Together, our results suggest that a generic “expertise” system spanning categories is computationally implausible.

For related findings supporting this conclusion, see also this recent preprint by Galit Yovel's lab: https://www.biorxiv.org/content/10.1101/2022.12.01.518342v1

Next we used the dual-task CNN we previously trained on both face discrimination and object categorization, which spontaneously segregated itself into two sets of filters, one more causally engaged in face recognition, and the other more causally engaged in object recognition (https://www.science.org/doi/full/10.1126/sciadv.abl8913).

Using lesioning analyses, we again find that the car task relies more on generic object than specific face filters.

Here we stepped back to ask whether the “expertise” hypothesis make sense computationally in the first place. In particular, would we really expect expertise for a new visual category like cars to make use of the same neurons used in face discrimination?

To find out, we trained one CNN on object categorization and another on face discrimination, and then asked how well each network performed on fine-grained car discrimination.