Extended Data Fig. 2: Encoding models training and testing.

a, Encoding models training. For each participant and area, we trained end-to-end encoding models that take images as input and predict the corresponding fMRI responses, using the single-trial NSD responses for the 9,000 participant-unique images. During training, the model predictions were compared to the single-trial target fMRI responses, and the resulting error was backpropagated to update the encoding model weights. b, We tested the encoding models on an independent portion of the NSD data not used for training, consisting of fMRI responses for 515 images seen three times by all participants, averaged across the three trials. Photos from the COCO image dataset/Flickr⁸⁵. c, We compared the noise of the in silico fMRI responses with the noise of the in vivo fMRI responses from the NSD, by comparing how much variance these two data types explained for a third, independent split of the in vivo NSD responses. Because the in silico fMRI responses did not capture all signal variance in the NSD responses, the in silico fMRI responses explaining more variance than the in vivo NSD responses would be indicative of the former being less affected by noise^19,29. We carried out the comparison through three sets of predictions, using the in silico and the in vivo NSD fMRI responses for the 515 test images. Each prediction involved explaining in vivo single response trials from the NSD experiment with a different predictor. In the first set of predictions, the predictor consisted of one of the two remaining in vivo NSD experiment response trials. In the second set of predictions, the predictor consisted of the average of the two remaining in vivo NSD experiment response trials. In the third set of predictions, the predictor consisted of the in silico responses from the trained encoding models. d, Single NSD response trials noise-ceiling-normalized explained variance, for the three predictors of the noise analysis. The variance explained by the in silico responses is higher than the variance explained by both single and averaged NSD trials, indicating that the in silico fMRI responses are less affected by noise compared to the NSD responses. Colored asterisks indicate significant difference between the noise-ceiling-normalized explained variance scores of two predictors (within-participant permutation test, one-sided, P < 0.05, Benjamini/Hochberg corrected over 2 tests for each area; population prevalence test, one-sided, P < 10⁻¹⁰, indicating within-participant significance in all 8 participants), for each area. Error bars reflect 95% confidence intervals.