Fig. 1. Encoding model and ensemble model architecture.

a The encoding model architecture. A feature extractor adopted from ResNet-50⁸ extracts features from the input image and a linear readout maps the extracted features to the responses for a specific brain region. b The ensemble model architecture. A group of n pretrained encoding models (where n = 7 or 8) are used to obtain a set of predicted activations, which are combined via a linear model fitted via ordinary least squares to optimally predict the query subject’s regional brain response.