17.8277},
}
Figure