Three-dimensional spatiotemporal receptive field structure in macaque area MT

A Zaharia, R Goris, J A Movshon and E P Simoncelli

Published in Computational and Systems Neuroscience (CoSyNe), Feb 2017.
Neurons in area MT are selective for direction and speed of visual motion. Some of these neurons are “pattern invariant,” meaning they are velocity-selective regardless of spatial pattern. These properties distinguish them from neurons in primary visual cortex (V1), which are selective for orientation and spatial frequency. Historically, pattern invariance has been quantified by comparing neurons' responses to plaids (two sinusoidal gratings superimposed) with the summed responses to the constituent gratings presented in isolation. More generally, moving images can be described in terms of the distribution of their 3D spatiotemporal frequency content: translating patterns consist of frequencies that lie on a plane. Recent efforts have characterized MT responses using stimuli rich in frequency content, including ``motion-enhanced'' natural movies (Nishimoto and Gallant, 2011) and ``hyperplaids'' (multiple, overlapping gratings; Inagaki et al, 2016). These two studies did not focus on pattern invariance, and arrived at models differing in their descriptions of MT receptive field properties.

We recorded single unit responses in alert macaque MT to novel hyperplaid stimuli, consisting of superpositions of multiple moving gratings that continuously and efficiently sampled 3D frequency space. To increase the dynamic range of neural responses, the stimulus was tailored to each neuron's estimated frequency preferences so that at least half of the components were sampled near the neuron's preferred frequencies.

We use a two-stage model of V1 and MT to capture the selectivity and invariance for spatiotemporal frequency in MT. Both stages include linear weights in 3D frequency space, followed by a nonlinearity. We find this model can both predict hyperplaid responses and account for pattern invariance in responses to single gratings and plaids. Moreover, we find that positive linear weights concentrated near the plane corresponding to the preferred velocity of the neuron, as well as nonlinear tuned suppression, are necessary elements of MT pattern computation.

  • Listing of all publications