Aeneas: AI Model Revolutionizes Historical Analysis of Ancient Inscriptions

Aeneas, a groundbreaking AI model from DeepMind, is transforming how historians interpret, attribute, and restore ancient inscriptions. This new tool, introduced in a recent paper published in Nature, accelerates the complex process of identifying parallels among fragmentary texts, providing historians with a powerful resource to connect the past.

What Changed Technically?

Aeneas builds upon DeepMind's earlier work with Ithaca, an AI model that focused on restoring, dating, and placing ancient Greek inscriptions. However, Aeneas takes this a step further by:

• Contextual Reasoning: Aeneas can reason across thousands of Latin inscriptions in seconds, retrieving textual and contextual parallels.
• Adaptability: The model is adaptable to other ancient languages, scripts, and media, including papyri and coinage.

Why It Matters to Practitioners

Historians traditionally rely on their expertise and specialized resources to identify "parallels", texts that share similarities in wording, syntax, standardized formulas, or provenance. This process is often time-consuming and requires extensive knowledge. Aeneas significantly speeds up this work by:

• Accelerating Research: Historians can quickly access relevant parallels, allowing them to interpret and build upon the model’s findings.
• Enhanced Accuracy: By leveraging a large dataset, Aeneas improves the accuracy of historical interpretations and restorations.

Technical Details

Aeneas is built on advanced deep learning techniques and leverages a vast corpus of Latin inscriptions. Here are some key technical points:

• Dataset: The model was trained on a comprehensive dataset of Latin inscriptions, curated in collaboration with historians from the University of Nottingham, Warwick, Oxford, and Athens University of Economics and Business (AUEB).
• Architecture:
  • Transformer-based Model: Aeneas uses a transformer architecture, which is well-suited for handling long sequences of text and capturing complex relationships.
  • Contextual Embeddings: The model generates contextual embeddings for inscriptions, allowing it to identify and retrieve relevant parallels efficiently.
• Performance:
  • Speed: Aeneas can process and retrieve parallels in seconds, a significant improvement over manual methods that can take hours or days.
  • Accuracy: Initial benchmarks show high accuracy in identifying textual and contextual parallels, with the model outperforming traditional methods in many cases.

Interactive Version and Open Source

To ensure this research benefits as many people as possible, DeepMind has made an interactive version of Aeneas freely available at predictingthepast.com. This tool is accessible to researchers, students, educators, and museum professionals. Additionally, the code and dataset are open-sourced on GitHub, encouraging further research and development in this field.

Future Directions

Aeneas represents a significant step forward in using AI for historical analysis. The model's adaptability to other ancient languages and media opens up new possibilities for connecting diverse historical evidence. DeepMind is committed to continuing this research and exploring how generative AI can further enhance the work of historians.