Fig. 4: Strong tests of hypothesized selectivity using high-throughput screening and GAN-based image synthesis.
From: Computational models of category-selective brain regions enable high-throughput tests of selectivity
a Schematic of the high-throughput image screening procedure. Here we screened ~3.5 million stimuli from several public image databases through the encoding models and investigated the stimuli that the models predict strongly activate a given fROI. b Histograms showing the distribution of predicted responses by the encoding models for the different regions of interest. Each color indicates a different stimulus database for the predictions. Source data are provided as a Source Data file. c Representative images predicted to strongly activate each of the fROIs. See Figs. S7-S11 for stimuli subsampled from the top 100,000 images. Note that the actual images have been replaced with copyright-free illustrative versions here. See https://osf.io/5k4ds/ for the actual top images. d Schematic for the GAN-based image synthesis procedure. We coupled a Generative Adversarial Network (BigGAN) as the prior along with our Resnet50 (the encoder) to optimize pixels and synthesize new stimuli that predict yet stronger responses in each of the desired fROIs. e Stimuli synthesized from this GAN-based synthesis procedure that the models predict maximally activate the FFA, EBA, and the PPA.
