Abstract
The contributions of visual cortex neurons to feature binding are not well understood. A central issue is distinguishing models of feature binding in which separate feature-coding neurons are linked together from models in which binding is done by neurons that explicitly encode feature conjunctions. Here, we measured brain responses with fMRI as subjects viewed an annulus containing colored, moving dots and detected the presence of target dots defined either by a single feature (color, or direction of motion), or by a conjunction of those features. Each trial consisted of a 500 millisecond stimulus presentation, followed by a one second response period. Data were acquired in a blocked design; the blocks contained 16 trials and were separated by 25 seconds of rest. Patterns of activity during the three conditions could be distinguished with machine learning algorithms, as early as V1. Analysis of the voxel-based weights output by the classification algorithm revealed that voxels informative for classification of the conjunction task were less informative for the feature tasks, and vice versa. In other words, when subjects searched for a feature conjunction (e.g. red and down), informative cortical locations were different than when subjects searched for the same color (red) or motion direction (down) separately. This result suggests that separate populations of neurons were activated during binding relative to feature detection. A whole-brain analysis revealed involvement of parietal cortex in binding, even though no task explicitly required shifts of spatial attention. Overall, these data are consistent with a feature binding mechanism sensitive to conjunction-coding neurons in early visual cortex.
his research was supported by grant 1000274 from the Israeli Binational Science Foundation, and by NIH grants 2RO1MH058383-04A1 and 1RO1MH062331.