computer king

The ancient scrolls (Scrolls of Herculaneum) from 2000 years ago were finally deciphered by a 21-year-old computer boy.

This ancient scroll can be said to be extremely special——

It was buried and carbonized by a volcanic eruption in 79 AD, and was not unearthed until 1752.

And because the ancient scroll itself is already too fragile, no one has dared to touch it for hundreds of years. If it is not handled properly, it may turn into dust.

Herculaneum and the famous Pompeii are 8 kilometers apart and located at the foot of Mount Vesuvius southeast of Naples, Italy; the former is closer to the crater than Pompeii.

So how to see what is recorded on it has become a big problem that has troubled scientists for hundreds of years.

Now, with the computer boy’s successful deciphering, he has become the first person in 2000 who has never opened the scroll but “saw” the text on it.

This task actually comes from a challenge called Vesuvius Challenge:

Contestants are required to find at least 10 letters within a 4 square centimeter area of ​​the scroll.

As the first person to complete the task, the teenager was officially awarded the US$40,000 First Letters Prize by the event officials.

Many friends must be curious, what are the words deciphered by the boy?

It is the ancient Greek word πορφυρας, which means purple dye or purple clothes.

After the results came out, Nat Friedman, the sponsor of the competition and technology entrepreneur, even touched Musk and proudly introduced the young man:

He is your SpaceX intern!

AI has made great contributions to deciphering thousands of years of ancient scrolls

The boy’s name is Luke Farritor, a computer major.

Before telling the story of this young man, we also need to explain the background of this game.

As early as 2019, Professor Brent Seales of the EduceLab at the University of Kentucky performed imaging work on the Herculaneum scroll in a particle accelerator and generated 3D CT with a resolution of up to 4 μm. scanning.

▲ Professor Hills and his team scanned at the particle accelerator

The professor’s doctoral student, Stephen Parsons, has long been working on using machine learning models to detect ink in CT scans.

So he and his team scanned and photographed fragments of detached scrolls with visible ink, providing a ground-truth collection.

▲ Stephen’s doctoral thesis: training machine learning models on ground-truch data sets

This research successfully attracted the attention of technology entrepreneurs Nat Friedman and Daniel Gross; thus, with their sponsorship, the Vesuvius Challenge was launched race.

They launched an open competition in March this year, with a total prize of US$70,000, with the aim of accelerating the deciphering of ancient scrolls.

In August of this year, a scholar named Casey Handmer wrote a blog about his discovery of a “crackle pattern” that looked like ink.

This can be said to be a very important breakthrough. Even Stephen had only seen direct evidence of ink on separated fragments before, but not on a scroll.

Luke then overheard the news and challenge during a PODCAST, and also saw Casey’s crack pattern being widely discussed on Discord. This aroused great interest in Luke, and he began to use his evening time to train a machine learning model about “crack patterns”.

During the training process, he discovered dozens of ink strokes and some complete letters. Luke marked them and used them as training materials.

▲ Left: Cracked ink visible against a background of papyrus fibers; Right: The resulting binary ink label.

It didn’t take long for the model to reveal traces of cracks that were invisible to the naked eye; these marks became clues to the final letters and actual words.

The picture below is a work submitted by Luke to the challenge. You can vaguely see the shape of the word “ΠΟΡΦΥΡΑϹ” (porphyras).

When Professor Hills and his team saw this picture, they exclaimed:

Although the letters were blurry, they could immediately read the word “porphyras.”

This word is not very common in ancient literature, but it can withstand scrutiny. It probably means “purple”.

One expert explained:

The sequence πορφυ̣ρ̣ας̣ may be πορφύ̣ρ̣ας̣ (noun, purple dye or purple cloth) or πορφυ̣ρ̣ᾶς̣ (adjective, purple).

Due to the lack of context, πορφύ̣ρ̣α ς̣κ or πορφυ̣ρ̣ᾶ ς̣κ cannot be ruled out.

It is worth noting that ancient words are different from modern words. The text at that time did not have spaces, so it was more difficult to determine the boundaries of words.

The official of the final challenge believed that Luke deserved to win the Initial Letter Award; and he himself was very excited when he learned the news:

Another graduate student made the same discovery

Coincidentally, inspired by the results of Casey and Luke, another contestant, Youssef Nader, a graduate student from Egyptian biorobotics in Berlin, used a different method and obtained the same result.

He screened the winning entries for the Ink Detection Prize on Kaggle, which focused on improving Stephen Parsons’ method for machine learning in separated fragments.

On this basis, he used domain transfer techniques to adapt these models to ancient scrolls: unsupervised pre-training on the scroll material and then fine-tuning the fragment labels.

Using this modified model from the Kaggle competition, he managed to find some letters, although not at all relying on Casey’s method of manually finding cracks. He then annotated what looked like letter shapes into the label data.

▲ Left: The earliest images of Youssef’s work; Right: His first set of hypothetical ink labels.

After repeated optimization, Yusef submitted the following works to the official:

In the end, Yusef won the second prize of the Initial Letter Award, which was worth US$10,000.

After seeing Yusef’s work, the experts further confirmed the words in the ancient scroll discovered by Luke.

They even began to speculate that the words above and below might be ανυοντα (realization) and ομοιων (similarity).

Just a few days ago, Yusef’s model produced even more amazing results:

In this image, you can clearly see four and a half columns of text, separated by some margin.

While not all of the letters were immediately identifiable to the experts, at least more were visible.

Officials said experts are conducting further research and will have new results soon.

source: