Thank you. Welcome to NVIDIA's AI podcast.
Today, we're going to talk about archaeology for computer science.
And we're going to do it with the guy who's been described as somebody who can read the unreadable, the Indiana Jones of... of the computer world.
And he is the chair of the University of Kentucky Computer Science Department and a professor there, Brent Seals.
Brent, welcome. Thanks so much. It's great to be here.
So what you have been doing, and this is fascinating, has been applying artificial intelligence to the task of reading English. ancient scrolls.
Reading things that really are unreadable.
I mean, over centuries and even longer than that, we've been trying to read these ancient texts and I guess the problem is always like in trying to read them, we destroy them.
So I want to know how it is that you even thought to apply. the ideas and the notions and the approaches of computer science to something like reading an ancient text. because they don't seem to fit naturally together.
Well, you know, computational methods have changed my world in the last 40, 50 years.
And it began for me with computation applied to the library, basically building a digital library.
I discovered in building digital libraries back in the 90s that some things are actually really hard to make digital.
And most of those things are things that are very badly damaged.
They're so difficult to handle that even taking a photograph of them is going to be problematic.
That's what got me started with technology together with antiquities.
And when you say digitizing libraries, we're talking literally like making them available to read on a screen or... in this case, on your mobile phone these days.
But as simple as that, or as straightforward as that, because it's never simple.
Well, yeah, I mean, that's how it started is that we had all this information that wasn't digital, but we're dealing with this internet of information. where things have to be digital.
So we started narrowing that gap, make things digital and then you can access it.
And at the beginning that was just digital photography, exactly.
So how does it go from the rather straightforward business of photographing pages and getting things digitized to... getting into the trickier aspects of reading these really ancient texts that, you know, you just can't look at and read, frankly, or take a picture of and read.
Well, what I found in visiting the archives is that the curators and the conservators love to bring out things to show you that aren't normally on exhibit.
Because those are the most challenging things in their world.
And so I would go to do work and And then in the cracks, in between, the conservators would show me the most profoundly damaged and interesting things they had.
We realized that there was no method for reading those things, for approaching those things, for digitizing them.
So how then to attack that problem? How to bring image recognition and the tools of computer science to what you saw in front of you?
Those moments happened for me right at the time when a lot of things had been developing.
New methods for imaging were things even when you can't open them up and expose them to the camera.
For example, computed tomography or CT scanning, I mean, it was in my brain in the 90s because of all the advances on the medical side.
And I started thinking... What if you could image things from antiquity that you can't even open with some kind of penetrating technology, like a tomographic technology?
X-rays, basically. And so how did you get permission even to start trying that?
Or did you kind of model things yourself and try to read something without actually opening it up?
Yeah, well look, I mean that's why this stuff, it's not an overnight kind of success.
You can't just stroll in and say, hey, I'm from the University of Kentucky.
That's the other UK, right? And I'd like to scan your thing with x-ray.
Give me, yeah. Give me your priceless antiquities and let me scan it with an x-ray.
Yeah, right. So instead, what we did is a lot of work in the lab with things that we designed specifically and then scanned and then pushed all the way through a software pipeline to sort of show proof of concept to show safety, that this was truly non-invasive and would not harm the objects.
And I probably did five years' worth of that work before I was able to make the pitch to a real institution, a friendly partner.
Are you – did you take Latin and Greek as a kid or, I mean, are you an amateur archaeologist?
I mean, how does the kind of computer science – part of you align with these, clearly it's the humanities, it's history, it's archaeology, but why? go there as part of your research and as part of what you've applied your smarts to?
Well, it was a very organic transformation.
I am a musician and I have always been interested in libraries and museums.
I'm also a Christian, and so biblical material has always been something I've been interested in.
But at the time, really, it was kind of just following the lead of this damaged material and thinking about it systematically as a computer scientist, well, if we are gonna actually solve this, what's the next step?
And then the next step takes you sort of deeper and deeper into the archives until at some point you wake up and sitting in front of you is a Herculaneum scroll.
