PLACE: Patch-Level Agnostic Concept Extraction

Code for the NeurIPS 2026 paper PLACE: Patch-Level Agnostic Concept Extraction. PLACE is a label-free method for extracting concepts from the patch tokens of Vision Transformers and foundation models. It produces concept heatmaps without needing a classifier.

The method, results and paper are on the project page: https://gabrieleonorato.com/PLACE

Notebooks

• concept_extraction_demo.ipynb: a short demo on the ImageNet church class with ViT-B/16, DINOv2 and ViT-MAE.
• task_agnostic_experiments.ipynb: reproduces the task-agnostic experiments, where one concept bank is learned from many classes.

Setup

git clone https://github.com/Restuccia-Group/PLACE.git
cd PLACE
uv sync
uv run jupyter lab

Or use pip install -r requirements.txt.

The data, classifier heads and saved concept banks are too large for git. Download task_agnostic_assets.zip from Zenodo and extract it in the repository root:

unzip task_agnostic_assets.zip -d .

Citation

@inproceedings{onorato2026place,
  title     = {{PLACE}: Patch-Level Agnostic Concept Extraction},
  author    = {Onorato, Gabriele and Silvestri, Fabrizio and Bastian, Nathaniel D. and Restuccia, Francesco},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  url       = {https://openreview.net/forum?id=8QlhKxCYSB}
}

License

CC BY-NC-ND 4.0

Acknowledgements

This work was supported in part by the U.S. National Science Foundation (NSF) under Grants CNS-2312875 and OAC-2530896, the Defense Advanced Research Projects Agency (DARPA) under Cooperative Agreement D25AC00374-00, the Office of Naval Research (ONR) under Award N00014-23-1-2221, and the Air Force Office of Scientific Research (AFOSR) under Award FA9550-23-1-0261. Special thanks to the authors of CRAFT, FACE and RSVC for releasing their code.