Fig. 5. Linear ensemble models capture individual regional preferences in NeuroGen framework.

a The NeuroGen framework concatenates a image generator (BigGAN) with an encoding model to synthesize images that achieve a desired predicted response in one or more brain regions. One synthetic image is generated from a 1000 dimension, one-hot encoded class vector (corresponding to one class in ImageNet) and a noise vector, which will be identified through optimization Different synthetic images may have different class vectors. The output image is then fed into the encoding model to obtain a region’s predicted response for that image. b–d Synthetic images from the version of NeuroGen using the linear ensemble encoding model; the 10 images were designed to maximize activation for each of the six NeuroGen individuals (one individual per row) in OFA, FFA1 and FFA2 regions, respectively. e A scatter plot indicating for each of the 3 face regions for each of the six NeuroGen individuals, the observed t-statistic of OFA, FFA1, and FFA2’s responses to images of animal versus human faces (x-axis) and the ratio of the number of animal images minus the number of human images, divided by the sum of these two numbers, out of their top 10 synthetic images created via NeuroGen (y-axis). Each color represents a subject and each shape represents a different face region.