The CSAIL-led method outperformed baseline methods describing individual neurons in a variety of vision models & a new dataset of synthetic neurons w/known ground-truth descriptions.

→ View original post on X — @mit_csail,