Incredible, artificial intelligence decodes the papyri burned during the eruption of Vesuvius in 79 AD

Incredible, artificial intelligence decodes the papyri burned during the eruption of Vesuvius in 79 AD

Two thousand years ago, a volcanic eruption buried an ancient library of papyrus scrolls, known as the “Papyri of Herculaneum“. These precious scrolls were charred following the historic eruption of Vesuviuswhich occurred in 79 AD The manuscripts were discovered in the 18th century and for over 800 the library of Naples jealously guarded them.

Examining them, the papyri in question have the appearance of small stumps: in reality they are masses of charred ash which, obviously, cannot be opened without causing irreparable damage. The challenge, which until now seemed impossible to overcome, consisted in reading the contents of the papyri leaving them unaltered, in their original condition, therefore completely rolled up on themselves.

The winners of the challenge: read the contents of the Herculaneum Papyri, carbonized in 79 AD

In March 2023, Brent Sealesprofessor of computer science at the University of Kentucky, together with entrepreneurs Nat Friedman and Daniel Gross, launched the Vesuvius Challenge, an initiative that aimed precisely to decipher the content of the Herculaneum Papyri. Now, just one year later, the goal has already been reached. At least as far as the initially set objective is concerned.

The entrepreneurs offered hundreds of thousands of dollars to the benefit of anyone who proposed an effective solution to read the contents of the papyri without causing any damage. And today the winners of the incredible competition can finally reap the fruit of their labor. Youssef Nader, Luke Farritor and Julian Schilliger cash in $700,000 but, above all, they demonstrate that they have identified the right “keys” to access the contents of the manuscripts, two thousand years later.

3D reconstruction of the Herculaneum papyrus scroll

3D artistic reconstruction taken from Vesuvius Challenge.

Artificial intelligence and machine learning are the protagonists of the decoding operation of the precious Vesuvian scrolls

As this paper explains, the researchers used a combination of advanced techniques, including computed tomography (CT), computer vision and machine learning to decode the Herculaneum Papyri.

Transported with care to the Diamond Light Sourceand particle accelerator near Oxford (UK), the Herculaneum scrolls were subjected to a X-ray scan at high intensity. In this way, using the parallel beam thus generated, the researchers were able to obtain a rapid, accurate and high-resolution replica (this video shows a sectional roll).

X-ray images have in fact made it possible to create a three-dimensional representation (3D) of the papyri, allowing researchers to virtually examine them layer by layer without damaging the original supports.

The software called Volume Cartographer he therefore allowed to to track the various ones manually layers of the papyrus processing the 3D scan. A step that proved essential for identifying the structure and arrangement of the papyrus sheets within the roll.

Training the model used to decode the contents of papyri

During the summer of 2023, another researcher, Casey Handmer, discovered the presence of a detail model called crackle within the X-ray images. This pattern appeared to form letters and signs.

Crackle papyrus Herculaneum

Image source: Casey Handmer’s blog.

Farritor, one of the winners of the Vesuvius Challengethus proceeded to train a model of machine learning to detect the shapes associated with the text starting from the images.

The other winner of the grand challenge, Nader, extended the application of adaptation techniques built from machine learning models previously trained on open fragments of papyrus. This insight helped detect ink in 3D images by clearly highlighting it on the charred scrolls.

By using the approach known as TimeSformer it was therefore possible to obtain even sharper images. The TimeSformer is a type of architecture based on a neural network that is proposed as a variant of the classic Transformer concept, introduced by a Google team in 2017 (and today widely used in generative modelsincluding those of competitor Microsoft-OpenAI).

The main feature of TimeSformer it is his ability to process 3D sequences and spatiotemporal data. While traditional Transformer models are primarily designed for two-dimensional and sequential data, TimeSformer extends this architecture to manage three-dimensional data sequences, as in the case of X-ray images obtained from papyrus scanning. Here you will find theoutput product downstream of processing with TimeSformer.

Various measures related to the optimization of the algorithms artificial intelligence they allowed us to avoid the typical problems due to overadaptation and hallucinations. The overfitting occurs when a pattern of machine learning it adapts too much to the training data, also learning the details of the noise and random variations present in that specific data. Performance on “new data” can therefore be negatively affected.

The results of the studies on the Herculaneum Papyri are now under the eyes of the world

On October 12, 2023, Faritor, a 21-year-old university student interning at SpaceX, managed to decipher a word from the Herculaneum manuscripts for the first time: it was the term “Πορφύρας” (purple, purple).

The work then continued incessantly, succeeding in the extremely complex task of decipher four words with an accuracy greater than 85% by the end of 2023. The team composed of scholars Nader, Farritor and Schilliger managed to read more than 15 columns of text, for a total of over 2,000 characters. An enormity for this first attempt, considering the state in which the scrolls are. By clicking here, you can check the columns recovered from the papyri so far: 95% of the text has yet to be explored.

What are the papyri about?

The deciphered text concerns the “source of joy“, the supreme good in Epicureanism. This is a concept dear to the ancient Greek philosopher Epicurus, who taught a particular conception of happiness and pleasure. That “source of joy” was closely linked to the pursuit of pleasure, but in a deeper and more lasting sense. It wasn’t just about seeking momentary pleasures, but about achieving a long-lasting, stable form of pleasure that led to an overall happy life. For Epicurus, happiness was achieved through moderation, the satisfaction of essential needs, and the cultivation of friendships and meaningful relationships. The “source of joy” could be found in the tranquility of the soul, in the satisfaction of basic needs, and in the minimization of worries and anxieties.

Challenge renewed for 2024: the bar is raised even higher

The challenge, however, is only just beginning. For 2024, the organizers have set an ambitious goal of deciphering 85% of the contents of the papyri by the end of the year. To try to cross the finish line, there is talk of the imminent installation of one scanner useful for examining significant volumes of rolls, every day.

In 2023, researchers went from 0% to 5% of a scroll. In 2024, the goal is to bring 5% of the content currently decoded in a single roll to 90% of all four rolls available. This leap forward will lay the foundation for successfully read all 800 scrolls.

Il bottleneck that remains to be overcome concerns the process of tracing the surface of each sheet of papyrus within the individual scroll. This process is now manual, therefore requiring a truly immense expenditure of effort. The work cost more than $100 per square centimeteronly to get to read the 15 columns published today.

With these exorbitant costs, segmenting all the rolls would cost hundreds of millions or perhaps even billions of dollars. At this point we need to invest inautomation and therefore on the so-called self-segmentation.

If you are interested, you can contribute with an investment and thus join the craziest archaeological project in existence“, we read in the celebratory post of Vesuvius Challenge.

Who are the investors who contributed to the initiative

At the beginning we mentioned Nat Friedman and Daniel Gross. The first is an entrepreneur and manager…

Leave a Reply

Your email address will not be published. Required fields are marked *