A claim to have ~solved the huge difficulty of Roman Imperial die studies.

[Deinomenid]

I am off-piste here, but  I attended a lecture at Oxford yesterday where researchers said they had made very substantial progress in die-linking the staggeringly vast  numbers of Roman  Imperial  (and  Republican) coins. I have seen other similar claims, but with the presumed gentle blessing of the Ashmolean, the introduction by the new Keeper of Coins, various luminaries etc nodding away I'm assuming  it was not a fly-by-night effort.

The quick version is that through extensive automated image analysis combined with numismatic skill to help the process (relevant landmark points on coins, manually checking the small percentage of queries etc) and crucially through specifically Bayesian modeling they have been able to slash the time taken to perform a given Emperor die study. The view is that Esty, Bransbourg, Crawford etc either take or explicitly assume a frequency based approach, with its many suppositions with the result that manual die studies would be needed instead but as  Crawford said "The practical problem is that to count all the dies used to strike during the Republic would be the work of several  lifetimes".

It was all very polite, but their view was the frequentist approach especially with subjective overlays by some (die life, wear, inferences over historical periods (Crepusius?), vagaries of the archeological process etc) was  substantially inferior.  I had seen a couple of articles  on this online in the past (they are Germans working for the National University of Singapore/Yale/Cornell) but had been sceptical until I saw this.

Their point seemed to be just to try to make the backbreaking work of die studies easier and therefore give better data for everyone to go and have the arguments over,  though reading between the lines  their hope seemed to be to demonstrate the  true extraordinary scale of monetization of the economy with some sort of birth of capitalism inference. There were lots of questions - nothing really challenging it as an approach - & a couple of comments at the end along the lines of Bob's used this for his work on post-reform Neronioan aurei production but as a mere Sicilian coin fanboy this was all a bit err Greek to me.

I just flag it here in case this comes up more generally a) that there is  hope for all of you yearning deep in your souls for effective die studies and b) to take this approach with possibly less scepticism as it seems  likely  it will be used more in the field. The main oddity to me was that they did not review any  specific study they had completed... 

.........................

Some links  in case of interest.

https://arxiv.org/abs/2112.00290 which calls it unsupervised deep learning for die analysis, but as above there *is* pre- and post- supervision.

One of the 2 authors - https://fass.nus.edu.sg/events/fass-brown-bag-seminar-quantifying-roman-the-money-supply-towards-a-genealogy-of-capitalism/

Cohesion and Repulsion in Bayesian Distance Clustering  https://www.tandfonline.com/doi/full/10.1080/01621459.2023.2191821  (they said the academic approach came  initially from life sciences)

[Heliodromus]

I don't get the impression that the frequentist vs Bayesian distinction is really meant to be a putdown of manual die studies, but rather just a technical detail (along with their Gaussian process-based keypoint features/comparisons) of how they approached such a complicated task using machine learning. The distinction is that in a frequentist approach you're ultimately deciding whether two points agree or not, whereas in the Bayesian approach it's all about probabilities which are used to update an underlying hypothesis of overall similarity or not.

I'm not sure why they'd be referring to Esty (@Valentinian) here, since the job of clustering (i.e. assigning specimens to dies) is entirely separate from what you then do with that data. For anyone interested in this aspect of estimating die populations from specimen counts, I recently came across this interesting paper puts a bit of a different twist on the meaning of singleton and low-frequency specimens by modelling coin production as a combination of high output and low output dies - the low output dies having suffered infant mortality.

https://www.sciencedirect.com/science/article/pii/S0305440321000765

 

 

[Deinomenid]

1 hour ago, Heliodromus said:

putdown of manual die studies

My strong  impression was absolutely NOT that they were putting down manual die studies. It was  just that it was a near  Herculean task to perform them in any number  given the vast quantities.  However, I have no pony  in the game, I was  just reporting  on what they said. Usually I'm rather suspicious of a new and improved anything.

I've gone through manual  die studies at some  length in some of "my" areas, and they are often difficult to even follow, and that's with relatively  tiny number of coins in eastern Sicilian mints 🙂.

Also there was nothing  but politeness to other approaches  - indeed Messrs W. Esty and A. Turing were mentioned  in the same breath!

 

Edited March 1 by Deinomenid
typo

[Rand]

Thank you for pointing out this work; it is interesting.

I wish the AI approach could solve the problem. It may help in the future, but it is unlikely at the present state of development.

