Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Machine Learning–Based Identification of Lithic Microdebitage - Ep 207
913 Plays1 year ago
We talk to Dr. Markus Eberl about his team’s use of a particle scanner to analyze micro-debitage. They used machine learning to analyze the data set and tried to learn more about early life than we could otherwise.
Transcripts
For rough transcripts of this episode go to https://www.archpodnet.com/archaeotech/207
Links
Contact
- Chris Webster
Twitter: @archeowebby
Email: [email protected] - Paul Zimmerman
Twitter: @lugal
Email: [email protected]
ArchPodNet
- APN Website: https://www.archpodnet.com
- APN on Facebook: https://www.facebook.com/archpodnet
- APN on Twitter: https://www.twitter.com/archpodnet
- APN on Instagram: https://www.instagram.com/archpodnet
- Tee Public Store: https://www.teepublic.com/stores/archaeology-podcast-network?ref_id=5724
Affiliates
Motion: https://www.archpodnet.com/motion
Motley Fool
- Save $110 off the full list price of Stock Advisor for your first year, go to https://zen.ai/techpodfool and start your investing journey today!
- *$110 discount off of $199 per year list price. Membership will renew annually at the then current list price.
Laird Superfood
- Are you ready to feel more energized, focused, and supported? Go to https://zen.ai/thearchaeotechpod1 and add nourishing, plant-based foods to fuel you from sunrise to sunset.
Liquid I.V.
- Ready to shop better hydration, use my special link https://zen.ai/thearchaeotechpod to save 20% off anything you order.
Recommended
Transcript
Introduction to Episode 207
00:00:01
Speaker
You're listening to the Archaeology Podcast Network. Hello and welcome to the Archaeotech Podcast, Episode 207. I'm your host, Chris Webster, with my co-host, Paul Zimmerman.
Interview with Dr. Marcus Eberl
00:00:15
Speaker
Today we talk to Dr. Marcus Eberl about his team's efforts to analyze microdebitage using particle analysis and machine learning. Let's get to it.
00:00:27
Speaker
Welcome to the show, everyone. Paul, how's it going, man? OK, I think I'm really, really sleep deprived, but happy, I guess. You know, came back from Saudi a week ago, just over a week ago, and I've got another week and a half or so before I head back. Oh, man.
00:00:44
Speaker
We're going to have to find some more co-hosts for you because my schedule is awful lately. How are you doing, Chris? And where are you? Yeah, I'm doing all right. We're in northwestern Washington. If you see a temperature map of the United States, we're the only blue part that is not over 100 degrees.
Context and Recording Details
00:01:01
Speaker
I mean, in fact, it's 70 degrees right now, the highest 72, and I am loving it. I don't ever want to leave here just because of that.
00:01:07
Speaker
given where everybody else is doing. So just so everybody knows too, we are recording video. So if you want to watch this rather than listen to it, you can go over to archaeology podcast network on YouTube and check it out. Let's get into the show. I'm going to introduce our guest through his bio right now, and then we'll start talking
Dr. Eberl's Archaeological Expertise
00:01:25
Speaker
to him.
00:01:25
Speaker
Marcus Eberl is an anthropological archeologist and epigrapher. He has conducted archeological fieldwork in Germany, Israel, Mexico, and Guatemala. And he currently directs archeological projects in Tamarindito and Zikan and Sakan, both ancient Maya sites in Guatemala's tropical lowlands. In the laboratory, he specializes in soil and ceramic analysis. Recently, he acquired a dynamic image particle analyzer for his micro artifacts lab. He now uses machine learning to identify human made artifacts in soil samples.
00:01:55
Speaker
His book publications include Community and Difference, Change in the Late Classic Maya Villages of the Patex Bhutan region, 2014, and War Owl Rising, 2017. Okay, now that we know Marcus, we'll introduce him. Marcus, how's it going? Perfect. Thank you so much for having me on your show.
00:02:14
Speaker
No problem. So yeah, Marcus is fresh in from the field, like literally a few hours ago as we're recording this. So we appreciate you coming on and, and recording this interview with us. But what led you to us was an issue in advances in archeological practice that we will link to in the show
The Role of 'Slow Data' and Machine Learning
00:02:29
Speaker
notes. And the title is machine learning based identification of lithic micro debitage, which just sounds super cool. So I'll kick it off with a question here.
00:02:40
Speaker
And we'll get into a bunch of other stuff. But you mentioned right in the beginning of this article, slow data. And honestly, I don't think that's a term I had heard before. So what do you mean by slow data?
00:02:51
Speaker
I think as archaeologists we all tend to work for long periods of time at the same site where we really produce very deep and rich data about this specific location that is really contextually very important and super rich and of course where we spend a lot of time interpreting it.
