Modeling Spatial Integration and Normalization Underlying Motion Processing in Receptive Fields of MT Neurons

Abstract

Information processing in sensory neural systems relies on a hierarchical architecture. The peripheral sensory organs encode physical inputs as neural signals, which propagate through a multi-stage network. Progression in the network is associated with increasingly abstract representations, more complex functionality, and larger receptive fields. A classic example is the visual motion processing pathway in primates. The abstraction of motion information initiates when the spatiotemporal characteristics of 1D contours are extracted locally in the primary visual cortex (V1). The resulting neural encoding goes through complex processing to generate a cohesive representation of motion directions of surfaces and objects in the mid-level cortical motion area (MT) through integration of convergent direct and indirect V1 projections. Such integration is spatial – signals from many spatially offset V1 receptive fields combine to form MT receptive fields that have diameters up to ten-fold larger. Strikingly, most modeling studies of MT signal processing have overlooked the profound impact of spatial integration in favor of explaining motion computation related to simple, isolated, and spatially uniform visual stimuli, leaving a gap in our understanding of the implications and mechanisms of spatial integration in mid-level motion processing. To address this issue, I have implemented the first image-computable MT model that includes realistic spatial integration of V1 complex units in a flexible framework. The response of the model to a series of dynamic, compound motion stimuli has led to insights into the spatial processing of motion signals in MT. (1) A random wiring paradigm for V1-MT connectivity can mostly account for the heterogeneity of the sensitivity profile of MT receptive fields observed in vivo, but falls short in explaining the diversity of motion direction selectivity reported in experimental data. (2) The well-established non-linear responses of MT neurons to two spatially offset motion stimuli within their receptive fields can be largely explicated by a cascade of mechanisms spread across model stages, including V1 surround suppression, spatially dependent signal weighting and nonlinear signal integration of inputs arriving in MT, whereas MT population-level normalization is less likely to play a role. (3) The spatial scale of the MT neuron's ability to integrate across local component motions to resolve the true direction of motion of a stimuli is not limited by surround suppression, as proposed in the literature, but is more likely to depend on opponent direction interactions at the V1 level. These findings highlight the intricate interplay between spatial and motion integration in the MT receptive field, and shed some light on the computational processes and the hierarchical architecture of the visual cortex.