Reflection from my (humble) die analyses

• Getting images/records is far more time-consuming than die-matching itself.
• Some coins are damaged, and there is not enough surface to identify dies.
• Photos could be of inadequate quality.
• Dies could be repaired, with repaired dies looking more like a different die.
• There are often multiple entries from the same coin, and it may be challenging to be sure if the old photo of a cast is a different coin or the same but affected casting and cropping.

While the analyses are interesting, a major objective of die studies is to have a documented list of all produced dies (and their combinations) as a numismatic reference. Their number on its own is exciting but of lesser importance. If the AI methods produce individual die lists, they would still need manual validation.

The tools I have tried so far did not work particularly well.

[Rand]

My favourite AI tool for die studies - sadly only works on my own coins.

PS. This post is meant to be humour. It does not mean a loop is an AI tool - it is not.

Edited March 2 by Rand

[Kaleun96]

25 minutes ago, Rand said:

Thank you for pointing out this work; it is interesting.

I wish the AI approach could solve the problem. It may help in the future, but it is unlikely at the present state of development.

Reflection from my (humble) die analyses

• Getting images/records is far more time-consuming than die-matching itself.
• Some coins are damaged, and there is not enough surface to identify dies.
• Photos could be of inadequate quality.
• Dies could be repaired, with repaired dies looking more like a different die.
• There are often multiple entries from the same coin, and it may be challenging to be sure if the old photo of a cast is a different coin or the same but affected casting and cropping.

While the analyses are interesting, a major objective of die studies is to have a documented list of all produced dies (and their combinations) as a numismatic reference. Their number on its own is exciting but of lesser importance. If the AI methods produce individual die lists, they would still need manual validation.

The tools I have tried so far did not work particularly well.

A lot of those points would apply to manual die analysis using only photos too, which a lot of die studies these days rely on. But I also disagree with some of them. 

For example, getting records/images can be as fast as a script that scrapes a database or polls an API. If you're recording them by hand then yeah it will take a long time but you can also download 10k records in a matter of minutes depending on the source and data quality.

The problems you mention like multiple photos of the same coin, poor image quality, or tooled coins do mean manual validation or data processing is always required or at least recommended. But a die matching model could also theoretically flag identical coins and once the model has identified coins by their dies, it's much easier to manually filter out duplicates or invalid results. These are problems I've faced when die matching manually from photos and having a model go through first and do some clustering would speed up the process significantly. 

I think these models are absolutely the future and will unlock die studies previously deemed impossible/too tedious. Manual validation will remain part of the process but 90% of the work could be done by models in a fraction of the time a human can do it and then the remaining 10% is pre-cleaning the data and validation. 

Edited March 2 by Kaleun96

[Rand]

14 minutes ago, Kaleun96 said:

A lot of those points would apply to manual die analysis using only photos too

As I said, this was my reflection on my (manual) die studies, so they all apply to manual studies.

Web scraping is easy, and lots of illegal scraping is going on. You would have to respect the rules of image owners, who are likely to forbid this and get their permission. But even if you ignore the rules, you will not be able to access many coin repositories that need personal access or are not online at all. Many catalogues and books have not been scanned and, if so, may not be free for scraping but free for manual studies.

My thoughts are about the current state of AI, and I am sure it will be a useful tool as it evolves.

[John Conduitt]

It must surely help with large die studies. Even analysing fifty coins is arduous, and it gets exponentially longer with each dozen as you have so many coins to check against. If there are 1000 coins it's impossible.

On the other hand, manually checking the AI's output would be that much harder and longer if you haven't already had to look at every nuance on every coin to work out what the important differences are. It would also be difficult to say what the 'landmark points' are without first doing this.

Perhaps it works when you manually sort the first couple of dozen dies to get your eye in, tell the AI why they're different, and it does the rest.

[Rand]

4 minutes ago, John Conduitt said:

If there are 1000 coins it's impossible.

Actually, sorting 1000 coins manually has not been a big deal after I worked out a set-up for type-varies-die records. I have sorted more than 3000 coins of Anastasius from about available 10000 records. About half of the records still need to be sorted due to continued data collection. Sorting coins is fun, but getting records is far more time-consuming to do meticulously. I have c. 14,000 records of auctions, books, museums, and so on, which I have checked for those coins. Many of them have not been or are no longer online, and a few are from museums without permission to publish without further approvals/payments.

[Edessa]

ACSearch's image search works pretty well but you can even just drop a picture into Google Lens and sometimes find a few die matches. I have only fooled around with it as I don't think I have the patience to attempt a real die study.

[Kaleun96]

On 3/2/2024 at 1:41 PM, John Conduitt said:

It must surely help with large die studies. Even analysing fifty coins is arduous, and it gets exponentially longer with each dozen as you have so many coins to check against. If there are 1000 coins it's impossible.

