Computational Model of Neurons in Visual Area MT

Electrophysiological studies indicate that neurons in the Middle Temporal (MT) area of the primate brain are selective for the velocity (both direction and speed) of visual stimuli. We've developed a descriptive computational model of MT physiology, in which local image velocities are represented via the distribution of MT neuronal responses. The computation is performed in two stages of identical architecture, corresponding to neurons in cortical areas V1 and MT. Each stage computes a weighted linear sum of inputs, followed by rectification and divisive normalization. The output of the model corresponds to the steady-state firing rates of a population of MT neurons, which form a distributed representation (population encoding) of image velocity for each local spatial region of the visual stimulus. One can think of the distributed set of responses as representing a probability density over local image velocity. The model accounts for a wide range of physiological data.


Model Structure

Model wiring diagram

The diagram shows the essential aspects of the model. In the first (V1) stage, each neuron computes an inner product of the image contrast with a space-time oriented receptive field. Our receptive fields are directional third derivatives of a Gaussian. The use of directional derivatives is a fundamental aspect of the model, as it allows us to interpolate responses at arbitrary space-time orientations from a fairly small (28) fixed population of V1 neurons. [NOTE: The Gaussian is chosen for simplicity: alternative choices would lead to beneficial properties such as causality or spatio-temporal separability, without fundamentally changing the steady-state responses]. These linear outputs are half-squared (halfwave-rectified and squared), and then divisively normalized. The normalization factor is a sum of a semisaturation constant (sigma) and the responses of neurons at all space-time orientations, and within a local spatial neighborhood.

The second (MT) stage of the model performs a summation over V1 afferents consistent with a given pattern velocity. These are illustrated in the spatio-temporal frequency domain. This is essentially a neural implementation of the "intersection-of-constraints" construction (see references). This construction is appropriate for so-called "pattern cells" in area MT. As shown in the paper below (Vis. Res. 1998), variants can be constructed for "component cells", by summing afferents of the subset of V1 cells with the same direction tuning.

Note: this is a model of steady-state responses to spatio-temporally homogeneous stimuli.


Software

A matlab implementation of the model is available here. A brief description may be found in the README file, and updates and changes are listed in the ChangeLog file.


Partial List of References

This Model

Related Computational Motion Models

Updated: February 19 2016.
Created: May 1997
top