00:03:13
Speaker
But from my perspective, and I think many others, we also realize that it's very hard to generalize from this data set to something, to patterns that are broader, larger, where we can say, oh, this is something that is not only happening here locally, but where we can look at patterns that are visible on a much larger scale.
00:03:36
Speaker
Again, I'm not critiquing slow data. I'm just literally coming back from the field where I'm producing slow data. What I try to do in this paper together with my colleagues is to point out how we can complement the slow data with other approaches that include, for example, machine learning. Why machine learning in particular? You have an interest in it, obviously, but what makes you gravitate toward it?
00:04:05
Speaker
So as we explain in this mouthful of an article, my interest was in a lot of understanding lithic, or I should say stone working. So in the area where I'm working, the classic Maya, they never really developed metal tools, as we see it in many other complex ancient societies.
00:04:28
Speaker
But they really relied on stone napping, napping of chert, of obsidian and possibly other stones to produce all their tools and many other implements in their life. And for me, it was always interesting to understand, well, who would
00:04:44
Speaker
did this stone napping, how did they produce their stone tools? And this brought me then to the interest into microdabitage because modern stone nappers tend to get rid of their debris that they produce during stone napping simply because it's very sharp and you don't want to walk over it or you don't want your children or others being hurt by it.
00:05:08
Speaker
And so actually, among the classic Maya and other societies, we know relatively little about how the stone napping actually works. Where do we have workshops? And one of the approaches that archaeologists have been toying around for decades now is the use of microdabitash, meaning to look into the microscopic debris that these stone workers are producing.
00:05:36
Speaker
And the advantage of microdabitage is that in most contexts, it's very hard to get rid of them. So unlike larger flakes that you can easily pick up on surfaces, microdabitage gets easily lodged into floors and other areas.
00:05:55
Speaker
So this means that by analyzing microdabitage, we can get an idea about where ancient stone knuckles have been working. My interest is now well to analyze microdabitage, to recognize microdabitage in soil samples, and to find out where ancient stone knappers have been working. Okay. In the past, or should I go on a little bit about how people have done this so far? Or
00:06:23
Speaker
Well, let me let me ask a clarifying question first. I mean, how do you define micro
Understanding Microdebitage
00:06:29
Speaker
debit? I read in the article, but just so we can get it out on the on the podcast here, what how are you defining micro debit? And how does that debit differ than say regular debit? You know, the tertiary primary secondary flakes. However, we identify those. Does it have different characteristics, not just size?
00:06:44
Speaker
Yeah, so and this is again, and actually, I mean, I can talk about a paper and progress that I'm just writing about exactly that. So there's a lot of discussion, you know, what is micro devitash, or this paper, the machine learning, we defined it as all flakes that are smaller than 6.3 millimeters.
00:07:02
Speaker
or a quarter inch. So this would be what would be seized out during regular screenings in excavations. So this is more like a heuristic approach. And in the paper that I'm right now writing, I argue that probably a tighter definition is warranted. But at least for the key ideas,
00:07:22
Speaker
that this is the stuff that falls through our sieves when we are excavating and looking for artifacts. One of the other aspects that you mentioned, this is one of these critical ideas that microdabitage supposedly shows the same characteristics as regular dabitage, like the choidal fractures. And this is, again, one of these issues that I can come back to.
00:07:48
Speaker
where I got a little bit skeptical looking at microdabitage because for me, it's much more variable than the regular flakes that we would pick up from a stoneworking workshop. Why would that be? Why is there so much more variability?
00:08:01
Speaker
I mean, one is simply the size. So what we did is we work with modern stone nappers. So we go to nap-ins. And for this particular paper, we work with a good friend of mine, Michael McPride. He's like, reproducing stone tools. So we asked them, you know, could you make a biface or arrowhead, whatever you want to do? And can we collect your debris?
00:08:26
Speaker
And most of the stone nappers that we approach, they are super happy. I mean, you know, for them, it's just, well, it's debris. And so then really care about it. But for us, for my team, it's really super interesting to collect all these thousands, if not tens of thousands of particles. And what we then do, and I can come back to that, we run all of these particles through a particle analyzer. So this is a machine that allows me to take photos of all of these particles,
00:08:56
Speaker
and describe them in various ways. And the interesting part to bring it back to Paul to your question, when we looked at the variability, like, you know, simple dimensions, how long is each particle? How wide? What is the transparency angularity? It turns out they don't fall in this
00:09:17
Speaker
clear, like a tightly defined class that we would expect. Instead, there is a huge variability. And this is one of the issues that makes dealing with micro-debitage so difficult. They're much more variable than people previously assumed. And is micro-debitage that you've derived from human activity, from flint napping,
00:09:41
Speaker
Is it statistically different, is significantly different than other little stone bits you might find? I mean, are you able to determine that? So this is exactly where it becomes really interesting because what we did for this article is we compared the experimentally produced macrodabitash to a regular soil sample that I picked up many years ago during my dissertation in Guatemala.
