Hacker News Reader

Comment by NitpickLawyer | Hacker News Reader

> Both images were independently reviewed by the Vesuvius Challenge papyrological team, led by Federica Nicolardi. The simultaneous reproduction of the title image from multiple sources, along with independent scholarly review, provides a high degree of confidence in the reading.

You have 2 teams using the same data, getting to the same conclusion. You also have an author that's known from other sources, with writings that we already "have". Then you have a team of experts reviewing this. Chances are these are real findings and not "hallucinations". Not everything in ML is gen-ai...

View on HN

To add to this, the main ML parts, as I understand it, are for the initial unrolling of layers, and for the detection of ink vs. no-ink (the position of the 'splotches'). Both of these are trained and calibrated from human observations.

All interpretation of ink as Greek letters is done purely by human inference. This may lead to errors, especially in parts where the ink is preserved especially poorly or where the text is totally different from expectations, but it would be classic human error instead of AI hallucination.

> You have 2 teams using the same data, getting to the same conclusion.

> You also have an author that's known from other sources, with writings that we already "have".

> Then you have a team of experts reviewing this.

Only the first of those points is evidence against the result being hallucinated.

>You also have an author that's known from other sources, with writings that we already "have".

Well, that's exactly what you'd expect from a hallucination no? If the model is overfit enough on the relevant corpus, a title that already exists should be much more likely.

The ML models are looking at tiny patches for areas of ink vs. no-ink, trained on the boundaries of more visible letters found by humans. They don't know what proper Greek letters look like, and they definitely don't know what correct Greek words would be (in particular, they have no "corpus" of words). Any possible overfitting is ordinary human overfitting.

If they have learned on ink areas shaped as letters, what prevents them from having a bias towards such shapes?

Because they don't get to see the entire letter shapes. The page at [0] shows the basic idea: they're forced to make decisions based on each part of each stroke of the letter. If they were heavily overfitting on the strokes of the letters in the training sample, then they'd be so inaccurate outside the sample that the assembled outputs would hardly resemble letters at all, much less words.

(Also, this is heavily-damaged handwriting, not clear print, so each letter isn't even uniform in shape. A model trying to cheat at ink detection would have an uphill battle trying to guess what all the variant letter shapes might be.)

[0] https://scrollprize.org/grandprize#how-accurate-are-these-pi...

Reply on news.ycombinator.com