Tldr (the original paper); Fine tuned ResNet-50 and VGG-16 classifiers are pretty good at identifying ancient pottery fragments found in the South Western USA while human expert seem quite poor at it.
Also, archeology papers have nice maps and photos. Open access paper here:
> A neural network would likely do reasonably well at sorting other types of decorated pottery, but so-called plainware—ceramics without any visible decoration or markings—would probably be a bridge too far.
I am think it might well work with microscopy. A 224x224 pixel zoomed out piece of plan white pottery isnt going to have any features worth extracting, but there could be a wealth of features at a ultra high zoom levels.
It's curious to see that they chose ResNet50 as an architecture. Although there was EfficientNet mentioned in the References.
No word on frameworks, although they used GradCAMs for Keras/TF. So, I am guessing Keras.
They are using voting. And achieving 98% accuracy in 10 votes. Is the model overfit?
This might be novel in archaelogy, but this is a very very easy task to perform for a DL practitioner. But they took their time to read a lot about ML interpretablity, tSNE, transferable features, etc and they cited them. Wanted to be rigorous.
Archaeologists aren't always the most technically adept people. This is just a demonstration paper to promote acceptance of ML techniques in the somewhat conservative field of southwestern archeology. I'm not a big fan of this paper, but they made a good attempt to address reasonable concerns.
> but they made a good attempt to address reasonable concerns.
I am guessing that, too.
It is weird how you can do extremely easy things in a field of application that is novel to the method and get huge amount of praise. I bet they will be called to speak on many archaelogy conferences to talk about it.
I am not jealous at all. And I truly respect the people behind the paper for doing something that no one else is doing. They are at least doing something new. And they deserve praise from people in their fields.
It's a well-known criticism among archaeologists of our field that we tend to borrow a lot more methodologies than we export to other disciplines, so this case isn't unique. There's been a lot of interest in ML applications, but personally as a former archeologist now working at an ML-heavy company in the bay, I'm a bit more skeptical. Just my 2c though.
I took a single archeology class in college maybe secretly hoping for a dose of Indiana Jones. Instead I got potshards, so many potshards. Sounds like a reasonable problem for a neural network.
Maybe a project like Zooniverse [0][1] could be used to gain even more leverage. Scientists would create a few labelled examples that are used to train human volunteers. These volunteers then classify enough images that can be used to train an image recognition system at which point classifying these images becomes almost free.
I was trained in Archaeology as an undergrad. I have considered neural nets as a possibility for classification tasks that may otherwise be time consuming. But the material still needs to be collected, cleaned and catalogued by human hands. There are steps along the way where skilled humans can classify objects. The neural net might be a way to confirm or flag human decisions but I would not rely on it as a primary means of classification. The things that are interesting are on the edges of any classification. Transitional artifacts or anomalies. For these you will need human interpretation. Also, I do begin to wonder why we would even bother if no human will be interacting with a given artifact, why not leave it in the ground undisturbed? If that’s even possible. But I start thinking, with the ease of AI, if we are painting ourselves out of the picture. Like using GPT3 to respond to emails as a solution to the problem of too many emails. And imagine the respondent doing the same.
Anyone know of digital archaeology projects in this vein that could use avocational volunteers? I love the field and think there's a ton of potential, but sadly grad school isn't in the cards right now.
My SO is an archaeologist and I work in ML, so naturally we've discussed how the two could be combined. This is one of the applications we've discussed as well as sorting/categorizing bags of bones and bone fragments into animal types (zooarchaeology is her specialty so this is a common task). As the article seems to indicate, the main hurdle for these tasks is data collection and annotation, especially since even experts often disagree on the classifications. There have been a number of times where my SO had to discard the data she was using after getting access to the materials they were gathered from and realizing that the researcher categorizing the animal bones had no idea what they were doing.
22 comments
[ 3.3 ms ] story [ 68.8 ms ] threadAlso, archeology papers have nice maps and photos. Open access paper here:
https://www.sciencedirect.com/science/article/pii/S030544032...
I am think it might well work with microscopy. A 224x224 pixel zoomed out piece of plan white pottery isnt going to have any features worth extracting, but there could be a wealth of features at a ultra high zoom levels.
Would be amazing to see 3D models from photogrammetry added to the mix as well.
No word on frameworks, although they used GradCAMs for Keras/TF. So, I am guessing Keras.
They are using voting. And achieving 98% accuracy in 10 votes. Is the model overfit?
This might be novel in archaelogy, but this is a very very easy task to perform for a DL practitioner. But they took their time to read a lot about ML interpretablity, tSNE, transferable features, etc and they cited them. Wanted to be rigorous.
I am guessing that, too.
It is weird how you can do extremely easy things in a field of application that is novel to the method and get huge amount of praise. I bet they will be called to speak on many archaelogy conferences to talk about it.
I am not jealous at all. And I truly respect the people behind the paper for doing something that no one else is doing. They are at least doing something new. And they deserve praise from people in their fields.
[0] https://en.wikipedia.org/wiki/Zooniverse
[1] https://www.zooniverse.org/
Maybe if imaging pot sherds is part of your standard procedure it'd make save time. But the imaging itself could be quite laborious.
In any case, this relies on a pre-existing model--trained on labeled data.