Insightful analysis of historical sources at scales beyond human capabilities using unsupervised Machine Learning and XAI

Oliver Eberle, Jochen Büttner, Hassan El-Hajj, Grégoire Montavon, Klaus-Robert Müller, Matteo Valleriani

Abstract

This paper presents a novel machine learning framework designed to analyze the Sacrobosco Collection, a massive corpus of early modern astronomy textbooks. By moving beyond traditional "close reading," the authors utilize an unsupervised "atomization-recomposition" approach to identify patterns in the circulation of scientific knowledge. The study reveals how institutional competition and book marketing drove the homogenization of science across Europe between 1472 and 1650, providing a "macroscope" for historians to observe scientific evolution at scale.

TL;DR

Researchers developed an "atomization-recomposition" machine learning framework to analyze over 76,000 pages of historical astronomy textbooks. By breaking complex tables down into numerical "atoms" and rebuilding them into semantic histograms, they uncovered hidden patterns of how scientific knowledge was homogenized and spread across Europe, driven by both institutional competition and early-modern book marketing.

Problem Statement

Historians of science often struggle with a "scale vs. depth" trade-off. While significant research exists regarding prominent figures like Copernicus, much less is known about the "core knowledge" taught to the average university student. The sheer volume of material—tens of thousands of pages of dense, numerical tables—makes traditional manual analysis impossible.

Standard Machine Learning (ML) and OCR (Optical Recognition) often fail in this context because historical documents are highly heterogeneous, featuring non-standard fonts, varying print qualities, and complex layouts. Furthermore, training deep learning models usually requires massive amounts of labeled data, which is rarely available for niche historical corpora.

Methodology

The authors introduce a novel atomization-recomposition framework designed for unsupervised settings with sparse annotations. The process, illustrated in Figure 1, follows four key stages:

  1. Atomization: Instead of recognizing entire tables, a robust digit recognizer (0-9) is trained using just a few hundred labeled patches. The model utilizes Equivariant Steerable Pyramids, a type of CNN invariant to the rotations and scales common in hand-scanned historical books.
  2. Recomposition: The model identifies "bigrams" (sequences of two digits) and their spatial relationships to capture mathematical content rather than just visual layout.
  3. Bag-of-Bigrams: Each page is represented as a histogram of these bigrams, creating a layout-invariant "fingerprint" for the table.
  4. Explainable AI (XAI): Using BiLRP (Bidirectional Layer-wise Relevance Propagation), historians can verify the model's decisions by seeing exactly which pixels and number combinations drove similarity assessments.
The atomization-recomposition pipeline, from book dismantling to machine-readable histograms.

Datasets

  • The Sacrobosco Collection: 359 early modern printed editions of astronomy textbooks (roughly 76,000 pages).
  • Sacrobosco Tables Dataset: A subset of 9,793 pages identified as containing numerical tables.
  • Ground Truth Sets: Small, expert-annotated datasets used for validation, including "Sun zodiac" and "climate zone" tables.

Key Insights and Ideas

The framework enables multi-faceted analysis of the corpus, revealing both temporal shifts and geographical variations in scientific content (see Figure 4).

  • Knowledge Homogenization: Scientific knowledge became increasingly uniform over time, fueled by intense competition between institutions like the Jesuit Collegio Romano and the French Royal Chair of Mathematics.
  • Marketing as a Driver of Science: A sharp "entropy drop" (a surge in identical tables) occurred around 1550. This was driven by printers exploiting privilegium (early copyright) to flood the market with re-issues before their exclusive rights expired.
  • The 9-Zone Variant: While tradition recognized 7 climate zones, the model identified a specific 9-zone variant. This was a "Wittenberg innovation" used by Protestant reformers to include northern latitudes in the "habitable" world, demonstrating how science was adapted for local identity.
Validation metrics and corpus-level analysis showing temporal entropy shifts and geographical distributions.

Key Results

  • High Fidelity: The model achieved Pearson correlation scores between 0.84 and 0.93 when compared to human-annotated ground truth, proving its accuracy in reading numerical content.
  • Superior Retrieval: The bigram-based representation significantly outperformed standard baselines (like VGG-16) in identifying semantically identical tables across different editions and printers.
  • Discovery at Scale: The framework successfully clustered 10,000 pages of tables, allowing historians to identify "families" of knowledge that were previously invisible.

Limitations and Future Work

The current model is specialized for numerical tables. While the framework can theoretically be adapted for visual or textual elements, this research focuses primarily on the "mathematization" of history. Additionally, challenges remain regarding high-quality digital scans, specifically with folded pages or severe ink bleed-through. Future work aims to integrate these numerical insights with textual and visual analysis to create a truly multimodal genealogy of historical knowledge.

Summary updated

BibTeX
@inproceedings{eberle2024insightful,
  author = {Oliver Eberle and Jochen Büttner and Hassan El-Hajj and Grégoire Montavon and Klaus-Robert Müller and Matteo Valleriani},
  title = {{Insightful analysis of historical sources at scales beyond human capabilities using unsupervised Machine Learning and XAI}},
  booktitle = {Science},
  year = {2024},
  doi = {10.1126/sciadv.adj1719},
}
Figure