MAIA is built from a pre-trained vision-language model equipped w/a set of commonly used vision interpretability tools for synthesizing and editing inputs, computing maximally activating exemplars from existing datasets, and summarizing & describing experimental results.

→ View original post on X — @mit_csail,