Spatial Preference-Weighted Representation of Multiple Stimuli in Cortical Area MT

Segregating objects from one another and the background is essential for perception. In natural scenes, it is common to encounter spatially separated stimuli, such as distinct figure-ground regions, adjacent objects, and partial occlusions. Neurons in mid- and high-level visual cortex have large receptive fields (RFs) that often encompass multiple, spatially separated stimuli. It is unclear how neurons represent and segregate multiple stimuli within their RFs, and the role of spatial cues. We recorded neuronal responses in the middle-temporal (MT) cortex of male rhesus macaques to spatially separated random-dot stimuli that moved simultaneously in two directions. We found that across motion directions, responses to bidirectional stimuli were systematically biased toward the component located at the neuron’s preferred RF subregion. The sign and magnitude of this spatial-location bias were predicted by the neuron’s spatial preference for single stimuli presented in isolation. The bias persisted when attention was directed away from the RF, changed predictably when the motion border was shifted within the RF, and developed over time, indicating a stimulus-driven spatial weighting mechanism. Bidirectional responses were well described by a spatial preference-weighted normalization model, in which the responses elicited by individual components are weighted and combined according to the neuron’s spatial preference. Bidirectional stimuli also reduced model-inferred gain variability, suggesting potential engagement of response-stabilizing mechanisms. These results show that MT neurons capitalize on RF spatial selectivity to represent multiple motion components. Across a neuron population with diverse spatial preferences, this coding strategy provides a neural substrate for segregating spatially separated moving stimuli. Significance Statement Elucidating how neurons represent multiple visual stimuli is essential for understanding neural coding in natural scenes. We found that MT neurons do not simply average spatially separated motion components within their RFs. Instead, small spatial preferences measured with single stimuli were amplified into substantial spatial biases when two stimuli were presented together, causing neurons to preferentially represent the component at their preferred RF subregion. This response pattern was captured by a spatial preference-weighted normalization model, extending the normalization framework by showing that RF spatial selectivity determines how multiple motion signals are weighted. A possible circuit implementation further suggests how spatially selective inputs, broad direction pooling, and divisive normalization could support segregation of spatially separated moving stimuli.