On the other hand, manually checking the AI's output would be that much harder and longer if you haven't already had to look at every nuance on every coin to work out what the important differences are. It would also be difficult to say what the 'landmark points' are without first doing this.

Perhaps it works when you manually sort the first couple of dozen dies to get your eye in, tell the AI why they're different, and it does the rest.

You shouldn't need to tell the model what to focus on per se, it can work that out itself. There are tweaks you can make so it doesn't focus on the wrong thing or you can add weights to certain features etc, depending on what type of model you employ, but theoretically you wouldn't need to be that familiar with the coinage. 

Being familiar would help with the validation and optimisation of the model though.

You may need some sort of training dataset to validate the model on and that would require manually die matching some examples and then checking accuracy but it's probably possible to do this completely unsupervised too. You feed the model the images and it learns the differences, you then make tweaks to make sure it's identifying unique dies rather than examples, and then manually check its results after. I don't know whether supervised vs unsupervised models are more common for die matching but I don't see why both wouldn't be possible. 

[Kaleun96]

On 3/2/2024 at 11:04 AM, Rand said:

As I said, this was my reflection on my (manual) die studies, so they all apply to manual studies.

Since you said AI is unlikely to help currently and then listed problems with die studies that would also apply to AI models, I assumed that those were reasons why you think AI models would not help currently. But if those are just general reasons, what are the reasons for why you think AI models specifically aren't yet useful?

I think these models are an emerging technology when it comes to die studies but I don't see any inherent limitations that mean they aren't yet useful. The main problem is that most people capable of a die study aren't statisticians or proficient in coding and we're probably a little ways off from a generic model being useful for most coin types, meaning bespoke models trained or optimised for different sets of data is still required. 

On a dataset of ~450 coins, I've probably spent half the time invested just identifying dies and as mentioned the time to check one die increases the more examples you add. A model could check a new entry's dies against the rest of the database in milliseconds, and at the least probably rule out the vast majority of possible matches and leaving just a couple close matches. The only reason I haven't tried to make a model myself is that it requires upfront investment to learn that aspect of machine learning and I've already done the bulk of the die matching. It's been in the back of my mind to try one day, maybe when I start a new die study from scratch. 

[Rand]

Thank you @Kaleun96 and I do agree with your points.

AI can be and surely will be helpful with die studies, but not in a similar way like, say, in medicine. Apart from data access, the major challenge is the nature of coins as the object of AI.
Unlabeled AI tools are unlikely to work: 
We do not just want to know the number of dies - we know them sorted by types and varieties and exclude fakes and objects that look like coins but are not.
Coin images will need some labelling, starting from 'this is an ancient coin' but ideally having a well-labelled training dataset detailing individual types. And here we have 100k+ ancient coin types for labelling! This contrasts with the output of often a small number of dies for a particular type. Developing an AI model may take far more time than doing this manually (of course, there would be template models to use).

Also, there is a finite number of dies for ancient coins, which will not increase once they have been documented (not any time soon). Once a die list is known for each coin type, like some Syracuse masterpieces, do we need AI to link a new coin? A dealer or museum would add labels, and a statistical tool would update the die projections.
Photo images do not have an agreed standard and would be much more difficult to use than, say, retinal images or mammograms. This may be solved, but it would need a good skill set.
I hope I do not sound like an AI sceptic. I am not.

[Heliodromus]

22 minutes ago, Rand said:

Photo images do not have an agreed standard and would be much more difficult to use than, say, retinal images or mammograms.

Yes, it's not apparent what quality of images is currently needed by this technique, but I'd expect that for a given quality of data, machine learning will do better than humans since "perception" is an area they excel at, and this has been the case in other areas of application.

The specific approaches used by the Heinecke et al paper linked by @Deinomenid do seem to mitigate the issue of poor data (whether due to photography, poor strike, worn coins, etc etc). In particular the Bayesian approach let's them make the most likely die classification in the face of uncertainty (much better than a human would do just by eyeballing it). They automatically determine the maximally informative comparison points via their Gaussian process based approach, which again seems at least as good as what a human would do in identifying device alignment points that are easy to compare (but perhaps NOT maximally informative).

I don't see how manual labelling would help here. That is normally a step that might be used during training of an image classifier, for example if the goal is to develop a classifier capable of recognizing coin types. I assume the coins being presented for a die study have already been determined to be of the same type. If there was a value in taking a set of photos of mixed coin types (an entire hoard, say) and doing a multi-type die study, then the way to do this would be to first use a separately trained classifier to separate into types, then use this specialized die comparison technique for the die identification.

