Abstract
Are there general neural mechanisms that perform visual transformations across all objects, or are transformations highly specific to individual objects? For example, in watching two objects become larger, we can appreciate that both objects are undergoing the same kind of transformation, but is the underlying neural transformation the same or different? We developed a novel computational approach to reveal visual transformations in visual cortex. We hypothesized that it should be possible to compute a mathematical transformation matrix that links object representations before and after physical change. If this represents a general mechanism, then when applied to a new object, the transformation matrix should yield a predicted representation that is consistent with the actual object representation after the physical change. We scanned 14 participants who viewed three different objects undergoing five different physical changes (enlargement, reduction, left-rotation, right-rotation, no-change). We measured activity patterns evoked by the objects in lateral occipital cortex before and after the physical change. Using regularized linear regression, we then predicted the response in each voxel in the post-change pattern from the response in all voxels in the initial pre-change pattern. Computing all regressions on training runs yielded a transformation matrix that was applied to the pre-change pattern in a test run of either 1) an unseen instance of the same object or 2) a new object. We compared the predicted post-change patterns to the true post-change patterns. We found that these transformation matrices produced post-change patterns that matched the true post-change patterns for unseen instances of the same object (p< 0.001). Critically, these transformation matrices generalized to new objects undergoing the same physical change (p< 0.001), suggesting that these transformations are abstract to some extent. These results demonstrate neural evidence for visual routines and show that such routines can be measured with fMRI.
Meeting abstract presented at VSS 2016