Visual cortical mechanisms of perceptual grouping: interacting layers, networks, columns, and maps

William D. Ross, Stephen Grossberg, Ennio MingollaNeural Networks 2000 · DOI

TL;DR

This paper presents a neural model of the visual cortex that explains how we group visual elements into coherent objects, even when they are partially occluded or only suggested by context. The model integrates the layered structure of the cortex, neural networks, and spatial maps to demonstrate how "illusory contours" and other Gestalt phenomena arise from the interplay of excitatory and inhibitory neural signals. This work offers a computational framework for understanding fundamental aspects of visual perception.

What Problem Does This Paper Address?

The human visual system is remarkably adept at perceiving complete objects and surfaces from incomplete or ambiguous visual information. This ability, known as perceptual grouping, allows us to see continuous lines, shapes, and surfaces even when they are only suggested by local features like edges or textures. For instance, we can perceive a complete square formed by four Pac-Man shapes, even though the sides of the square are not explicitly drawn. While psychophysical studies have long described these phenomena (e.g., Gestalt laws), understanding the underlying neural mechanisms within the visual cortex has been a significant challenge. Specifically, how the layered architecture (laminae), functional units (columns), and spatial representations (maps) of the visual cortex cooperate to achieve this context-sensitive grouping, especially reconciling bottom-up sensory input with top-down feedback, remains unclear. This paper aims to bridge this gap by proposing a detailed neural model that simulates these processes.

Methods

The authors developed a computational neural model that simulates the interactions within and between cortical areas V1 and V2, as well as their input from the Lateral Geniculate Nucleus (LGN). The model is built upon several key principles:

  • Laminar Circuits: The model explicitly incorporates the layered structure of cortical areas (layers 2/3, 4, 6). Different layers have distinct roles, with layer 4 receiving primary input, layer 2/3 handling horizontal interactions for grouping, and layer 6 providing feedback.
  • Bipole Property: A crucial computational element is the "bipole property," where a neuron responds only when stimulated on two sides of its receptive field (indicating a potential contour between inducers), but not when stimulated on only one side (preventing extrapolation from single points). This is implemented through a balance of excitatory and inhibitory connections, particularly in layer 2/3.
  • Folded Feedback: The model utilizes a "folded feedback" mechanism where signals from higher layers (2/3) feed back to lower layers (4 via 6). This allows grouping signals to interact with and modulate the initial sensory input without losing analog sensitivity (i.e., preserving graded responses).
  • Analog Sensitivity: The model emphasizes preserving analog sensitivity, meaning neural responses are graded and reflect the strength of the input, rather than being purely binary (on/off). This is achieved through shunting inhibitory networks.
  • Hierarchical Processing: The model simulates processing across multiple stages, from LGN to V1 and V2, with V2 having larger receptive fields to account for longer-range grouping.

The model's equations describe the dynamics of neural activation within these circuits, and simulations are performed by iterating these equations until equilibrium is reached.

Datasets

The paper does not explicitly mention using specific datasets for training or evaluation in the traditional machine learning sense. Instead, the model's performance is evaluated by comparing its simulated outputs to existing psychophysical and neurophysiological data. These data are derived from experiments described in the cited literature, often involving specific visual stimuli designed to elicit perceptual grouping phenomena (e.g., Kanizsa squares, line ends, noisy images).

Key Insights and Ideas

  • Layered Computation is Key: The model demonstrates that the specific arrangement of neurons in different cortical layers is fundamental to perceptual grouping. Layer 2/3's horizontal connections enable cooperation, while layer 6's feedback to layer 4 and LGN allows for selection and refinement of groupings.
  • Feedback is Crucial for Selection: Top-down feedback from higher cortical areas is not just for top-down control but is essential for selecting the "winning" perceptual grouping when multiple interpretations are possible, without sacrificing sensitivity to the original sensory input.
  • The Bipole Property Explained: The model provides a concrete neural implementation of the bipole property, showing how a balance of excitation and disynaptic inhibition can achieve interpolation (completing contours between inducers) while preventing extrapolation (creating contours from single inducers).
  • Analog Coherence: The model suggests that perceptual grouping is not just about detecting local features but about assessing "analog coherence" – the spatially distributed evidence for a feature's likelihood or strength. This allows the system to form robust percepts even in noisy or ambiguous conditions.
  • Scale Matters: The difference in receptive field sizes between V1 and V2 is modeled to explain why V2 can group elements over larger distances than V1, aligning with experimental observations.

Key Results

The model successfully simulates a range of psychophysical phenomena related to perceptual grouping, including:

  • Illusory Contour Formation: The model generates illusory contours in response to stimuli like Kanizsa squares, demonstrating the bipole property in action.
  • Gestalt Laws: It replicates the Gestalt laws of proximity and good continuation, showing how grouping strength changes with the arrangement and density of inducing elements. For example, it shows how adding more horizontal elements can bias perception towards a horizontal grouping (proximity) and how line ends can be interpolated to form continuous contours (good continuation).
  • Context Sensitivity: The model demonstrates how the strength of perceived contours is sensitive to contextual factors, such as the "support ratio" (ratio of real to total contour length) in Kanizsa squares, and the number/density of line-end inducers.
  • Signal-to-Noise Enhancement: The model shows how its grouping mechanisms can enhance the signal-to-noise ratio, allowing for the perception of faint boundaries embedded in noise.
  • V1/V2 Differences: The model accounts for differences in grouping capabilities between V1 and V2, with V2 showing longer-range grouping due to its larger receptive fields.

Limitations and Future Work

  • Simplifications: The model uses simplified representations of neurons (e.g., single voltage compartments) and assumes certain parameters are fixed. The authors note that the layer 2/3 complex cells are more sensitive to parameter variations than other cells.
  • Specific Stimuli: While the model reproduces many phenomena, its direct application to more complex 3D perception or dynamic scenes is not fully explored. The paper mentions that depth perception, a complexity not handled by the current model, might restore vertical groupings behind bars in one example.
  • Future Work: The authors suggest that the model provides a framework for further investigation into how cortical circuits control perceptual grouping and preserve analog coherence under diverse stimulus conditions. They also point to the potential role of corticogeniculate feedback in endstopping and the broader implications for understanding feature binding and attention.

Summary updated

BibTeX
@article{ross2000visual,
  author = {William D. Ross and Stephen Grossberg and Ennio Mingolla},
  title = {{Visual cortical mechanisms of perceptual grouping: interacting layers, networks, columns, and maps}},
  journal = {Neural Networks},
  year = {2000},
  doi = {10.1016/s0893-6080(00)00040-x},
}
Figure