Automatic Field Guides: Sparse Autoencoders and Heuristic Ranking for Dichotomous Key Learning

A fork of saev, modified for dichotomous key recovery and discovery in images of Heliconius butterfly mimics.

Using this Repository

All scripts for training SAEs, generating sparse features, scoring features for selectivity and W1 distance, and viewing dichotomous key candidates are located in contrib/mimics/exps/001-features/

First, generate ViT patch embedding shards using generate_dino_shards.sh. Then, train SAEs using train_dino_saes.sh and generate sparse features using inference_dino.sh. Finally, score features by running the following command:

uv run contrib/mimics/launch.py score

Visualization can be done using uv run marimo edit contrib/mimics/exps/001-features/dino_viewer.py.

Citation

If you want to cite the software, please use the "Cite this repository" link on the GitHub page, which will provide you with the appropriate citation format.