Abstract
How view invariance is achieved in recognizing 3D objects has been a matter of some debate. To date, human functional neuroimaging studies of view invariance have been somewhat inconsistent in their results, reporting different levels of across-view tolerance within the object-selective, lateral occipital complex (LOC). Critically, many of these studies used familiar objects as stimuli, thereby confounding prior experience of multiple views with nominal view invariance in measured neural responses. Using fMRI, we investigated the encoding of object viewpoint within human LOC across different timepoints while learning novel 3D objects. In a learning phase, participants were trained to recognize the objects in two views separated by a 90 deg rotation in depth. Following extensive training, participants were asked to recognize the now-familiar objects in a generalization phase featuring four additional, novel views. MRI scans were run during the early and late stages of the learning phase, as well as the generalization phase. Surprisingly, within the LOC, we observed that the signal-to-noise ratio (SNR) of the BOLD response to the trained views decreased over learning, and, moreover, was negatively correlated with behavioral accuracy. Consistently, during generalization, the BOLD response to the two familiar views was significantly lower than the response to the four unfamiliar views, indicating that the observed learning effect was limited to familiar views. Finally, we applied linear support vector machine classifiers and found that the multivoxel patterns across the LOC showed above-chance discrimination of different views, suggesting that the LOC does carry information about object viewpoint. In sum, our results are consistent with the LOC encoding view-specific information about objects. Beyond this basic finding, the observed SNR reduction in the BOLD signal during training, which reflects a reduction of averaged neuronal activity, is suggestive of a shift towards sparser neural representations of objects with increasing familiarity.
Meeting abstract presented at VSS 2014