00:10:07
Speaker
And I should say this soil sample contained all kinds of stuff. So, you know, like little plant twigs, little ceramics, whatever was, you know, what you would find in a soil sample. And so what we did for this article is to compare the experimentally produced microdavitage to this other, so which contains not only rocks, it contains sand, twigs and whatnot.
00:10:33
Speaker
And the interesting part is statistically, we applied not only the machine learning algorithms to these two samples, but we also ran statistics. And one of my graduate students is writing a paper about that. And we can show that these are statistically really different. There is some overlap. But overall, if we look for specific dimensions, we can really differentiate between these two categories very nicely.
00:11:00
Speaker
Okay. You mentioned when you first started getting into this that people have tried to study microdebitage in the past. Aside from your approach of using these tight scanning methods and machine learning to suss this out, what have been some of the past approaches to this? That's what I tried about 20 years ago when I was dealing with my dissertation. I was working at a really interesting site in my area and my mentor encouraged me to look into soil sample.
00:11:29
Speaker
And I was like, well, that would be really interesting because we knew that there was a stone napper working at this particular location where we excavated. And so I, you know, I used the microscope that we had in the lab. And what I did was the traditional approach of analyzing soil samples. So I pulled, I think I have a head about
00:11:49
Speaker
50 soil samples, and I sift them into different size fractions. And then I looked at each size fraction under a microscope, hoping to find, oh, what is soil and what is microdabotage? And this experience, I mean, it was super tedious. I mean, literally, I spent hours and hours just looking through a microscope at soil samples.
00:12:12
Speaker
And ultimately, I mean, you know, this experiment never made it into my dissertation because I really questioned myself, you know. After looking for hours and hours at the same soil sample, I was like, do I really see that or am I just imagining that this is a tiny flake or is this really just imagination speaking?
00:12:36
Speaker
Just to give you a few numbers, what I didn't realize at that point with the particle analyzer, I regularly find 300,000 to 500,000 particles in a soil sample. This is literally less than a handful of soil. Now you can imagine the
00:12:57
Speaker
idea would be, or the ideal for an archaeologist would be to look at each of these hundreds of thousands of particles and say, oh, this is soil, this is microdabotage. And at least for me, this is where I became skeptical, where I said, you know, I mean, I can do this for perhaps 100, 200 particles.
00:13:18
Speaker
But to really scale it up to 500,000 or even millions of particles, I said, you know, this is, I mean, at least for me, I couldn't say that this was what I could really get a quantifiable result out of that. Wow. That's really cool. Well, why don't we go ahead and take a break? In the meantime, check out arcpadnet.com forward slash members to join us for our next culture or share event and see all the other stuff that we have for you back in a minute.
00:13:47
Speaker
Hi, welcome back to the archaeotech podcast episode 207. Today, we're talking with Marcus Eberl about a recent article of his about computer vision and lithic micro-debitage. And Marcus, just before we went to break, before we got a little glitchy there, talking about some of the shortcomings of doing this kind of work.
00:14:05
Speaker
with a more traditional manual process of using microscope and counting particles by hand. So much so that you gave up on that initially the first time through with it. So aside from just being slow and maybe inaccurate, are there any other major shortcomings to that kind of more traditional approach? And how do you address those with the computer vision approach?
00:14:26
Speaker
Thanks for asking that. I think another
Challenges in Manual Analysis
00:14:29
Speaker
shortcoming is also the interobserver error. If I want to train other people to do this type of analysis,
00:14:39
Speaker
how well will different people see the same thing in the same sort of sample? And again, I mean, I have colleagues like Isaac Ola, who really trained and other observers, but this has been a problem in the past. We really get a standardized observation that different people see and classify the same things. And I think especially with micro-debitash, we easily run the risk of
00:15:09
Speaker
that different people see different things and count different things. Another thing I was interested in in what we were seeing here is how the ground surface influences the accumulation of microdevitage. Like if you've got a really smooth, clean surface, I mean, you can tend to sweep all that out and not have anything, whereas if you're just on the ground, you know, and you've got more crevices and things. So just tell us a little bit about that and your experience in this collection and what you can find.
