Learning to Associate Orientation with Color in Early Visual Areas by Associative Decoded fMRI Neurofeedback

Abstract

Associative learning is an essential brain process where the contingency of different items increases after training. Associative learning has been found to occur in many brain regions [1-4]. However, there is no clear evidence that associative learning of visual features occurs in early visual areas, although a number of studies have indicated that learning of a single visual feature (perceptual learning) involves early visual areas [5-8]. Here, via decoded fMRI neurofeedback termed "DecNef" [9], we tested whether associative learning of orientation and color can be created in early visual areas. During 3 days of training, DecNef induced fMRI signal patterns that corresponded to a specific target color (red) mostly in early visual areas while a vertical achromatic grating was physically presented to participants. As a result, participants came to perceive "red" significantly more frequently than "green" in an achromatic vertical grating. This effect was also observed 3-5 months after the training. These results suggest that long-term associative learning of two different visual features such as orientation and color was created, most likely in early visual areas. This newly extended technique that induces associative learning is called "A-DecNef," and it may be used as an important tool for understanding and modifying brain functions because associations are fundamental and ubiquitous functions in the brain.

Figure 1. Procedure of a trial in the A-DecNef training stage

An achromatic (gray-black) vertical grating was presented during the induction period (6 sec), during which participants were trained to induce BOLD-signal multi-voxel patterns in V1/V2 corresponding to the color red, unbeknownst to participants. Immediately after the fixation period (7 sec), a feedback disk was presented for 1 sec. The disk size roughly represented how similar the BOLD-signal patterns in V1/V2 induced during the induction period were to the patterns evoked by the target color stimuli (red-black gratings) presented in the color classifier construction stage. See also Figure S1.

Figure 2. Change in the red likelihood by A-DecNef training (N = 12)

On Day0 the color classifier was constructed, whereas on Day1, 2, and 3 the A-DecNef training was conducted. Gray dots represent individual data and red dots represent the average across participants. The red likelihoods on Day1, 2, and 3 were significantly higher than on Day0 (P<0.001 for all three days and the average; t₁₁=5.64 for Day 1; t₁₁=7.25 for Day 2; t₁₁=7.07 for Day 3; t₁₁=7.42 for the average across the 3 days; paired t-test with Bonferroni correction). The average value is for Days 1-3.

Figure 3. Stimuli and results of the post-test stage

(A) Stimuli. The color in the inner gratings varied in 8 steps between a reddish tint (x=0.323, y=0.310, Y=17.9) and greenish tint (x=0.313, y=0.323, Y=17.9), passing through a neutral gray, while the outer grating was kept achromatic. (B) Mean (±SEM) chromatic psychometric functions for the vertical (black circles), oblique (white circles), and horizontal (black triangles) gratings for the A-DecNef group (N=12). (C) Mean (±SEM) chromatic psychometric functions for the vertical (black circles), oblique (white circles), and horizontal (black triangles) gratings for the control group (N=6). (D, E) Individual and mean (±SEM) red response% for the vertical and horizontal gratings at the neutral gray for the A-DecNef group (D) and for the control group (E). See also Figures S2 and S4.

Figure 4. Accuracy of the color classifier and predictability of the red likelihood in V1/V2 using a searchlight analysis [14]. Color coded voxels correspond to P<0.05

(A) Distribution map of the color classifier accuracy during the color classifier construction stage. The accuracy was computed by moving a sphere ROI across the whole brain. The classifier has the ability to extract information for the red likelihood in both V1/V2 and ventral areas including V4. The color scale bar indicates the accuracy (%). (B) Distribution map of the predictability of the red likelihood of V1/V2 during the color classifier construction stage. The predictability was the highest in V1/V2 and to a moderate degree in ventral areas including V4. The color scale bar indicates a coefficient of determination between the V1/V2 red likelihood predicted by BOLD-signal patterns within an ROI and the actual V1/V2 red likelihood (based on the V1/V2 BOLD-signal patterns) in the searchlight analysis. See Supplemental Experimental Procedures for technical details. (C) Distribution map of the predictability of the red likelihood of V1/V2 during the A-DecNef training stage. Not much significant predictability was found outside V1/V2. The color scale bar indicates a coefficient of determination, as in (B). See Supplemental Experimental Procedures for technical details. See also Figure S3.