I work on vision transformers: retrieval under visual abstraction, and linear probes for content and style on frozen CLS tokens.
Research
-
Poster writeup: DAS retrieval, CLS probes for content and style, and incorrect-neighbor examples.
-
Things100 × ArtBench: a linear probe on frozen CLS improves style retrieval versus raw late-CLS cosine kNN.
-
Diffusion Abstraction Spectrum (DAS) on Toronto Scenes: retrieval from line drawings to photographs across ViT-B models.