The authors examined dozens of pretrained vision transformers spanning different architectures, scales and objectives.
Patch-token decoding does not reliably explain vision transformer decisions
Across pretrained models, controlled experiments and a theoretical example, identical classification ability can coexist with aligned, inverted or semantically uninformative explanations.
Top University
Ronen Schaffer · Hila Manor · Tomer Michaeli
Technion – Israel Institute of Technology
Research Digest··2 min read
Schaffer, Manor and Michaeli test whether decoding individual vision transformer patch tokens reliably reveals where a model represents a class.
§