I'm somewhat familiar with ML/AI - I work in software, and have previously written my own neural net framework (think Torch for C++).

 

[Rand]

1 hour ago, Heliodromus said:

That is normally a step that might be used during training of an image classifier, for example if the goal is to develop a classifier capable of recognizing coin types.

Yes. This is correct. I have been referring to the broader using of AI as a solution for die analyses and classification questions in numismatics. AI will help many individual tasks withing this ambition. 100K+ types of antient coins is one of a major challenges.

 

Please note, the cited study used 'visual validation' as the grand truth (please see my AI [an eye] tool above). Applying the method to Nero's obverses makes sense to me (>2000 images for 1135 denarii). It would be less useful for multiple reverses, which could be easier to handle manually. Each of the coins passed manual labeling (this time selection as a particular type of obverse). Imagine doing this for all antient types.

I showed these coin photos before. Humour aside, I need to know if this is the same coin or not for die number projections. There are no other coins known to me of this variety 'helmet with a cross on PERP type'. The seller did not have info on the provenance. If this is the same coin (which I think it is) a bonus would be a proventance to a famous Ratto sale.

Could the proposed AI method help?

Edited March 4 by Rand

[Heliodromus]

I think the way to handle this would be as a separate "data cleaning" step prior to the die clustering.

In my experience the hardest photos to disambiguate are when at least one of them is from an old catalog where it's a plaster cast being photographed and therefore the shape of the flan appearing different can't be relied on. If the differences between two images of the same coin are just down to photography/lighting then that seems easier to handle by creating custom datasets for this purpose (same coin photographed many different ways). In either case a pre-filtering/cleaning of the dataset based on a separate model trained for that purpose would seem to be part of he solution. Of course this wouldn't always be necessary if one knew a priori that the coins were all distinct (e.g. a hoard study).

Another option when duplicate data is suspected would be to run the die clustering model first, then just manually inspect the identified double die links, which should be a far smaller set of the coins than having to manually clean the dataset before hand.

The best way to think about ML (I hate the overused term AI - maybe only generative AI  - ChatGPT etc - deserves this label) is as automation. It's generally a way to train a computer to do something that a human can do. If a human can determine if two images are the same coin or not, then a model can probably be trained to automate this and do just as well.

 

 

Edited March 4 by Heliodromus

[Deinomenid]

4 hours ago, Kaleun96 said:

most people capable of a die study aren't statisticians or proficient in coding

The study I reference indeed has this issue. So there are 2 of them. One more of numismatist, the other the stat and coder person. They co-presented. It sounds like  it would be a bit of a curate's egg if  it were  just one of them...

 

1 hour ago, Heliodromus said:

They automatically determine the maximally informative comparison points via their Gaussian process based approach

Exactly! They spent some  time on this. Related, there were questions about  how it could work given not every coin is incredibly perfectly photoed and  their answer was to filter DOWN to an acceptable but informative level.

 

4 hours ago, Kaleun96 said:

theoretically you wouldn't need to be that familiar with the coinage

Yes, except for oddities that get thrown in by mistake, but these are flagged and manually reviewed. This was  one of the reasons they said a given study would  take 2 weeks rather than 90 seconds on a laptop.

 

8 minutes ago, Heliodromus said:

from an old catalog where it's a plaster cast being photographed

This was a focus of some concern and there was a work-around they had. I am sorry  I can't remember  what it was in all cases,  but in some it was in the small manual section.

 

I'm in over my head here, so should  probably shut up. One last incredibly  broad point of  interest to a layman like me was that the stats/computer code man said specifically that some of the  initial impetus for their work was from  studies in life sciences (as above)  on working  out  how many - his example - insect species remain  undiscovered in the Amazon.

 

 

[Kaleun96]

On 3/4/2024 at 2:09 PM, Rand said:

Thank you @Kaleun96 and I do agree with your points.

AI can be and surely will be helpful with die studies, but not in a similar way like, say, in medicine. Apart from data access, the major challenge is the nature of coins as the object of AI.
Unlabeled AI tools are unlikely to work: 
We do not just want to know the number of dies - we know them sorted by types and varieties and exclude fakes and objects that look like coins but are not.
Coin images will need some labelling, starting from 'this is an ancient coin' but ideally having a well-labelled training dataset detailing individual types. And here we have 100k+ ancient coin types for labelling! This contrasts with the output of often a small number of dies for a particular type. Developing an AI model may take far more time than doing this manually (of course, there would be template models to use).