00:15:35
Speaker
You know, I mean, this is another really critical aspect. I mean, for example, in the area where I'm working with classic Maya, the upper levels of society had the nicer houses, often with very smooth stucco floors. And a colleague of mine,
00:15:52
Speaker
did some experiments with microdabitash and she found that very little actually stuck to the stucco floors. I mean, they were very hard and you wouldn't really expect to find macrodabitash in whatever crevices you have there because the surfaces tend to be too dense and too smooth for like a good analysis of microdabitash.
00:16:15
Speaker
On the other hand, the area where I'm working, so as a Maya archaeologist, I tend to specialize on the non-elite section of the population, where we have mostly pebble floors, possibly filled with packed soil. In these types of environments, microdabitage tends to stick much better.
00:16:37
Speaker
I mean, so we have here like areas that couldn't be where you could clean the larger flakes off very easily. But I think where people would struggle to clean off microdabotage because it's really it's it's sinks into the soil.
00:16:52
Speaker
I mean, two points on that. First off, I mean, just in our own society today, I mean, you tend to see that... I don't know how to say this in any way that's not PC, but certain levels of society tend to just have more things around anyway and don't clean up as much to begin with, right? So I have a feeling that that's probably similar in other cultures. And along those lines, what are the chances the people in the elite of Maya society in particular are even flint knapping and not just having things created for them?
00:17:22
Speaker
I mean, that's actually really one of the aspects that I and my colleagues are really interested in, because at this moment, we don't really know. I mean, again, you know, it makes sense what you say, you know, like, like, oh, if you're like a Lord, I mean, you won't be napping.
00:17:37
Speaker
things. But the point is, you know, it's really hard to prove that point. And that's what I find fascinating about archaeology in general. And this approach in particular, you know, I mean, I can go back and say, you know, if I work in a palace, you know, do I have any evidence of microdibitage? Or is there a potential that, you know, a few stray flints are still visible? So for me, it's more like a question to be asked instead of
00:18:04
Speaker
assuming, okay, we, you know, people at that level, they don't care about napping church or whatever. So, yeah. So a traditional method of trying to identify workshops would be looking at larger flakes, for example, or could be across any variety of different kinds of artifacts of any kind of workshop that you're looking at. But obviously, you'd be looking for lithic flakes. If you're looking for these foot napping workshops, in absence of larger flakes,
00:18:34
Speaker
is looking at the microdebitage. Is that a reliable indicator of workshops?
00:18:40
Speaker
And this is going to be an interesting discussion in the future. So for some researchers, they would like to see both, meaning that we have here microdibitage, and we have some kind of evidence, like visible evidence for a workshop. That could be, of course, larger flakes. But this could be also specific other tools that were used for flint knapping. That could be a specific layout.
00:19:08
Speaker
And I personally would veer more towards that. I mean, I think clusters of microdabitash can be very useful, but I would like to see additional evidence. For example, in the area where I'm working, these are larger residential groups with multiple buildings grouped around the plaza. And it would be great to see elevated concentrations of microdabitash in specific areas of this group.
00:19:38
Speaker
but i would like to see additional evidence for example you know do we have other types of tools that we find for example in the middle of this group that would point out oh we have here very likely a stone napper who is working there so for me again you know i mean as i said you know i'm not against slow data
00:19:57
Speaker
So, you know, I'm perfectly happy to do an extensive excavation, look into mittens, and then to really use, for example, my macro debitage analysis with machine learning to compliment that with my excavations of mittens and others. And at least that's my gut feeling right now. Man, I got to go back to this elite thing real quick, because
00:20:21
Speaker
I just, first off, I love how much a simple question can actually put a lot more things into play, right? It just leads you down this path because I'm thinking, okay, so if they are Flint napping and somebody's cleaning up after them, right? Whether it's them or somebody else cleaning up after them, you think you would see
00:20:38
Speaker
some sort of common place that they would sweep up all this debris from inside the house, dust, skin, lithic debitage, whatever it is, and then dump it in a place. So now you're getting to where's their trash collection for this elite place. And could you analyze that to see, well, in this, what looks like household trash, we have microdebitage.
00:20:58
Speaker
You're absolutely right. And the problem often is, so there have been several archaeologists who did like experiments or ethno archaeological studies. The problem with this, that modern stone nappers, and I say stone nappers, because in the area where I'm working, people would be both using flint and obsidian. So these stone nappers, they often go out of their way to dispose of the trash.
00:21:24
Speaker
So it's not something you know where you simply you know where you just walk behind the house and you see oh there must be the mitten and you know this is where I have to excavate. I mean at least in some of these modern cases stone nappers really knew I mean you know they wouldn't have their family handling these sharp debris and they really walked like hundreds of meters away from the
00:21:47
Speaker
area where they were working. So in the sense, that would then add the challenge, you know, as an archaeologist, where do I find the trash and especially the workshop trash? It could be a different place than where they would put the kitchen debris, you know, all the vegetables and fruits that were thrown out.
