Abstract
Previous work showed that features interpolated during apparent motion (AM) are represented in the population-level feature responses in primary visual cortex, indicating that the brain fills in details that are absent in raw sensory inputs but are reconstructed during dynamic object transformations via top-down processes (Chong, Familiar, & Shim, 2016). Predictive coding accounts hypothesize that feedback can suppress responses in early sensory cortex when incoming sensory information is predicted by top-down expectations. However, it remains unclear how top-down, filled-in neural representations in early visual cortex are modulated by the predictability of the sensory input. Using fMRI and an inverted encoding model, we examined how neural representations of interpolated features during dynamic filling-in evolves as our prediction builds up, and how they are affected by the predictability of the moving object's trajectory. A gabor patch that was oriented radially to the central fixation, was sequentially presented in each quadrant to induce rotational AM along the circular trajectory. AM trajectory was either predictable, where the gabor appears to move in one direction (clockwise or counterclockwise), or unpredictable, where the direction appears to randomly change. Consistent with the previous finding, regions in V1 retinotopically mapped to the AM path showed feature-selective responses for orientation interpolated during AM. Such responses were absent in the first cycle of AM, implying that information about the overall motion trajectory needs to be extracted first. Crucially, after the first cycle, the feature-selective responses were stronger when the trajectory was unpredictable, compared to when it was predictable by top-down expectations. Our finding is consistent with the predictive coding hypothesis, and suggests that top-down representations of filled-in features in early visual cortex can be created after the initial prediction for the object dynamics is formed, but suppressed later via feedback when the uncertainty of the upcoming sensory input is low.
Meeting abstract presented at VSS 2018