Students Win $700K for Using AI to Decipher Ancient Roman Scroll
An unopened scroll that was buried in volcanic mud and ash during the eruption of Mount Vesuvius nearly 2,000 years ago has finally been decoded by a team of students who used machine-learning algorithms to reveal the contents of the ancient text.
The winners of the international Vesuvius Challenge — Youssef Nader, Luke Farritor, and Julian Schilliger — were awarded a grand prize of $700,000 this week for using artificial intelligence to decipher portions of the scroll, according to an announcement on Monday, February 5. Notably, the challenge’s Grand Prize winners had each previously been involved in the international competition since it began in 2023. Farritor and Nader each received awards for independently deciphering the Ancient Greek word porphyras, meaning “purple,” and Schilliger won three prizes for working on a tool known as the Volume Cartographer that further enabled the three-dimensional mapping of the scroll’s papyrus. For the grand prize, the three students teamed up, divvying up the work to ultimately crack the scroll’s code.
To date, researchers have managed to read approximately five percent of the scroll’s text, but aim to decode 90% by the end of 2024, as well as uncover more text from the other 800 scrolls that were enclosed in volcanic debris.
But roughly transcribed snippets already reveal that the first scroll’s author wrote about the pleasure derived from the accessibility of goods like food — such as the taste of capers — and music. In Epicurean philosophy, pleasure was viewed as the ultimate goal of all actions and holding intrinsic value. At one point, the scroll’s author even appeared to throw shade at his competitors, characterizing them as people who “have nothing to say about pleasure, either in general or in particular, when it is a question of definition.”
“My primary focus was ink detection,” Nader told Hyperallergic, explaining that he employed a technique called “supervised learning” which consists of teaching an AI model how to detect ink based on a series of examples. The model used cross-sections of images to detect repeating patterns and identify letters.
“[The AI] is sort of zooming in on the papyrus paper, trying to look inside into the depth of the paper,” Nader said. “By doing this, it learns how to separate what ink looks like and what papyrus looks like and where to find this ink.”
Nader noted that the greatest challenge was trying to grow the data set, as the team “virtually started from nothing.”
“We didn’t have any ink inside of the scroll,” Nader said. Another issue was determining which AI model to use before settling on a TimeSformer-based model — the same system he used to decipher the first word from the scroll’s text, earning him second place behind Farritor for the First Letters challenge.
Nader said he is interested in continuing to use AI to further archaeological research of ancient Egyptian papyrus — an inherently interesting subject to him, given his Egyptian heritage. He also said that he is excited by the potential for AI and how discoveries such as this can shape the public’s understanding of the novel technology.
“Hopefully we continue changing people’s minds, one scroll at a time,” Nader said.