AI deciphered text on these ancient fossilized scrolls. Here’s what it said.

The Herculaneum Papyri, an ancient scroll fossilized by the Mount Vesuvius eruption in 79 AD, is now decipherable, thanks to an open-source AI project.

Computer science student Luke Farritor “became the first person in two millennia to see an entire word from within an unopened scroll this August,” according to the Vesuvius Challenge, which offers cash prizes for those who can figure out what the scroll says.

This mass of carbonized scroll is now legible.
Credit: Vesuvius Challenge

In addition to the previous discoveries from researchers, these milestones were possible because of the open-source approach. The machine learning techniques contestants used aren’t particularly new. But bringing the “open-source mindset to an academic project” is what enabled such rapid achievements, said JP Prosma, a spokesperson for the contest. “By open sourcing the data and setting up the right incentives for contestants, contestants were able to explore many more ideas than a small team of academics could ever have done in the same timeframe,” Posma continued.

How was AI used to decipher ancient text?

The contest was created to accelerate the research and discovery of the scrolls after University of Kentucky researchers led by Dr. Brent Seales used computer vision to virtually “open” the scrolls in 2015. The volcanic ash from the eruption had carbonized the scrolls, preserving them, but also rendering them too delicate to unroll. Three-dimensional CT scans were able generate a full virtual image of the scrolls in 2019, which then prompted entrepreneurs Nat Friedman and Daniel Gross to fund a competition this year to open-source the research.

Tweet may have been deleted

Enter Farritor, who tapped into previous discoveries of ink “crackle patterns” by another contestant named Casey Handmer. Farritor trained a machine-learning model to learn the patterns, which then became data to improve the model’s recognition abilities. Eventually, Farritor’s model found a formation of letters which made up the word “porphyras,” which means purple in the ancient script.

Farritor was awarded $40,000 for discovering the first word. Another contestant, biorobotics student Youssef Nader, used another method involving an unsupervised pre-training model on the data, then fine-tuning the data on the “fragment labels.” Nader found letters possibly forming the words “achieving” and “similar”; he was awarded $10,000 for his discovery.

Farritor used a machine learning model to discover the word “purple.”
Credit: Vesuvius Challenge

The contents of the scrolls “likely contain texts from Philodemus, an Epicurean philosopher,” said Posma. By using this technique, researchers hope to decipher other carbonized scrolls found at the site, which would “double the number of works we have from ancient literature.”

Now that a major component of the puzzle has been solved, the race is on to decipher the rest. The Vesuvius Challenge will award the grand prize of $700,000 to whoever can read “four passages of text” from the scrolls. Those looking to get involved can get more information on the challenge’s website and Kaggle, and access code on Github. There’s also a lively conversation about the Vesuvius Challenge happening on Discord.