00:22:05
Speaker
Okay. I think maybe we should get into the part of the actual title of this with the machine learning stuff. I'm interested because you mentioned 20 years ago, you're interested in microdebitage. You're looking through a microscope at soil samples and you're figuring all this out. So presumably you've been thinking about this throughout and it's been one of those things that's just kind of like plaguing your mind. When did machine learning become viable enough for you to look at it and start saying,
00:22:31
Speaker
You know what? This could be a tool where we could actually solve this problem. Is it with this paper, or was it the genesis of that a little bit earlier?
Adopting Particle Analysis and Machine Learning
00:22:38
Speaker
Actually, it goes back a little bit earlier and really started with the particle analyzer. And this was, I don't even know what got me there, but this was like a late night Googling accident. Because I was always interested in different ways in which I could describe a flake. I mean, how could I get quantitative data on a flake?
00:23:00
Speaker
And as I said, some Googling accident late night brought me to particle analysis. And just to explain that, these are machines used by modern companies for quality control.
00:23:15
Speaker
So the company where I bought this machine from, they told me, oh, they sell it to glass speed manufacturers. So where people have 10,000 glass beads and they want to know whether this run of glass beads is exactly according to whatever specification. So they were actually super surprised and delighted to have an archaeologist knock on their door and say, oh, this is exactly what I want to use for my work.
00:23:40
Speaker
And this was really the thing where I realized, I mean, the company invited me to send them two samples, you know, to test the capacities of their machine. And when I got the results back, you know, they sent me two Excel spreadsheets, I realized, okay, I'm in trouble, because this is more data than I really expected. I mean, you know, they couldn't even send me a spreadsheet just as an attachment, I had to go to like Dropbox or similar servers.
00:24:10
Speaker
I mean, like, you know, there were like, tens of thousands of particles listed there. And each particle I have like 40 variables. So this is then where I realized, okay, I mean, you know, I mean, I'm into statistics, and I can handle that. But just looking at the spreadsheet, I realized, you know,
00:24:31
Speaker
This seems to be like something where machine learning or computers in general could do something really interesting. And I then linked to the data scientists who are the co-authors on this paper who were super delighted because for them, this was the ideal data set that they have been looking for, you know, like exponentially produced data where we know what something is. And then you just try to see
00:25:01
Speaker
this specific class in a different data set. So they have been amazing, you know, like, and I've been working with them for the last five years. And this is really where the machine learning then took off, because they said, you know, this is a perfect data set to apply machine learning algorithms. So in many ways, as an archaeologist, really stumbled into it. And through various colleagues, I really got into machine learning.
00:25:29
Speaker
But it's really like this coincidence of various factors that brought me to this topic and into machine learning. Nice. Well, that's fun. I kind of like that, yet again, archaeologists are picking on some other industry, in this case industry, to find tools that we want to use. I always joke that the symbol for our field is a brick layers tool.
00:25:56
Speaker
Yeah. Not even something of our own. Exactly. Exactly. Now let's talk about the machine learning aspect of it. Actually, you know what? Before we get there, I'm curious because you said a small sample that you sent this company and they send back tens of thousands of results. What is the smallest particle that their scanner can actually find and suss out and identify? And then I've got a follow-up to that one related to the debitage. So that first.
00:26:24
Speaker
Yeah, so so this machine that ultimately with the help of my university, I was able to acquire measures everything between 35 microns. So this is what 0.0.
00:26:38
Speaker
0.35 millimeters and 35 millimeters. So it really covers a wide range. And for the microdabitage analysis, I limit that. So I filter out all the smallest and the largest particles. So I limit that to 0.125 millimeters to 6 millimeters. So theoretically, this machine could measure much more.
00:27:07
Speaker
I also have to filter it out because I ran, I mean, you know, when I got the machine for the first time, you know, I had no clue and I was just playing around with it. And the machine several times broke down because when I just used the entire range that it can measure, I literally ended up with millions of particles, specifically because I have tiny clay particles that just create types of data points.
00:27:35
Speaker
Yeah. Wow. I mean, that, I guess that would be expected. I guess that led to another question. I was going to ask how small you can call something micro devotage before it starts becoming, I mean, potentially natural. You know what I mean? I mean, how do you even know it was human created when you're getting that small?
00:27:54
Speaker
Exactly. This is exactly where I'm really working through these issues, because exactly as small as it gets, the rounder these microdabitage particles become, and they are very...