Also, there is a finite number of dies for ancient coins, which will not increase once they have been documented (not any time soon). Once a die list is known for each coin type, like some Syracuse masterpieces, do we need AI to link a new coin? A dealer or museum would add labels, and a statistical tool would update the die projections.
Photo images do not have an agreed standard and would be much more difficult to use than, say, retinal images or mammograms. This may be solved, but it would need a good skill set.
I hope I do not sound like an AI sceptic. I am not.

"Unlabeled AI tools are unlikely to work ... Coin images will need some labelling, starting from 'this is an ancient coin'"

You of course need to do some data cleaning beforehand, as you do with any kind of die study, but this is not what is meant by "labelling". Labelling would be identifying the dies in advance and validating the performance against a test set. When we're talking about training an unsupervised model, we mean feeding it images of coins you want it to cluster together based on similarity. So there's no step where you need to tell the model "this is an ancient coin" unless you plan on feeding it images of bananas or cats.

" And here we have 100k+ ancient coin types for labelling!"

Perhaps there's a misunderstanding. From your reply I get the impression that you're thinking of some generic AI model that is capable of producing die studies for any coin you feed it. That's not what I have in mind nor is what most people have in mind. Maybe one day it will be possible to just feed a massive set of images to a model and tell it to identify all the dies but that's neither necessary for ML/AI die identification to be useful nor the intended use case currently. 

What the current models are proposing to help with is the following scenario: you have 10,000 images of coins of a particular type or series that you want to do a die study for. You scraped/downloaded these images from some database that contained other information, such as the ruler and possibly type attribution. You've gone through and cleaned out the irrelevant results, tidied up some of the attributions, and now you have 9,700 coins to perform a die study on. The next step is either to do a manual die study or feed these images to a model. You've spent a long time cleaning the data but the die study itself is going to take even more time. This is the point the model steps in. You'll want to train the model on specific types or groups of related types (i.e. similar obv/rev designs). But this step is not some additional step that only an ML model requires - you have to identify the types in advance for any kind of die study. 

So in short: no one is proposing to feed an ML model a trove of random coin images. The data cleaning and organising aspect of die studies isn't going anywhere anytime soon. You still need to know some basics like the dataset contains coins of the same emperor/ruler/iconography and are either the same type or multiple similar types. That data cleaning and processing does take awhile but the actual identification of dies takes a significant amount of time too and it's this specific part that supervised or unsupervised ML models can help with. That being said, you can absolutely have fakes/duplicates/misattributions in the dataset and not cause any significant problems - these will be easy to identify and remove once the model attempts to cluster them or will be part of the general post-clustering validation.

You can find an example of an unsupervised model for die identification here: https://www.researchgate.net/publication/356710176_Unsupervised_Statistical_Learning_for_Die_Analysis_in_Ancient_Numismatics

Once a die list is known for each coin type, like some Syracuse masterpieces, do we need AI to link a new coin?

I don't exactly get your line of reasoning here. There's not a dichotomy where we have to commit to either only using AI or never using AI. If all the possible dies are known then whether you need/want to use AI to identify the die of a new coin is entirely up to you. It may help for types with a large number of dies, and it'll definitely help when processing a new hoard even if there are no new dies in the hoard. Picture a hoard of 30,000 Athenian tetradrachms - would you prefer to go through the list of, say, ~1000 known dies and identify the correct die for each one, or would you rather a model do that work for you?

Edited March 8 by Kaleun96

[Rand]

23 hours ago, Kaleun96 said:

You of course need to do some data cleaning beforehand

For typical data projects, data cleaning takes most project time (80% has been quoted).

Yes, we referred to different ML/AI uses - I referred to AI as an universal solution of die analyse - my view the technology is not ready for this yet.

We both agree that it is helpful, in some cases, for issues when many similar dies were used. Athenian tetradrachms are a good example of an issue where an ML/AI approach is desirable and may be the only solution. There is no point or possibility of training an AI/ML model for EID MAR aureus when only three or four examples are known.

There will be a few situations when ML/AI may struggle and will need additional input from researchers - when:

• Similar dies were used for gold and silver issues, such as a few aurei and denarii, but only black-and-white photos are available.
• Similar styles were used on coins of different diameters (some late Roman Bronzes)
• The same dies were used for coins of different thicknesses, for example, some thalers and their fractions/multiples. 

https://www.acsearch.info/search.html?id=1478082

https://www.acsearch.info/search.html?id=3166387

https://www.acsearch.info/search.html?id=2708751

The above is now an issue for analyses of preselected images of coins of the same type.