Ammon News - Google DeepMind's Aeneas AI model helps historians decode damaged Latin inscriptions, outperforming experts in dating accuracy and predictive text

Google DeepMind has developed an AI model that can predict where ancient Latin texts originate, estimate their age and restore missing sections.

The model is called Aeneas and it represents a significant advancement in the field of epigraphy — the study of ancient inscriptions.

Developed by a team of researchers from universities in the UK and Greece, as well as a team from DeepMind itself, Aeneas addresses one of the most longstanding challenges in the fields of history, archaeology and anthropology.

"Aeneas can retrieve relevant parallels from across our entire data set instantly" because each text has a unique identifier in the database, says Yannis Assael, a research scientist at Google DeepMind.

Training on vast historical datasets
Aeneas was trained on inscriptions from three of the world's largest databases of Latin epigraphy.

The combined dataset contained text from 176,861 inscriptions — plus images of 5% of them — with dates ranging across a millennium from the 7th century BC to the 8th century AD.

The model comprises three neural networks, each designed for different tasks: restoring missing text, predicting geographical origin and estimating age.

Along with results, Aeneas provides a ranked list of similar inscriptions from the dataset to support its conclusions.

Outperforming human experts in accuracy tests
The team tested Aeneas by asking 23 epigraphers to restore text removed from inscriptions and to date and identify origins both independently and with the model's assistance.

Working alone, experts dated inscriptions to within around 31 years of the correct answer.

Dates predicted by Aeneas were accurate to within 13 years — significantly outperforming human specialists.

When specialists had access to the model's predictions and similar inscription lists, they achieved greater accuracy in geographical identification and text restoration than either working alone or relying solely on the model. Technology Magazine