Discussion

We have found that the computation of true pattern direction in area MT is in fact a dynamic process which evolves during the first 200 ms after a plaid stimulus is presented. Pattern cells show a range of behavior. Some pattern cells show component selective behavior in the beginning of their response, and later change to pattern selective. Other pattern cells, however, maintain their pattern selectivity throughout their response. We found no component cells which reliably behaved in a pattern selective manner at any time of their response. In addition, pattern cells show a longer response latency than component cells by about 6 ms. Our results support the notion that cortical processing of two-dimensional motion signals by MT neurons is a dynamic process which is shaped in the first few hundred milliseconds of their responses.

In a recent study of human subjects, Masson and Castet (2002) examined the short-latency ocular following response to grating and plaid stimuli. With conventional plaids (made from two orthogonal moving gratings) or single gratings, subjects showed a very fast response which was close to the true motion direction from the very beginning. With ``unikinetic plaids'' (made from one stationary and one moving grating separated by 45$^\circ$ in orientation), they found that the ocular following response initially followed the grating motion and later the plaid motion. The latency difference between these two signals was approximately 20 ms. This result demonstrates that the encoding of two-dimensional motion in plaid patterns takes additional processing time compared with gratings. Their latency of 20 ms is in a range similar to that found in our data. Because we do not know the way in which neurons are combined to produce this latency difference, it is hard to make a direct comparison between the latency found by their study and ours. However, the combination of quantitative measures from both psychophysics and physiology is essential in constraining models of how motion perception arises from neuronal responses.

Li et al. (2001) have recorded responses to bar stimuli in the posteromedial lateral suprasylvian area (PMLS) of the cat, which is considered to be analogous and perhaps homologous to primate area MT. They used fields of randomly placed iso-oriented line segments (modeled after Lorenceau et al. (1993)), and adjusted the length of the lines and the angle between their orientation and direction of drift. When the lines moved perpendicular to their orientation, the orientation and direction cues are consistent. When the lines are moved an an angle which is oblique to their orientation, the two cues do not match. Thus, as in a plaid, a cell is faced with the choice between responding to the orientation of the elements in the pattern or the true direction of the overall stimulus. They found that the initial responses of cells in their population, within 200 ms after the onset of the stimulus, tended to be more CDS than the later responses, from 200 ms to about 2000 ms. Increasing the length of the bars in the field shifted the direction tuning toward the CDS prediction. In the limit, if the bars extend beyond the aperture, the stimulus becomes like a grating and PDS behavior is not possible.

In area MT of the alert macaque, Pack and Born (2001) found behavior similar to that of Li et al. (2001). They too used a bar stimulus modeled after Lorenceau et al. (1993), in which the orientation of the bars was set to be either perpendicular to or at a 45$^\circ$ angle with the direction of motion. They found that when the bars moved obliquely to their orientation, their population of cells initially responded to the orientation of the bar segments, but later signaled the true direction of motion. This dynamic change between signaling the orientation and direction was completed by 150 ms in their population of MT neurons. They did not divide their cells into pattern and component on the basis of grating stimuli, so it is somewhat difficult to compare their data to ours. However, if we were to perform a similar experiment, combining our data across all cells, our results should agree with theirs for the following reasons. Component cells, which we have found to have shorter latencies and reach their characteristic response faster than pattern cells, would dominate the early time-average of the population. Pattern cells, with their longer latencies and increased time to evolve their response, would not affect the early time-average, but would slowly bias it toward the pattern direction.

This explanation goes part of the way toward accounting for the Pack and Born (2001) result. However, by 150 ms after stimulus onset, the dynamic changes they report are complete, and the preferred direction in the population is within $\pm 10^\circ$ of the true stimulus direction. In our explanation, as component cells would not show considerable dynamics over the course of the stimulus presentation, the population average would continue to reflect their response to the orientation components of the stimulus, ignoring the true direction of the pattern. The solution to this discrepancy would appear to lie in the stimulus differences between our two studies. As Li et al. (2001) report, changing the length of the bars in their stimulus can have a considerable influence on the pattern and component classification of cells. In addition, both psychophysical (Stoner et al., 1990) and physiological evidence (Stoner and Albright, 1992) indicates that with grating stimuli, changes to the contrast of the intersections results in changing perception and a corresponding change in neuronal response. We thus conclude that our results are broadly consistent with those reported by Pack and Born (2001), and that stimulus differences are the primary reason for any of the discrepancies.

One model that has been proposed to account for the behavior of MT neurons is an extension of the normalization model in V1 (Simoncelli and Heeger, 1998; Heeger et al., 1996). In this framework, the evolution of PDS behavior is roughly instantaneous. We have shown that the average PDS response evolves much slower than the average CDS response (by about 40-60 ms). This finding, along with the results from other laboratories described above, makes it clear that the population response of MT neurons evolves dynamically in the first few hundred milliseconds after stimulus onset. Clearly, current models of MT neurons will need to be revised to account for this new data. However, it is not yet clear what the circuitry is that underlies this delay. For at least two reasons, the magnitude of the delay may not be as large when dealing with a model of individual neurons. First, the population delay here includes significant differences between neurons in response latency (PDS neurons respond later on average than CDS neurons). Although this is important for any model of the readout of the population response in MT, it would be an overestimate of the time that it takes for any single neuron's response to reach a steady state. Second, the significance criterion used here for individual cells was also shown on the plots for the population average. This is merely for reference, as the population response of PDS and CDS neurons reaches significance much earlier. This time to reach significance is important for any model of behavioral decisions based on pattern discrimination, like experiments described above from Masson and Castet (2002).

Some of the studies on the perception of two-dimensional motion stimuli have speculated on the neuronal circuitry involved in generating the dynamic changes reported here and by others. In particular, there are several mechanisms proposed to generate these dynamics: (1) Feedback from higher cortical areas, which presumably decode more complicated stimulus features, might enter MT and influence their responses. The earliest responses may reflect feed-forward processing, while the dynamics may reflect the influence of the feedback signal. (2) There may exist a two-pathway mechanism by which the processing of different stimulus elements takes place (Wilson et al., 1992). If one of the pathways was slower, dynamics in MT might reflect the additional time needed for signals from this pathway to propagate. (3) Within MT, cells may vary in their dynamics based on the position they occupy within the local cortical circuitry (e.g., input layers vs. output layers). Additional studies will be necessary to determine the validity of these hypotheses. In particular, layer information would prove useful in developing a model of the input and output circuitry in MT. Furthermore, intracellular data could prove invaluable in determining how the inputs to MT neurons from V1 (and potentially other areas) are combined to produce PDS behavior. Our study provides some quantitative measurements of timing that will prove useful in ultimately determining the circuits which underly neuronal responses in MT.