Testing models of cortical area MT

Abstract

The visual motion of rigid objects can be analyzed through two stages, one that responds to the oriented image components, and one that finds the motion consistent with these components (method of intersection of constraints). These stages are considered to correspond to two classes of neurons in cortical area MT. Neurons are commonly classified by comparing their direction tuning for a plaid (composed of two drifting sinusoidal gratings) with the predictions of two abstract models. The response of the “component-direction-selective” model is just the sum of the responses to the plaid components alone. By contrast, the “pattern-directionselective” model predicts the same direction tuning for gratings and plaids, and computes the direction of a plaid with the intersection of constraints method. More recently Simoncelli and Heeger (1998) proposed physiologically more detailed models (hereafter the SH models) for both type of MT cells. In these models, MT cells combine the output of appropriate model V1 cells selected by an intersection of constraints rule. These models account qualitatively for a wide range of phenomena, but both their large number of parameters and the fact that their response is not given in a closed equation form make them difficult to fit to data. We have tested all these models by fitting large data sets obtained by Majaj, Carandini & Movshon (1998) in area MT of paralyzed, anesthetized macaques. Stimuli included drifting gratings and plaids with (1) variable direction of the two components; (2) variable direction of the first grating and variable contrast of the other. The abstract models were extended with a simple form of contrast-dependence and with a contrast-gain control mechanism acting through divisive inhibition (normalization), while closed form equations were developed for a reduced form of the SH models. The abstract model for component-direction selective cells extended with a contrast gain control correctly predicted the responses of a subset of cells. By contrast, even though it had the largest number of free parameters, the abstract model for pattern-direction selective cells was generally inferior to all the other models. This was particularly evident for those stimuli in which gratings had different contrasts. However, these data sets are possibly strongly influenced by contrast adaptation in area V1, an effect that has not been modeled in this study. Our results support the SH models, which together correctly predicted the responses of a large fraction of cells. The predictions of the abstract and the SH models for component-direction selective cells were similar if both were provided with a gain control. On the other hand, on a subset of cells, the SH model for pattern-direction selective cells performed better than all the other models. Our results further confirm that stimuli often used for the classification of MT cells may be inadequate. Indeed, the predictions of the SH pattern-direction selective and componentdirection selective models are substantially different only for slow moving gratings, not for fast moving gratings.