Imaging prior information in the brain

Abstract

In making sense of the visual world, the brain's processing is driven by two factors: the physical information provided by the eyes ("bottom-up" data) and the expectancies driven by past experience ("top-down" influences). We use degraded stimuli to tease apart the effects of bottom-up and top-down processes because they are easier to recognize with prior knowledge of undegraded images. Using machine learning algorithms, we quantify the amount of information that brain regions contain about stimuli as the subject learns the coherent images. Our results show that several distinct regions, including high-level visual areas and the retinotopic cortex, contain more information about degraded stimuli with prior knowledge. Critically, these regions are separate from those that exhibit classical priming, indicating that top-down influences are more than feature-based attention. Together, our results show how the neural processing of complex imagery is rapidly influenced by fleeting experiences.

Fig. 1.

Overall methodology. Before each fMRI run, two images—one containing a natural object and the other an artificial object—were shown. During the run, subjects saw Mooney images of four different objects—the two primed images and two novel images—and were asked to identify the contents of each image as either natural or artificial. For the next set of images, undegraded versions of the two previously unprimed images were used to prime the subject, and two additional novel images were introduced. In this manner, every image (except the first primed and last unprimed images) was shown to the subject in both primed and unprimed conditions.

Fig. 2.

Behavioral performance recognizing the primed and unprimed Mooney images during the fMRI scans. (A) Across all images, subjects performed significantly better for primed images in a 2-AFC task in which they were asked to indicate whether the image was natural or artificial. Shown is the expected value and 95% confidence intervals for the proportion that was correct over all images and subjects; P value was calculated from a one-tailed–paired McNemar test corrected for multiple comparisons over all subjects and image sets. (B) When images were selected to show a per-image priming effect (resulting in discarding 10 of 38 images), subjects were at chance at recognizing the unprimed images.

Fig. 3.

Difference between SVM and GLM results. (A) Group average response showing primed voxels, which show greater activation for primed versus unprimed images in a GLM analysis, here illustrated in red and yellow. (B) When these voxels are used as a region of interest for the SVM analysis, no increase in information decoding is seen, and SVM performance drops to chance. Note that, although we show the group average response here in A, each ROI was determined from a subject's individual data.

Fig. 4.

Object priming increases the information found in many regions of the cortex. (A) Using atlas-based anatomical regions of interest for the SVM analysis, several notable regions, including the pericalcarine cortex, inferior parietal cortex, lateral occipital cortex, and fusiform gyrus, show significantly increased decoding accuracy for primed versus unprimed images. No areas showed higher decoding accuracy for unprimed images. The magnitude of the color scale indicates the level of significance in log power (e.g., both +2 and −2 indicate P < 10⁻²), whereas the sign of the scale indicates whether primed images yielded higher accuracy (positive scale) or vice versa. Colored lines indicate the borders of anatomically based ROIs. (B) Example of the results of the decoding accuracy in a few notable cortical regions. Pericalcarine regions showed relatively high decoding accuracy and were significantly higher in the primed condition. Classification accuracy was lower overall in the fusiform gyrus, but significantly elevated in the primed condition in the right hemisphere only. Dashed lines indicate chance performance levels over the entire group of subjects.