Well, I did take Latin forever and I got to visit Pompeii and Herculaneum, but tell us, you know, the story of these Herculaneum scrolls and then... what problem it presented and how that was different than some of the other work you'd done before.
Yeah, sure. We'd done these proxies in the lab, and we'd gotten pretty good at very simple objects, pushing them all the way through.
So I used that to make the approach to the Institut de France, where they have six Herculaneum scrolls in their collection, and two of them are actually still intact, unopened.
Wow. Herculaneum, just for people who may not have visited or know, there was the famous eruption of Pompeii in what was the year?
CE 79, maybe? Yeah, CE 79. So a long time ago.
And it buried... Pompeii in feet of ash, Herculaneum, same story, it was a nearby city.
These towns, these cities were completely covered and lost to civilization for a long, long time.
But in that horrible eruption, there was lots of things preserved, it sounds like.
Well, there were. In fact, they discovered about 250 years ago the library at the Villa of the Papyri.
It's the only library from antiquity that's survived.
Wow. It's known to be a library because they found about 1800 of these carbonized scrolls that were still in situ, in place.
So when you say carbonized, scrolls describe for us what these things look like?
Because you kind of sort of think of like, oh, it's a little brittle, but it's, you know, recognizable as a scroll.
What do these things actually look like?
Yeah, I wish it were that easy. It's not that they're a little bit brown on the outside and otherwise... pretty much scroll-like.
Instead, they sort of look like charcoal briquettes, pieces, fragments, and you recognize instantly how friable there.
If you dropped one of these, it would completely shatter like a piece of glass because the carbonization...
While it didn't consume the papyrus, it transformed it the way charcoal gets transformed.
Have you ever gone to the grill? at the end and you kind of touch it and it all falls apart.
Right, right. Yeah. So you kind of get that feeling when you look at these things, it's the same kind of effect.
Okay, so you have these charcoal briquette scrolls and...
You have your theories about how to read them, and still, it's mind-boggling that you can actually read them, but how does it work?
What do you do? Well, tomography... is all based on computation.
Originally, it was done in an analog way, but Nobel Prize, actually, in medicine in for the process of doing computed tomography and it truly is a miracle because with x-ray And the wavelength of x-ray being so tiny, the resolution that you can get on the inside of an object using tomography is just fantastic.
I mean, on the order of... Microns. Smaller than a human hair.
However, I understand how you can kind of sort of see inside things, but that doesn't help you read it.
So how do you bring to bear... the computer science side of things and artificial intelligence to help kind of take images and then...
And images that are wrapped up, like you say, in kind of charcoal briquettes and then convert them to things that can actually be read and or viewed.
Yeah, well, it turned out that this problem fell exactly into my wheelhouse because I was doing 3D computer vision. and everything about how the geometry of shape could be approached computationally.
And that's exactly what you're talking about in this scan. of the internal structure of something that's wrapped up.
Because every one of those wraps is a three-dimensional structure. section of that surface that has writing on it.
Oh, so because we think of reading as a 2D thing, right?
It's a character on a page. But you are looking at it as a three-dimensional problem.
Exactly, it has to be that way because this thing which is wrapped turns around in three dimensions, and if you're going to untangle all that and read it again in two dimensions, you have to have a computational transformation that's going to take you from These little pieces that effectively what they look like is potato chips.
I mean, you follow the surface for a section and it curves like a potato chip.
And then you follow another one and you get another potato chip.
And those have to be flattened and then put together like a jigsaw puzzle.
And so how did the tools of computer science, and in particular, whether it was deep learning or some machine learning, how did you apply those to this problem?
But what we've been doing with learning is we've been figuring out how to follow those potato chip-like sections through that volume because it turns out that there's a lot of stuff going on inside there that would get you confused as to whether you're on the same page or going to a different page.
But we can train an algorithm to recognize when you're following that same layer along.
We can train it to be pretty good at following the same layer.
And that's one of the key things to be able to do.
