Abstract
Pattern motion sensitivity is a critical response property in MT, and the spatial scale of pattern motion integration in MT neurons provides a key constraint for MT models. Experimental studies have examined how pattern selectivity deteriorates for "pseudoplaids" where the component gratings of the plaid are spatially separated within the MT RF and varied in size (Majaj et al., 2007; Kumbhani et al., 2015). They concluded that pattern motion is computed at a spatial scale ~1/3 of the MT RF. Mechanisms proposed to account for this include V1 direction-tuned normalization, which may arise from surround suppression, or spatially-local computational subunits in MT dendrites. Previous MT models omitted spatial integration or were not image-computable, precluding the testing of alternative hypotheses. We developed an image-computable modeling framework to build pattern-selective MT units (Baker and Bair, 2016), and extended this framework to include V1 inputs with RFs distributed across visual space and summed to produce MT RFs ~5x the diameter of V1 RFs. We tested the hypothesis that the apparent subunit size of spatial integration is determined by V1 monocular opponency, the same mechanism that explains loss of pattern sensitivity for dichoptic plaids (Baker and Bair, 2016). In model units with V1 opponency and tuned normalization, we reproduced the main results of Majaj and Kumbhani, without including explicit spatial subunits. In particular, pattern sensitivity in our MT units degraded when component gratings were shown in a 2x2 grid lacking spatial overlap, but with finer 4x4 grids, pattern sensitivity began to recover. We tested our spatial integration model using binocularly-presented pseudoplaids with opposite motion in each eye, making the novel prediction that 3D-motion tuning remains robust even for pseudoplaids where monocular pattern sensitivity is lost. We are now implementing physiologically-plausible V1 surround suppression to test alternative hypotheses for how the spatial subunit is determined.
Meeting abstract presented at VSS 2017