A Dynamical Model of Context Dependencies for the Vestibulo-Ocular Reflex

NeurIPS 1995

TL;DR

This paper presents a dynamical model of the vestibulo-ocular reflex (VOR) that accounts for how context, such as eye position and head translation, modulates its gain. By integrating sensory information from the vestibular system and neural correlates of eye position, the model successfully reproduces experimental data from monkeys, offering insights into the neural computations underlying VOR adaptation.

What Problem Does This Paper Address?

The vestibulo-ocular reflex (VOR) is crucial for stabilizing vision during head movements, ensuring that images remain fixed on the retina. Ideally, the VOR gain (the ratio of eye rotation velocity to head rotation velocity) should be -1 for distant targets. However, real-world scenarios involve varying target distances, head translations, and eye positions, all of which necessitate dynamic adjustments to VOR gain for accurate image stabilization. Previous research often simplified this by assuming a fixed target at infinity. This paper addresses the challenge of modeling these context-dependent modulations of the VOR, aiming to understand how the brain integrates diverse sensory inputs to achieve precise eye movements.

Methods

The authors developed a dynamical model based on a previously established kinematic model of the VOR. The kinematic model describes the ideal compensatory eye rotation by requiring that a target remains stationary on the retina, incorporating head rotation (from semicircular canals), head translation (from otoliths), and eye position/vergence angle.

The dynamical model builds upon this by proposing a neural architecture and computational mechanisms. Key aspects include:

  • Integration of Sensory Inputs: The model combines signals from semicircular canals (head rotation velocity) and otoliths (head translation velocity).
  • Neural Integrator: Eye position is assumed to be provided by a "neural integrator," a known neural circuit in the brainstem responsible for maintaining eye position. The model uses a simplified representation of this integrator with a gain and a time constant.
  • Contextual Modulation: The core of the dynamical model lies in how it incorporates context. The cerebellum is hypothesized to play a significant role in this modulation, integrating eye position and vergence signals with vestibular inputs. The vestibular nucleus is also modeled to contribute, particularly with transient eye position information.
  • Vector Computations: The model employs vector mathematics to represent rotations and translations in three dimensions, reflecting the complex geometry of eye and head movements.
  • Parameter Fitting: The model's parameters were adjusted by hand to match experimental data, with the constraint that fitting one response often affected others, implying strong interdependencies.

Datasets

The paper evaluates its dynamical model by comparing its predictions to experimental data from monkeys, specifically the data collected by Snyder & King (1992). This existing dataset, which recorded monkey eye velocity responses to head movements with varying target distances and rotational axes, serves as the benchmark for the model's performance. No new datasets were introduced.

Key Insights and Ideas

  • Context is King for VOR: The VOR isn't a simple reflex; its effectiveness relies heavily on integrating contextual information like where the eyes are pointing and how the head is moving in space.
  • Cerebellum as a Modulator: The cerebellum is proposed as a key computational hub for VOR context modulation, likely combining multiple sensory streams to fine-tune eye movements.
  • Neural Integrator's Role: The "neural integrator" is essential not just for holding gaze but also for providing the eye position signal that dynamically shapes the VOR.
  • Vector Math in the Brain: The model's reliance on vector computations suggests that the brain might be performing complex geometric transformations to achieve accurate gaze stabilization.
  • Interplay of Inputs: The model highlights how different sensory inputs (rotation, translation, eye position) interact dynamically, with time delays and synergistic/antagonistic actions contributing to the observed VOR responses.

Key Results

The dynamical model successfully reproduced many characteristics of the experimental data from monkeys, particularly the modulation of eye velocity responses with target distance and the location of the head's rotation axis.

  • Target Distance Modulation: The model captured the fluctuations in eye velocity between 20-50 ms and the subsequent decrease and increase modulated by target distance (50-80 ms).
  • Rotation Axis Modulation: The model showed good agreement with data for specific locations of the rotation axis, with waveforms being close to experimental results.
  • Oscillations Explained: The model suggests that observed oscillations in VOR responses are likely due to acceleration components in neural signals, time delays in sensory processing, and the interplay of canal and otolith signals.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • Parameter Fitting: While the model fits many aspects of the data, the fitting process was done by hand, and the number of parameters is not small. The authors note that fitting one response often required adjusting parameters that affected other responses, indicating tight constraints on the parameter space.
  • Representation: The vector representation used in the model is a mathematical construct. The authors suggest that other representations, like quaternions, might also be relevant and that detailed comparisons with neuronal recordings are needed to determine the brain's actual computational strategy.
  • Learning Mechanisms: The paper poses the question of how the central nervous system learns these complex transformations. The cerebellum is a prime candidate for this learning, and the authors are working on a model where the cerebellum learns to anticipate sensory inputs and use prediction errors to modulate the VOR in real-time.
  • Experimental Validation: Direct comparison between the model's predictions and Purkinje cell recordings is proposed as a way to gain deeper insights into the specific inputs and computations within the cerebellum.

Summary updated

BibTeX
@inproceedings{1995dynamical,
  title = {{A Dynamical Model of Context Dependencies for the Vestibulo-Ocular Reflex}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1995},
}
Figure