Paria Mehrani presents “Learning a model of shape selectivity in V4 cells reveals shape encoding mechanisms in the brain” at Neuroscience 2021
Venues: Neurosciences 2021, Virtual
Paper:Learning a model of shape selectivity in V4 cells reveals shape encoding mechanisms in the brain
Abstract:
The mechanisms of local shape information transformation from V1 to more abstract representations in IT are unknown. Studying the selectivities in intermediate stages of transformation suggest plausible mechanisms. For example, findings in Macaque V4 responses suggest that these neurons are selective to convexities and concavities at a specific position in the stimulus (Pasupathy et al., 2001). Although such investigations reveal intermediate shape representations in the brain, they often do not suffice in capturing complex and long-range interactions within the receptive field due to imposing priors on tunings, e.g., fitting a single Gaussian to neuron responses. We propose a learning-based approach that eliminates the need for such strong priors. Specifically, we investigate shape representation in Macaque V4 cells and formulate shape tuning as a sparse coding problem according to previous findings of V4 neurons (Pasupathy et al., 2001). We emphasize that our goal is not to find a mapping from the stimulus to V4 responses but rather to study how V4 neurons combine responses of V2 neurons, in this case, curvature-selective V2 cells, to achieve their reported part-based selectivities. To this end, our proposed model takes responses of simulated curvature-selective V2 cells as input by combining two previously introduced hierarchical models 2DSIL (Rodríguez-Sánchez et al., 2012) and the RBO network (Mehrani et al., 2021). With simulated curvature signal as input, our algorithm learns a sparse mapping to V4 responses that reveals each Macaque V4 cell’s tuning and the mechanism by which the tuning is achieved. Our results suggest that V4 cells combine shape parts from the full spatial extent of their receptive field with similar magnitudes of facilitatory and inhibitory contributions. Additionally, our approach suggests a mechanistic model on how object-centered representations in V4 are achieved, providing a better understanding of shape encoding mechanisms in the brain.