Effects of Spike Timing Underlying Binocular Integration and Rivalry in a Neural Model of Early Visual Cortex

Erik D. LumerNeurIPS 1997

Introduction

When our eyes see similar scenes, the brain seamlessly merges them into a single, cohesive 3D view—a process known as binocular integration. However, when the eyes are presented with radically different images (for instance, horizontal lines in the left eye and vertical lines in the right), we do not perceive a mashed-up grid. Instead, our perception flips back and forth between the two images in a phenomenon called binocular rivalry.

While researchers have long known that rivalry-related activity occurs in "higher" areas of the brain (the extrastriate cortex), it has remained unclear how these areas "know" when to start competing. Because eye-specific information is largely lost by the time signals reach these higher areas, what is the signal that distinguishes "matching" inputs from "conflicting" ones? This paper proposes a neural model where the temporal synchronization of neuronal spikes in the early visual cortex—rather than just their overall firing rates—serves as that critical signal.

The Neural Model Architecture

To investigate this, the author constructed a simplified computational model of the mammalian early visual pathway. The model uses standard integrate-and-fire neurons, which simulate how real neurons accumulate electric charge and fire a spike once a threshold is reached.

The network consists of four hierarchical stages:

  1. LGN (Lateral Geniculate Nucleus): Relays raw visual input from the eyes.
  2. V1 Layer 4: Monocular regions that respond specifically to either the left or right eye.
  3. Vp (Primary Cortical Sector): The first stage where inputs from both eyes combine.
  4. Vs (Secondary Visual Area): A higher-order area where complex features (like line crossings) are extracted.
Schematic of the four-stage neural model

The model features local recurrent circuits with both excitatory and inhibitory connections. Crucially, inhibition is strongest between groups of neurons that prefer different orientations (e.g., horizontal vs. vertical). It also includes backward projections from higher areas (Vs) to lower areas (Vp), mimicking the brain's top-down influence.

The Role of Spike Timing

The core hypothesis tested by the model is that spike timing dictates whether signals are integrated or forced to compete:

  • Synchronization as "Glue": When both eyes see the same thing (congruent inputs), neurons in the early stages of the model fire in sync. This temporal coordination "protects" the signals as they move up the processing chain, allowing multiple features to be perceived simultaneously.
  • Desynchronization as a Trigger for War: When inputs conflict, the neurons in V1 become desynchronized. This lack of coordination makes the signals vulnerable to "winner-take-all" competition in higher cortical areas.

To prove that timing is causal, the author ran a control experiment driving the model with identical inputs but artificially introducing a 10ms delay between the two eye pathways. This tiny timing difference was enough to trigger rivalry, demonstrating that spike timing—not just the image content—is a primary driver of the phenomenon.

Key Findings and Network Dynamics

The model was tested using simulated visual stimuli: conflicting inputs (a left-tilted grating for one eye and a right-tilted grating for the other) and congruent inputs (identical "plaid" patterns presented to both eyes).

Firing rates across network layers for conflicting vs. congruent inputs
  • Vs Activity Patterns: During rivalry simulations (conflicting inputs), only one feature-selective group in the secondary visual area (Vs) remained active at a time, reflecting the perceptual "flipping" of binocular rivalry. Conversely, during integration (congruent inputs), all relevant groups sustained high, stable firing rates.
  • Rivalry Index: The author developed a quantitative "Rivalry Index" (0 for perfect integration, 1 for total exclusion). Conflicting inputs produced an index near 0.4–0.5 in higher areas, while congruent inputs stayed near 0.1, showing a clear mathematical distinction between the two states.
  • V1 Correlates: The model predicts that even if average firing rates in V1 look similar across both conditions, the coincidence rate (how often neurons fire at the exact same millisecond) is significantly higher during integration.

Limitations and Future Work

While the model successfully demonstrates how spike timing can initiate binocular rivalry, it relies on simplified dynamics. It focuses primarily on the onset of rivalry and does not deeply explore the long-term, stochastic switching (the way perception flips every few seconds), which likely involves complex adaptation and noise processes. Furthermore, the model simulates only four stages, whereas the real visual system involves dozens of interconnected areas.

Despite these limitations, the paper suggests that "timing-sensitive competition" might be a fundamental property of the cortex. Future work could explore whether this spike-timing mechanism applies to other senses, helps explain how the brain selects which stimuli to attend to, or influences motor action selection.

Summary updated

BibTeX
@inproceedings{lumer1997effects,
  author = {Erik D. Lumer},
  title = {{Effects of Spike Timing Underlying Binocular Integration and Rivalry in a Neural Model of Early Visual Cortex}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1997},
}
Figure