Interpretable Visual Feature Discovery using Multiple Instance Learning: A Case Study of Colonial Korean Print

Zenodo (CERN European Organization for Nuclear Research)

Abstract

Reproducibility bundle for the article Interpretable Visual Feature Discovery using Multiple Instance Learning: A Case Study of Colonial Korean Print, forthcoming in Computational Humanities Research (Cambridge University Press, ISSN 2977-8158). Cite the article Van de Pol, A., Prokic, J., & Mol, A. (forthcoming). Interpretable Visual Feature Discovery using Multiple Instance Learning: A Case Study of Colonial Korean Print. Computational Humanities Research, X(X). Cambridge University Press. Volume, issue, and page range will be added once the article is in print. The deposit contains: Source code archive (code-v1.0.0.tar.gz) with training, evaluation, and interpretability pipelines. Trained AttriMIL Flash model weights for both 16x16 and 8x8 patch configurations (5-fold each, optimizer state stripped). Trained ConvNeXtV2 Base and Swin S3 Base 224 baseline weights under both held-out-book GroupKFold and page-random KFold protocols (5-fold each). Page-level metadata for the 57,583-page corpus, with canonical printshop labels and source catalog links. Fold definitions for both evaluation protocols, including the page-to-book mappings used for the GroupKFold splits. Per-fold evaluation results for every released model. Interpretability artifacts: 8x8 patch embeddings with attention weights, UMAP and t-SNE coordinates, HDBSCAN cluster assignments, and a compact cluster-assignment artifact. Documentation: dataset card, manifest with SHA-256 checksums, and a deposit-level README. The page images themselves are not redistributed and remain governed by the Hyundam Mun'go collection terms; per-page URLs are recorded in the dataset metadata. Licensing: The source code archive is released under the MIT License. All other artifacts (dataset metadata, fold definitions, model weights, evaluation results, interpretability artifacts) are released under CC-BY-4.0.

Authors 3

1. Aron Van de Pol Aachen

   Leiden University

   Affiliation as printed

   Leiden University

2. Jelena Prokić Aachen

   Leiden University

   Affiliation as printed

   Leiden University

3. Angus A. A. Mol Aachen

   Leiden University

   Affiliation as printed

   Leiden University