Self-organization of spatiotemporal receptive fields and laterally connected direction and orientation maps

James A. Bednar, Risto MiikkulainenNeurocomputing 2003 · DOI

TL;DR

This paper presents a computational model that demonstrates how the visual cortex can develop orientation and direction selectivity, along with specific lateral connections, purely through self-organization driven by moving visual stimuli. The model successfully replicates key anatomical and functional features of the primary visual cortex (V1), suggesting a unified mechanism for these complex visual processing capabilities.

What Problem Does This Paper Address?

The primary visual cortex (V1) is known for its neurons' selectivity to the orientation and direction of moving visual stimuli. While it's understood that lateral connections between neurons play a role in this processing, their precise function in the development of directional selectivity and the organization of these maps has been unclear. Previous computational models often treated these lateral connections as fixed or focused solely on afferent input learning. This paper aims to bridge this gap by showing how both the receptive field properties of individual neurons and the network of lateral connections can emerge simultaneously through a self-organizing process driven by realistic visual input.

Methods

The authors extend the LISSOM (Laterally Interconnected Synergetically Self-Organizing Map) model, now called HLISSOM, to incorporate the ON and OFF channels of the Lateral Geniculate Nucleus (LGN) and their temporal delays. The core idea is that directional selectivity arises from LGN cells with different response latencies.

Here's a breakdown of the key methodological aspects:

  • Hierarchical Architecture: The model simulates a simplified visual pathway, starting with photoreceptors, followed by multiple sheets of LGN units (ON and OFF, each with different time lags), and finally a sheet of V1 neurons.
  • Spatiotemporal Receptive Fields: V1 neurons receive input from LGN cells. The model learns the weights of these afferent connections.
  • Activity-Dependent Learning: The model uses a Hebbian learning rule. When pre- and post-synaptic activities are high, the connection weights are strengthened. This learning is normalized to ensure stability.
  • Lateral Connections: V1 neurons also have excitatory and inhibitory lateral connections. These connections are also learned and adapt based on correlated activity. Weak connections are pruned, leading to patchy connectivity.
  • Moving Stimuli: The model is trained using sequences of moving oriented patterns, simulating natural visual experience.

What's novel here is the simultaneous self-organization of both the spatiotemporal receptive fields (leading to orientation and direction selectivity) and the patterned lateral connections, all driven by the temporal dynamics of the input.

Datasets

The paper does not explicitly mention using pre-existing datasets. Instead, the model is trained using synthetically generated visual stimuli: "moving oriented patterns," described as "two-dimensional oriented Gaussians or grayscale photographic images." The specific experimental setup involved a 54x54 photoreceptor sheet and four 36x36 ON-center and four 36x36 OFF-center LGN sheets.

Key Insights and Ideas

  • Unified Self-Organization: The most significant insight is that a single, activity-dependent learning mechanism can explain the emergence of multiple complex features in V1: orientation selectivity, direction selectivity, and the specific patterns of lateral connectivity. This suggests a parsimonious explanation for cortical development.
  • Role of Temporal Delays: The model highlights how temporal delays in LGN inputs are crucial for developing direction selectivity. Neurons can integrate signals arriving at different times to infer motion direction.
  • Space-Time Inseparability: The learned receptive fields are not simply static spatial patterns with a time-varying gain. Instead, they are inherently spatiotemporal, meaning the preferred stimulus is a moving pattern, not just a static shape.
  • Predictive Lateral Connections: The learned lateral connections are not random. They tend to connect neurons with similar orientation and direction preferences, and their structure reflects the correlations observed during self-organization. This suggests these connections are not just for local processing but also for global organization and potentially binding.

Key Results

  • Realistic Receptive Fields: After training, nearly all V1 neurons developed spatiotemporal receptive fields that were selective for orientation and, importantly, direction. These RFs were found to be "space-time inseparable," mirroring experimental findings.
  • Organized Maps: The model successfully generated smooth, topographically organized maps of orientation and direction preferences, exhibiting features like pinwheels and fractures, consistent with experimental observations in animals.
  • Patterned Lateral Connectivity: Long-range lateral connections were observed to preferentially link neurons with similar orientation and direction preferences, and to extend along the orientation axis, aligning with experimental findings.
  • Perpendicular Preferences: The preferred orientation of neurons was generally found to be perpendicular to their preferred direction of motion, another observation consistent with biological data.

Limitations and Future Work

  • Limitations: The paper acknowledges that the learned receptive fields "can only rarely be expressed as a function of a fixed spatial RF and a time-dependent scaling factor," implying that the spatiotemporal nature is fundamental and not easily decomposed. The model's "weakest connections are eliminated periodically," which is a simplification of biological pruning mechanisms.
  • Future Work: The authors propose several avenues for future research:
  • Experimental Validation: Further validation of the model's predictions regarding lateral connections against specific experimental results.
  • Interaction with Other Stimulus Dimensions: Investigating how orientation, direction, and lateral connectivity interact with other visual features like spatial frequency.
  • Adult Function: Studying how the self-organized connection patterns contribute to adult visual perception, specifically in object binding and segmentation.
  • Motion Aftereffects: Exploring how short-term adaptation of lateral connections could replicate phenomena like the motion aftereffect (e.g., the waterfall illusion).

This work aims to unify explanations of both the development and adult function of the visual cortex into a coherent, self-organizing framework.

Summary updated

BibTeX
@article{bednar2003self,
  author = {James A. Bednar and Risto Miikkulainen},
  title = {{Self-organization of spatiotemporal receptive fields and laterally connected direction and orientation maps}},
  journal = {Neurocomputing},
  year = {2003},
  doi = {10.1016/s0925-2312(02)00735-x},
}
Figure