Abstract
It is still an open question as to how the image velocity of a moving edge can be derived from the outputs of a small population of speed-tuned Middle Temporal (MT/V5) neurons. If the MT neurons span a range of spatial frequencies (sfs), then low sf units will dominate the output simply because of their larger receptive fields. The output distribution across the population of MT neurons becomes skewed at locations away from the edge and coding schemes such as winner-takes-all or the weighted vector average will produce an incorrect velocity estimate. We have overcome this problem in a model that uses inhibition between MT neurons of different spatial scales. Let MT1, MT2, MT4 and MT8 represent the outputs of model MT neurons tuned to speeds 1, 2, 4 and 8 deg/s (with peak tf = 4 Hz). We construct a basic ‘2nd derivative’ velocity estimator (e.g., dMT2) tuned to 2 deg/s by combining the outputs as follows: dMT2 = MT2 − .5 MT1 − .5 MT4. Our new code also uses another set of model neurons (rMT1, rMT2 etc.,), identical to the first, but which have their peak tf ‘retuned’ to 8Hz (Perrone, JOV, 2005). We have discovered that by modulating the MT input neurons to dMT2 with the outputs of the rMT neurons we are able to eliminate the spatial scale problem described above. Furthermore, the rMT neurons in our model have the same speed-contrast dependence shown in actual MT neurons (Krekelberg, et al., J. Neurosci., 2006) and the dMT neurons display similar center-surround contrast effects seen in the MT data of Pack, et al. (J. Neurophys., 2005). The new velocity code is also able to replicate the positive and negative shifts in grating speed estimates caused by changes in contrast (Thompson et al., Vis. Res., 2006).
Supported by the Marsden Fund Council from Government funding, administered by the Royal Society of New Zealand.