Fig. 1: The construction of 25D object space and stimuli for fMRI experiments.

a 200 k natural images from ImageNet, an online database, were passed through AlexNet, a neural network for object recognition. Independent component analysis was performed on the responses of units in the penultimate layer (fc7). 25 independent components (ICs) were extracted to build an object space. b For each IC, the angles between 25D vectors of 200k images and the IC axis were computed, and 50 images with the smallest and the largest angles were selected to represent the positive and negative ends of that IC, respectively. Representative images for IC1 are shown on the right. c Representative images for more ICs. d The difference between the average 25D vectors of the positive and negative representative images was computed for each IC. Besides the 50 most extreme images (image set 1), images with the 50th to the 100th smallest and largest angles were selected as a separate image set (image set 2). The cosine angles between the differential vectors estimated using two image sets and those between the differential vectors and the original IC axes are shown in the upper and lower panels, respectively. e A typical stimulus sequence for the fMRI experiment. In each block, 48 images were randomly selected from the 50 representative images and presented in random order. The dashed lines indicate the blocks used in the following analyses. f Boundaries of the temporal lobe and retinotopic visual areas were identified using retinotopic mapping. The visual stimuli are shown above the inflated surfaces of one monkey. Two different perspectives of the inflated surfaces are shown. The color scale bar indicates the common logarithm of the probability of error. STS, superior temporal sulcus. g The flat map of the left temporal lobe of the same monkey (purple patches in f). Note that due to copyright restrictions, the original ImageNet images in (b, c) are replaced with natural images from Pixabay (https://pixabay.com/).