Figure 6:

Emergence of abstract category structure over the hierarchy of layers, comparing similarity structure in each layer vs that present in V1 (black line) or in TE (red line). Both cases, which are roughly symmetric, clearly show that IT layers (TEO, TE) progressively differentiate from raw input similarity structure present in V1, and, critically, that the model has learned structure beyond that present in the input. This is is the simplest, most objective summary statistic showing this progressive emergence of structure, while subsequent figures provide a more concrete sense of what kinds of representations actually developed.