Abstract
Multivariate pattern classification can be used to decode the orientation of a viewed grating from fMRI signals in the human visual cortex. We have speculated that this orientation signal results in part from the fine-scale columnar structure of orientation-tuned cells. In addition, it has been known since the earliest demonstration of orientation decoding that other, more coarse-scale orientation signals might also contribute to successful decoding, such as the retinotopic bias for orientations radiating outward from the fovea found in the retina, LGN, and V1. More recently, Freeman, Brouwer, Heeger and Merriam (2011) made the stronger claim that this radial bias is completely necessary for orientation decoding. They claimed that if this bias is mathematically removed, then orientation can no longer be decoded. However, their study relied on a temporal phase-encoding procedure in which a grating was rotated through the orientation domain over time. With this design, voxel responses depend on both orientation selectivity and on temporal characteristics of the hemodynamic response, and we show that conflating these sources may lead to erroneous conclusions. We successfully replicate Freeman et al.'s findings with a fast 24-s rotation period, but find persisting orientation signals with a slower rotation period that minimizes temporal blurring of the BOLD response. In addition, we show that simply blurring the edges of the wedge stimulus used to map the radial bias reduces temporal artifacts, and again reveals a persistence of orientation signals even at the faster rotation period. Our results imply that Freeman et al.'s conclusions were in fact due to the confounding of orientation signals and hemodynamic factors, and consequently depended on very particular stimulus conditions. When we take this confound into consideration, we find robust evidence of orientation information in fMRI signals in the human visual cortex that do not depend on the radial bias.
Meeting abstract presented at VSS 2014