Modeling Complex Cells in an Awake Macaque during Natural Image Viewing

William E. Vinje, Jack L. GallantNeurIPS 1997 · OpenReview

TL;DR

This paper presents a computational model of complex cells in the primate visual cortex (V1) that better predicts their responses to natural images than previous models. By incorporating nonclassical suppressive mechanisms, the model captures more of the temporal dynamics of cell firing, particularly in response to dynamic natural scenes.

What Problem Does This Paper Address?

Understanding how the brain processes visual information is a fundamental challenge in neuroscience. Complex cells in V1 are a crucial early stage in this processing, responding to specific visual features like orientation and spatial frequency. However, their responses are not simply a linear combination of inputs; they are influenced by complex, non-linear interactions, including suppressive effects from surrounding visual information. Previous models often struggled to accurately predict V1 cell responses to natural, dynamic visual stimuli, falling short in capturing the nuanced temporal firing patterns. This paper aims to bridge this gap by developing a model that accounts for both the classical receptive field properties and these important nonclassical suppressive influences during natural image viewing.

Methods

The researchers developed a model of V1 complex cells built upon a classical receptive field (CRF) energy mechanism. This core mechanism, which detects oriented and spatially filtered stimuli, is then modulated by two nonclassical suppressive mechanisms: a divisive gain control field (G) and a texture contrast field (T). These suppressive fields are hypothesized to reflect influences from outside the classical receptive field and are incorporated with a delay parameter to account for potential latency differences.

The model was applied to "review movies," which are sequences of image patches extracted from natural scenes, designed to mimic the stimulation a V1 cell receives during natural free viewing. The model's parameters were fit to the recorded spike train data (represented as peri-stimulus time histograms, PSTHs) from three individual V1 cells. They explored fitting the CRF energy mechanism alone and then progressively adding the gain control and texture contrast mechanisms, noting the impact of delays in the suppressive surround.

Datasets

The study used data collected from three awake, behaving macaque monkeys. For each cell, a static natural scene was presented, and the animal's eye position was tracked as they freely viewed it. Image patches centered on the cell's CRF were then extracted to create 9-second "review movies." These movies were replayed multiple times, and the cell's average response was computed. The paper does not introduce a new benchmark dataset but rather uses existing single-unit recording data from natural scene viewing.

Key Insights and Ideas

  • Beyond the Classical Receptive Field: The study strongly suggests that nonclassical suppressive mechanisms (gain control and texture contrast) are essential for accurately modeling V1 complex cell responses to natural stimuli, going beyond simpler energy models.
  • The Importance of Temporal Dynamics: The inclusion of a delayed suppressive surround significantly improved the model's ability to match the temporal firing patterns (PSTH) of real cells, particularly in capturing the phasic responses observed at the onset of new stimuli.
  • Naturalistic Stimuli Reveal Limitations: Using dynamic "review movies" derived from natural viewing highlighted limitations of simpler models, especially their inability to capture the phasic nature of real cell responses to saccades.
  • Model Simplicity vs. Explanatory Power: The model, despite using relatively few parameters (7 free parameters), achieves a notable level of predictive power for natural stimuli, suggesting a good balance between complexity and explanatory capability.

Key Results

  • The CRF energy mechanism alone provided modest but significant correlations (r values ranging from 0.35 to 0.59) between model predictions and actual cell responses.
  • Adding the suppressive surround mechanisms (gain control and texture contrast) led to improvements in correlation, with the best correlations achieved when the suppressive effects were delayed.
  • In one specific case (cell 97A, movie B), the inclusion of a delayed suppressive surround dramatically improved the correlation, transforming a stepped model response into a more phasic one that closely matched the data.
  • The model struggled with cells that were not orientation-selective, as the fitting procedure could not reliably determine the optimal CRF orientation. Directionally-selective cells also posed a challenge, likely due to the model's space-time separable receptive field missing motion energy cues.

Limitations and Future Work

  • Approximation of CRF(t): The paper acknowledges that approximating the temporal component of the CRF as a delta function leads to a relatively constant model PSTH during fixations, contrasting with the phasic responses of real cells.
  • Directional Selectivity and Color: The current model does not account for directional selectivity or color tuning, which are known properties of V1 cells.
  • Space-Time Separability: The assumption of a space-time separable receptive field structure may limit the model's ability to capture more complex spatio-temporal interactions.
  • Future Directions: The authors plan to implement more sophisticated energy mechanisms that allow for nonseparable space-time receptive fields, incorporate more detailed surround mechanisms (e.g., asymmetric surrounds), and add a broadband texture contrast term. They also envision extending this modeling approach to extrastriate visual areas (V2 and V4).

Summary updated

Abstract

We model the responses of cells in visual area VI during natural vision. Our model consists of a classical energy mechanism whose output is divided by nonclassical gain control and texture contrast mechanisms. We apply this model to review movies, a stimulus sequence that replicates the stimulation a cell receives during free viewing of natural images. Data were collected from three cells using five different review movies, and the model was fit separately to the data from each movie. For the energy mechanism alone we find modest but significant correlations (rE = 0.41, 0.43, 0.59, 0.35) between model and data. These correlations are improved somewhat when we allow for suppressive surround effects (rE+G = 0.42, 0.56, 0.60, 0.37). In one case the inclusion of a delayed suppressive surround dramatically improves the fit to the data by modifying the time course of the model's response.
BibTeX
@inproceedings{vinje1997modeling,
  author = {William E. Vinje and Jack L. Gallant},
  title = {{Modeling Complex Cells in an Awake Macaque during Natural Image Viewing}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
  url = {https://openreview.net/forum?id=SyWlgFbdZB},
}
Figure