Abstract
It is unknown how the visual system encodes local image velocity. Neurons in cortical area MT/V5 respond selectivity to image speed independently of spatial frequency (Perrone & Thiele, 2001), but their activity is tuned to particular speeds rather than linearly related to image velocity. Thus, extracting image velocity presumably involves a computation based on a population of MT neurons, such as vector averaging (Lisberger & Movshon, 1999; Priebe & Lisberger, 2004). Here, we identify a major problem with these vector average schemes and explore possible solutions. Using a model of MT neurons that can be tested with 2-d image sequences (Perrone, 2004), we have found that the performance of vector average schemes is compromised by the problem of spatial scale. The problem arises because MT neurons tuned to high speeds have larger receptive fields than those tuned to slow speeds. As the distance from a moving edge increases, neurons tuned to slow speeds are disproportionately omitted from the summation operation. Consequently, the vector average result is biased towards higher velocities in the region surrounding a moving edge. We have tested an alternative mechanism that incorporates a set of 12 different MT neurons (4 spatial frequencies x 3 temporal frequencies) at each image location, (x, y). An edge moving at speed V forms a ridge of peak activity in the MT array. At (x, y), the maximum output occurs in the MT neurons tuned to speed V. The pattern of activity is different for image locations away from (x, y). Therefore the ridge of maximum MT activity occurring at (x, y) can be isolated using derivatives (both across and along the ridge) and a veridical estimate of local image velocity can be found despite the differences in receptive field sizes across the population of MT neurons.
Supported by a Royal Society of New Zealand Marsden Fund grant.