Future Meets Past

Ithaca, an AI-driven deep neural network tool from DeepMind, restores lost text in ancient inscriptions, identifies their place of origin, and date

It is always a challenge for historians and archaeologists to reconstruct the past. While artifacts like potteries, weapons, and similar objects are indirect sources, documents are their primary source. However, surviving ancient documentation are almost always fragmented and crumbling. And they are often inscribed on terracotta, stone, or metal surfaces – which, being inorganic materials, cannot be put through radiocarbon tests to determine the age.

A three-in-one solution

But an artificial intelligence (AI)-driven solution has come forward to the rescue. Google DeepMind (now a subsidiary of the Google parent company Alphabet) has developed a new AI-powered deep neural network tool that can suggest the missing textual portions of damaged ancient inscriptions, identify their original location, and help establish the date they were created. Named “Ithaca”, it is the first such AI model that can perform all these three core tasks central to epigraphy (the study of ancient inscriptions). In a multi-author paper published in Nature, the development team announced that Ithaca can reconstruct missing text with an accuracy level up to 62–72%, depending on data input.

The original paper describes the solution as: “…a deep neural network for the textual restoration, geographical attribution and chronological attribution of ancient Greek inscriptions. Ithaca is designed to assist and expand the historian’s workflow. The architecture of Ithaca focuses on collaboration, decision support and interpretability.”

AI to the rescue

It is typical to use predictive AI modelling for restoring missing data. This is possible because deep neural models can be easily trained to identify very complex patterns by feeding them a huge body of similar data to analyse. For example, OpenAI’s GPT-3 can effectively suggest missing words in a sentence or missing sentences in a paragraph. AI-based image processing tools are also routinely being used to restore damaged films and pictures.

In all such cases, the essential technique remains the same – to intelligently predict what could probably have been lost from the original by analysing what is still left of it. In an email interview to Lifewire, the CEO of AI developer Singulos Research, Brad Quinton explained the advantages:

“… AI can look through a large number of ‘known good’ examples to find patterns between, for instance, a given text and its date and location of creation… …Often, these patterns are so complex that they would not be obvious to a human expert… …Conceptually, researchers could use similar techniques to determine the date and origin of art or tools, or other historical man-made artifacts where there is an expectation of change in the underlying style and technique over time and by location of origin.”

From Pythia to Ithaca

And that is just what Ithaca has done – but with far more accuracy for ancient languages than ever achieved before. Named after the Greek island described in Homer’s Odyssey as the kingdom of Odysseus, Ithaca takes forward the work started by “Pythia” – the previous machine learning tool developed by the same team in 2019. That was the first model to use deep neural networks to restore ancient text.

However, Ithaca has raised the bar by several notches when compared to its predecessor. While Ithaca performs three key tasks, its accuracy levels are way ahead of Pythia. Historian Thea Sommerschield, one of the authors of the paper and key collaborator on the project, sums it up succinctly as reported in the media: “Not only does it advance the previous state-of-the-art set by Pythia, but it also uses deep learning for geographical and chronological attribution for the very first time and on an unprecedented scale.”

Flexible and accurate

The following pointers would explain why Ithaca is such an important development:

• The program has been trained on a dataset of around 78,000 ancient Greek inscriptions to look for patterns that help generate the targeted output.
• The final tool was tested using over 60,000 well-studied ancient Greek texts, dated between 700 B.C. and 500 C.E. Portions of the texts were randomly obscured to evaluate how accurately Ithaca predicts the missing parts.
• The findings reveal that without any external assistance, the tool could: (i) restore letters in damaged texts with 62% accuracy; (ii) determine an inscription’s geographic origins among 84 regions of the ancient world with 71% percent accuracy; and (iii) date texts to within 30 years of their known year of writing.
• Scholars and students working without the tool could restore texts with only 25 percent accuracy. But when assisted by Ithaca, the historical accuracy of their own work went up to 72%.
• Although trained on Greek inscriptions, Ithaca is flexible and can be configured to deal with any number of ancient writing forms, even Mayan and cuneiform, and any medium of writing.
• Data visualisation features built in the tool make analytical interpretations easy for researchers.
• The software is free and available online, along with its open-source code, for anyone to use.

A machine-human collaborative tool

Ithaca has already been used to date inscriptions from the Acropolis at Athens. However, refinements are always possible, and some experts are pointing out that a 62% accuracy means, it is still inaccurate about one-third of the time.

Replacing specialized academics with the program is not the target yet. But it can be an excellent support tool, especially because as Sommerschield says “… it’s really difficult for a human to harness all existing, relevant data and to discover underlying patterns.”

That’s the exact area where machine learning comes in handy, and the tool does open up an avenue for machine-human collaboration. And that is what the authors of the paper hope:

“This research shows how models such as Ithaca can unlock the cooperative potential between artificial intelligence and historians, transformationally impacting the way that we study and write about one of the most important periods in human history.”

Read the full paper at: https://www.nature.com/articles/d41586-022-00702-6

Anyone interested can access Ithaca at: https://ithaca.deepmind.com/