Winners of 2023 Vesuvius Project announced

[guy]

The 2023 winners were announced of the Vesuvius Challenge. The goal was to read portions of the Herculaneum scrolls that were left damaged after the eruption of Vesuvius in AD 79.

First breakthroughs occurred in October 2023 (see post below),

A Grand Prize was awarded to the first team to recover 4 passages of 140 characters from the Herculaneum scroll. Smaller prizes were also announced.
 

 

https://scrollprize.org/

 

The Vesuvius Challenge for 2024 is the goal of reading 90% of a scroll by the end of the year.

 

 

Edited August 10 by guy

[caesar novus]

I wonder if they can make an estimate of how accurate their AI results are. Pattern matching approaches may be biased to find known stereotypes. Maybe they can calibrate by running against remains of an old Torah vs. modern, which I gather doesn't change even punctuation marks for ages.

I had involvement in early AI approaches, which were quite elegant in tracking degree of certainty and could even explain it's reasoning. But sometimes this approach hits a wall, and fuzzy pattern matching approaches became popular, which I didn't respect much.

An example of the difference is innocence vs reasonable doubt of guilt (= not guilty). An exact AI system may flail forever trying to determine innocence with uncertainty of alibis etc. A pattern matcher can step back and rule not guilty if the perp looks anywhere from 0 to 98% guilty (with 99-100% being guilty). BTW I hate those "innocence projects" that crow about tiny adjustments of reasonable doubt.

Wikipedia unexpectedly has a very lucid description of various current AI approaches https://en.wikipedia.org/wiki/Artificial_intelligence which puts me more at ease that more than just pattern matching is used. The sloppy AI narrations common on youtube probably shouldn't blacken the reputation of AI in general.

Edited February 11 by caesar novus

[guidoLaMoto]

Amazingly clever use of technology.

Maybe I don't know the definition of "AI," but his is no more intelligent than a vending machine used to get a candy bar,  just more complicated. 

The video in the Nature article cited shows that they did a CT scan from end to end of a scroll, then used a program to transform the scan into  flattened 2D images, layer by layer along the axis of the scan. That program included a proportionality factor to exaggerate differences in x-ray transmissibility, giving the image of dark letters on a lighter background. ...Basically the same principle used in medical CT scans. No "thinking" involved.

Ma, che furbi, 'sti studenti!

 

 

Edited February 11 by guidoLaMoto

[caesar novus]

I understood that they first used conventional approach to virtually unwrap and examine,  as done for the dead sea scrolls. Then let the contestants try to find still elusive words:

Quote

“crackle” — that seemed to form the shapes of Greek letters. Luke Farritor, an undergraduate studying computer science at the University of Nebraska–Lincoln, used the crackle to train a machine-learning algorithm

That was from skimming the Nature article; maybe I missed something from the videos which are too linear for my patience. The news article may have oversold/clickbaited the AI angle.

[Danielerefs]

Were there any ancient equivalents to the modern Super Bowl ads in ancient Rome, and if so, what were they promoting and how were they presented to the audience?