Figure 8. DNN layer and DNN training comparisons in terms of correlation with fMRI responses in scene-selective cortex.
Panels show convolutional and fully-connected (FC) layer-by-layer RDM correlations between (A) an object-trained (ReferenceNet) and a scene-trained (Places) DNN; (B) both DNNs and the a priori selected feature models; (C) the object-trained DNN and scene-selective ROIs; (D) the scene-trained DNN and scene-selective ROIs (all comparisons FDR-corrected within ROI; See legend of Figure 2B for explanation of the statistical indicators and noise ceiling). While the decreasing correlation between DNNs indicates stronger task-specificity of higher DNN layers, the original fc7 DNN feature model correlated most strongly with high-level layers of both DNNs. The object-trained and the scene-trained DNN correlated similarly with PPA and OPA, with both showing remarkable good performance for mid-level layers. The RDMs for each individual DNN layer are provided in Figure 1—source data 1. Searchlight maps for each layer of the object- and scene-trained DNN are provided in Figure 8—video 1 and Figure 8—video 2, respectively.