00:28:10
Speaker
they become very hard to distinguish from regular sand kernels that you have in soil samples. So this is really one of the issues that I'm working on right now, where I say if you go too small, it becomes very hard to distinguish between these two categories. And it doesn't make sense to call very small particles microdabotage, because it's really hard to distinguish them from regular soil particles.
00:28:34
Speaker
Yeah. I'm sure they experienced erosion much faster too. You know, the edges get knocked off, things like that, just because of their size, I would imagine. Yeah.
00:28:43
Speaker
All right. Well, let's take a break. And on the other side, yeah, we'll wrap up this discussion and maybe talk a little bit more about the algorithms and the actual methods that you guys used to do this. So we'll do that. In the meantime, I'll mention once again, check out arc pod net.com forward slash members to become a member of the archeology podcast network. We've got a Kelturo share event coming up. That's the Sunday after this podcast releases about underwater maritime archeology. So check that out. If you miss it, you can watch it. If you're a member anytime you want, we'll be back in a minute.
00:29:15
Speaker
Welcome back to episode two Oh seven of the archaeo tech podcast. And we are talking to Marcus here about this paper that is linked in the show notes. So go check that out. I won't read the whole thing out again, but we're talking about micro debitage machine learning and figuring all this out. So Paul, you had a question to bring in on the side.
00:29:31
Speaker
Yeah, I did. I was curious. Marcus, you were saying that size is a great filter that you use once things get too small. You don't want to count them because they're unreliable. Part of that is that they become rounder. Are there other thresholds that you use about the shape of the particles that you can use to filter or basically size and roundness, the main one?
00:29:55
Speaker
Actually, I mean, for the moment, and this is really like the ongoing project, I'm still trying to figure out, you know, what are good characteristics to identify microdabotage, such as to repeat one aspect, this particle analyzer that I'm using measures about 39 variables,
00:30:14
Speaker
And I say about because you can turn on specific measures and others. So you have a lot of variables to play with. So you can look into angularity, different types of length, width, and so forth. And this is one of the aspects that I'm trying to work through right now to answer seemingly basic questions. How do microdabitash look like? And how can I identify it in statistically useful ways?
00:30:43
Speaker
And so currently, I'm still using basically all these variables. And we are just like focusing on a few like for example, in this particular paper, transparency came out as the key variable, which probably for archaeologists to look at flints or like like
00:31:02
Speaker
five faces is pretty obvious, I guess, because normally flakes are relatively thin, so light shines through them. And the same applies to microdabitage. So for our machine learning algorithms, transparency was by far the key variable.
00:31:21
Speaker
Others came in, but this is also where these different algorithms then differed, which type of variable they preferred. But as I said, transparency was really the one that stood out among all four algorithms that we compared.
00:31:37
Speaker
You know, that's really interesting to me because you've got in the article, you used a handful of different algorithms, right? To these machine learning algorithms. But in order to make a machine learning algorithm smart and, and, and give you the results that it's, that it's going to come out with, you have to teach it. But if we don't really know, if you guys don't really know exactly how to characterize one of these microdebitage flakes, what kind of material are you giving the algorithm to say, this is right. And this is wrong.
00:32:06
Speaker
So what we do is, and this is why the data scientists on the team were so delighted, because we have experimentally produced microdabitash. So basically what we gave the machine is, OK, this is a sample of experimentally produced microdabitash. And this is what you have to as a training material. So this is how microdabitash looks like.
00:32:29
Speaker
And then we give it the archaeological soil sample and say, OK, where can you find this microdevitash in an actual soil sample? And can you figure out, are there similar particles in this soil sample? So this is the key part, that we have an experimental sample that can be used for training, and then we can apply it to an actual soil sample. I'm imagining somebody napping in like a radiological suit in a clean room or something like that, so you don't contaminate it with anything else.
00:33:03
Speaker
I bought like one of these 25-30 feet tarps and put it up in a small room and the flint knapper was then working diligently in front of the tarp and we just folded it up and put it all into Ziploc bags.
00:33:20
Speaker
It's been earlier, flint and obsidian, and we're talking about the training group. How much variability do you have between different kinds of stones that would be used? Because I can imagine that might be significant between jerks and flints and obsidians and whatever else. People are quartzite, whatever else.
00:33:39
Speaker
You know, this is again, you know, like, in a short, I would have to say, we don't know, because, you know, like, like, people have always assumed, as I said before, with the manual approach, you know, that every microdabitash particle should look like a regular flake.
00:33:54
Speaker
And nobody has really studied differences, for example, in material. And I'm just getting into that. Is obsidian the same from Flint? Are different sources of obsidian, do they produce the same type of debitage? So this is really what I'm really starting to throw the statistics out to see whether we can see these differences or whether they are really similar.