So right, it's following a sentence or a word or a series of characters down a page in the right order.
Actually, even before you get to the writing, it's following the surface on which there probably is writing.
Oh, I see. All right. Well, so take us through it.
How does the algorithm then get from, you know, following one potato chip to the next to then at the end, we get to... read or know about what life in Herculaneum was like in AD 79.
Yeah. That's where we're trying to get. The following of the layer is a huge application of machine learning because we've We've done a supervised network that is really good at following those layers and then giving us the final result.
And then the second thing that we do is try to paint the ink that should be on there from the original volume where that surface was sitting.
And when we do that painting, basically, we apply another machine learning algorithm that knows how to sort of amplify what ink probably looks like inside that volume.
And is the ink then another object that you can identify and kind of track separately from the wrap and the three dimensions of these scrolls.
Well, the ink and its effect in this x-ray is much more subtle than anything else that you might be looking at.
And that's why we actually look for the ink as a secondary step.
First, we follow the surface itself because it's really easy to see the fibers of the papyrus, for example.
And then once we know where the fibers of the papyrus are, then we start looking for the much more subtle effect of this carbon that has been deposited into those fibers.
That subtle effect is amplified using machine learning.
Oh, and that subtle effect is writing. That's fascinating.
Yeah, we have results in the lab right now that are pretty encouraging and we're gearing up to go back and collect the full data from the real material.
So that we can push everything through the end.
So we're kind of halfway there, I guess.
And by halfway, I don't mean another two decades, but I mean maybe another year. is this technique or is there opportunity to apply this technique to, to other texts and other schools and other, you know, other things for that matter around the world?
Absolutely, this application of machine learning to tomography and making tomographic evidence more sensitive, more prominent, basically recognizing really complex structures inside tomography.
I mean, it's a huge area of advancement right now, and it's playing a role in medicine.
For example, tumor detection for cancer. and of course in antiquities now.
The people that you work with in France and in Italy and other parts of the world, what do they hope to learn or what Do you hope we can learn from really starting to decipher these ancient texts?
Well, for Herculaneum, being that it's the only library from antiquity, I mean, I think we owe it. to humanity, really, to try to wring what we can out of that cache of material.
We have such a limited amount of material from that era.
And you're talking about 300 intact scrolls and the possibility of further excavation where they might find another...
2,000 scrolls. So I think that's front and center.
As far as other material, there's just a ton of stuff from a limited cache of material that remains unanalyzed.
And we should be cranking through it to find out what it says.
Again, is there something that you would like to uncover?
Is there some text that's out there that that you would like to apply this to?
I mean, what motivates you to really do the sort of impossible Well, I'll tell you, last year was pinnacle because I did work with a scroll from the Dead Sea Scroll Collection.
And as a kid, I mean, that was... an iconic part of biblical archeology for us, you know, that that cache of materials had been discovered and then it confirmed readings that were right there in our Bibles every Sunday.
So that really was a dream for me. When I look at Herculaneum material, I don't know if you know this, but The Bible in the chapter in the book of Acts actually talks about the Apostle Paul stopping off at the Bay of Naples and At the time when he had been doing all the writings to the churches,
And it's always been intriguing to me to think about those writings, those early Christian writings, what happened to them.
I mean, would it be possible to discover any early Christian material in the Herculaneum cache?
You know, when I realized what Herculaneum was and the era, you know, because I'm not a historian, but I've learned about this.
I realized, wow, that was so much was happening in the ancient world at that time.
It's really fascinating to think about. Well, it's an exciting time to be a computer scientist and somebody who can bring to bear the tools of of AI to these kinds of problems.
And I think it's great that these ancient worlds and this very modern world can kind of help each other out.
So I want to thank you for helping us out, and we'll keep you on this and look forward to more progress.
Professor Brent Seals, thank you so much.
I thank you. PostScript here. If you are listening to this podcast on iTunes, Google Play Music, or whatever your favorite podcast player is, do us a favor and leave us a review.
Thanks for listening.
