Abstract
It's surprisingly easy to recognize objects at different sizes, orientations, and positions in the visual field. Robust object recognition across such transformations suggests that the visual system achieves an (almost) invariant representation at some level. While understanding the computations that underlie view-invariant representation is an ongoing topic of research (DiCarlo & Cox, 2007), it remains unclear exactly where and how invariance is achieved. Here, we aim to provide some insight into this question by exploring the emergence of orientation invariant representations within the visual system using fMRI. We used a rapid-event related paradigm in order to attain highly reliable brain patterns for 40 items (8 distinct objects at 5 orientations each; average reliability across items and subjects, r=.79). We then divided the cortex into several regions of interest, including V1-V3, lateral occipital complex (LOC), and the broader occipitotemporal cortex (OTC). We divided the data set into all possible halves of the 12 runs, and computed the correlation between item patterns for each half for all combinations of the 40 items (1600 correlations). This analysis revealed strong orientation dependence in V1-V3: for a given object, pattern similarity decreased as the difference in orientation between items increased. In contrast, LOC showed evidence for complete invariance: the average pattern for a particular object on one half of runs (e.g., upright face) was just as similar to the same object at the same orientation (upright face) as it was to the same object at any other orientation (45, 90, 135, 180 deg) on the other half of runs. The same degree of orientation invariance was observed in OTC. These results show that object responses in occipitotemporal cortex are completely invariant, not just tolerant, to changes in orientation, and suggest that these invariant representations emerge abruptly in the transition between early visual cortex and LOC.
Meeting abstract presented at VSS 2014