00:34:23
Speaker
But yeah, it's again, one of these simple questions that lead down a rabbit hole of, well, we don't really know. We have to jump into that and look into different sources, different raw material sources and compare. I'm also imagining different workshops using slightly different techniques or different tools might also affect that.
00:34:43
Speaker
For example, I have a graduate student who's interested in gender differences. I don't know whether you ever went to one of these nap-ins. I find them fascinating because they are basically 95% male. So this is often also what we think. When we think about stone napping, okay, well, these are male stone nappers.
00:35:04
Speaker
But again, do we really know that? And so I have a graduate student who looks for specifically female stone nappers to collect their debris to see whether she can find, for example, different differences among the debris of male versus female stone nappers. Again, an issue that simply hasn't been studied. And I'm looking forward to see that, whether she finds something.
00:35:33
Speaker
Yeah, it might be, not to be too on the nose, but literally microscopic differences between male and female, but that the algorithm could, could suss out. Yeah. Yeah. It had to be said. You know, what we also do, we do interviews with modern stone nappers. And for example, we had several male stone nappers who say, you know, female colleagues that they have, they can put as much pressure on the stone.
00:35:57
Speaker
So again, you know, this is not quantifiable, but you know, anecdotally, they say, oh, there are differences, how male and female stone nappers work. And you know, if we can quantify that, I mean, that, that, that would be fascinating. You know, something Paul said earlier made me think.
00:36:14
Speaker
about another thing that could be looked for during this process. You mentioned different workshops, different tools. Obviously, there's different things used to make stone tools, right? You've got antler, you've got bone, you've got other rocks, things like that. Are you guys close to analyzing? Because those are going to break up as well, and you'll produce micro bits of debris from those. So are you able to find some of that in your samples, or are you looking for that kind of stuff yet?
00:36:40
Speaker
Not yet. I mean, but you bring up a really good point. I mean, different napping techniques. So another thing that we are working with modern stone nappers is to ask them, well, or what we do is we record the specific tools that they're using. And we also ask them, you know, could you use an antler instead of like, like an issue stick or whatever they're doing? Yeah. And that's, for us, really a critical part. Again, something that has not been studied. The other aspect
00:37:10
Speaker
I really, that's a really good idea. I mean, I haven't really thought about antler debris that could end up in a soil sample. Yeah, I mean, it's possible. I mean, honestly, I haven't talked with them, so I don't know. Yeah, there's so many directions this could go, really. It's pretty interesting. Bringing this all together. And I know you're really in preliminary stages on figuring out what we can actually find out from all this stuff. And we talked about this a little bit in segment one, but really,
00:37:41
Speaker
What have you learned from some of the things you've analyzed by studying microdebitage that you wouldn't have known otherwise about a site or about a technique or something like that? What is this really telling you that's different from other techniques?
00:37:55
Speaker
Well, I mean, one aspect is simply that I see a path forward to analyze hundreds or even thousands of soil samples. So the traditional manual approach has been really limited in the sample population. I mean, I think, you know, I did a review of the literature. I think the maximum number that I saw is about 160 samples. So people have been really
00:38:21
Speaker
Dealing with like often just a single room in a building or like a residential group but it has been really hard to use micro debitage analysis at least the traditional micro debitage analysis on the level of an entire archaeological site.
00:38:38
Speaker
or even like a region. And that's really what I'm getting very excited about. I mean, one of my previous graduate students, we are still collaborating, she, for example, took soil samples from an entire archaeological site in the Maya lowlands.
Scaling Microdebitage Analysis with Machine Learning
00:38:53
Speaker
And we are now still, I mean, these are now 500 and something samples. So we are still in the process of looking through them.
00:39:00
Speaker
But this is for me the first where I can say, OK, we can probably talk about how stone napping look like in an entire ancient city. You know, do we have different neighborhoods of stone nappers? Did they work with different stone materials to not compete with each other? And so they are really, for me, fascinating issues about the ancient economy of stone napping that previously could not be answered. And so that's what I'm really getting excited about.
00:39:28
Speaker
to scale up this particular approach and hopefully get insights into larger economic patterns in ancient societies. As we're wrapping up here, anytime somebody comes up with an efficient way to do something and you bought the equipment, you have the particle analyzer and you're developing the algorithms for this. Since this paper came out not too long ago, have you been approached by anybody else to say, hey, can you analyze my soil? Can you do this? You guys might be setting up a business here.
00:39:59
Speaker
I'm not yet there, but I'm getting really, I mean, people are getting really fascinated by it. And I should say, I mean, I'm also getting requests from archaeologists who suddenly hear about the machine and have completely new ideas. For example, a colleague in my department, his wife works with, she's interested in the development of maize, corn, across time.
00:40:25
Speaker
And so she, for example, excavates mittens where she has different layers from thousands of years back until more recently, different chart maze kernels. And so she asked me whether she could run these hundreds of maze kernels through this particle analyzer to study how the individual kernels, how they differ in size or dimensions across time.
00:40:51
Speaker
So the point is, you know, you can apply this particle analyzer not only to microdabitash, but you know, you can come up with really interesting new ideas about how to look at archaeology. So yeah, I mean, I assume that there will be more people knocking on my door and then asking about that. Could be an easy way to fund your research. If you would be that wealthy terminus archaeologist, you know, our funding limits.
00:41:22
Speaker
Indeed, indeed. Well, where do you go from here? What are you looking at next? Without giving away any papers or anything that are in production, what kind of questions are you hoping to answer after this? One thing is really that we tried to nail down quite a few of the questions that you just brought up. For example, is microdébitage the same across different raw materials?
00:41:46
Speaker
Do we see differences among male, female, or stone nappers, let's say, in the experience? A stone napper who has five years experience versus a stone napper who has 20 years of experience. So we are really, my team, and I should really say it's a team. I mean, it's not just me. I mean, there are graduate students and others who are working on that. We're really trying to answer these questions through this.
00:42:12
Speaker
So this is really one of the biggest hopes to find answers to these questions. Another thing that I'm actually what I did this summer, I mean, I came back from Israel. So for example, I'm also using the particle analyzer to study ancient mortars.
00:42:29
Speaker
So we got an NEH grant to look into the different components that ancient peoples put into mortars. And so I collected mortar samples and now we started already to run them through the particle analyzer. And we hope to be able to find out how mortars from different cultures differ in the composition of their mortars.
00:42:54
Speaker
Nice. Nice. That's really cool. In Iraq, Chris, you know well. I've been working there for the last couple of years. Yeah. And one of my colleagues is doing all sorts of soil cores, looking at environmental change and the advance in the retreat of the waters there. And I could see this being maybe a way of... He looks at very small faunal remains to tell us what kind of environment it was. And I could see this being... Or this set of tools being a way of looking at
00:43:22
Speaker
finer grade changes between one layer or another than he can currently do, just cause it's too slow to do it manually. No, I mean, and you know, if anybody's interested, I mean, you know, you can approach me. I mean, people can run their samples and go with that.
00:43:38
Speaker
Yeah. Awesome. All right. Well, this has been super fascinating. I hope we can follow you on this and really stay on top of some of these changes. We've talked about machine learning. We've talked about AI and doing some stuff like that before. And a lot of it is really fascinating, but this really hits one of my sweet spots. As a cultural resource management archeologist out here in the West of the United States, we deal with a lot of lithics and sometimes
00:44:04
Speaker
You know, we're on a site and we're like, Oh, okay. So we found like 20 flakes. That's pretty much it. Here's the story of the site. When in reality, there's 200,000 micro debitage flakes because somebody already picked up the tools. Somebody already even picked up some of the flakes and the real story of the site hasn't been told. I think that's just super interesting from a, from a management standpoint. You know, we, we say the site is not significant because we didn't find anything big, but what about all the small stuff that could really tell you about the site?
00:44:31
Speaker
I'm looking at it with interest because it's not about the tech. It's not about the stats. It's about how those are being used to analyze what people did, which is, you know, as anthropological archeologists here, that's really what it's all about. Absolutely. Yep. Yep. Exactly. All right. Well, Marcus, thank you. We really appreciate it. And always feel free to come on the show when you're, you know, ready to talk about some more fascinating tech stuff that you guys are doing and just chat about what you're doing.
00:44:57
Speaker
Thank you so much. Thank you for having me. Thanks for coming on. Take care.
00:45:06
Speaker
Thanks for listening to the Archaeotech Podcast. Links to items mentioned on the show are in the show notes at www.archpodnet.com slash archaeotech. Contact us at chrisatarchaeologypodcastnetwork.com and paulatlugol.com. Support the show by becoming a member at archpodnet.com slash members. The music is a song called Off Road and is licensed free from Apple. Thanks for listening.
00:45:32
Speaker
This episode was produced by Chris Webster from his RV traveling the United States, Tristan Boyle in Scotland, DigTech LLC, Culturo Media, and the Archaeology Podcast Network, and was edited by Chris Webster. This has been a presentation of the Archaeology Podcast Network. Visit us on the web for show notes and other podcasts at www.archapodnet.com. Contact us at chris at archaeologypodcastnetwork.com.