Feasibility Of Computer Grading

No more possible than it would be for a computer to judge famous works of art based on colors, brush strokes, etc.

Some things are simply too subjective, and too reliant on personal aesthetics, to be judged by computers. Computers are leagues better at technical and scientific analysis.

Even with all the leaps-and-bounds advances in AI, machine learning, and neural networks, the best AI's in the world can recognize an individual's face in a crowd of thousands, but can't tell you if the face is ugly, beautiful, or in-between.

Sure, there are aspects of a coin a computer could measure: luster, reflectivity, color/toning, depth of relief, surface imperfections, weight, diameter, alloy content, etc. With a lot of money and a much greater amount of time and "learning" they could even be trained to look for varieties such as overdates, die doubling, RPD's, and so on.

Where it all falls apart is in the question of "degrees." The computer can analyze two coins and tell you that one coin is 8.2% more lustrous than the other, but that the surface relief is 6.1% lower in key measurement spots for the coin type; that one coin is 60% reds, 18% oranges, 20% browns and 2% gray-black vs. another at 40% reds and 60% browns; so on and so forth.

What it can not do currently, and might not ever be able to do, is to tell you which coin is more likely to be pleasing to YOU. It also can't assign a numerical grade based on its analysis, e.g. how much of a % of a coin's fields can have minor nicks <0.2mm deep and how many of those nicks can be present to lower the grade from MS-65 to MS-64. It also fails in the face of conflicting data that we, as humans, assign arbitrary statistical weights to: is a coin with flawless fields, great luster, and a terribly weak strike higher grade than one with slightly nicked fields, average luster, and a razor-sharp strike?

The eyes of a human, combined with the knowledge of a human brain, are far better graders -- and far more trainable -- than any computer currently is, and this is not likely to change unless we can train AI to recognize eye appeal, an inherently subjective/aesthetic determination which is often at odds with strict logic.

I'm not sure why some want to replace every job with computers.

Computer grading isn't going to happen. It's been tried before and failed and would fail again. Even computer grading would just be subject to the subjectivity of the coder and much easier to trick.

What computers are being used for by PCGS at least is an AI program that helps with Counterfeit Detection and doctoring, yes some of the fakes have gotten that good.

The internet has always made that seem much easier and more common than it really is.




Grading is not and never has been a precision system of measurement. Grading is a subjective evaluation of condition/preservation. Nothing that involves an evaluation of condition is a precision act.




You're comparing apples and duck eggs here your example is so far from a comparable.

Couple of points.

First off I am not suggesting this is a preferred alternative I was just questioning if it was feasible. Given that PCGS, at a minimum, has tried some variation of this then I assume other, more qualified individuals than me, are also asking this. To me PCGS, and NGC, are the "gold standard" of grading services so the very fact PCGS utilizes some sort of computer grading and are working on improving it tells me that more of it will come into play down the road. To me it is like self driving cars which I think are being made for idiots but despite what I think are probably coming eventually anyway. I just won't buy one.

Second aesthetics does come into play when collecting, and grading, coins. But to a lesser degree than with, say, grading a painting. Almost all modern era coins, from the last few hundred years forward, are made by mechanically pressing a planchet between two dies. So it isn't the same as a painting as it is, at the end, a mostly mechanical process. Even the subjective part, if one coin has more "eye appeal" than another, is just that, subjective. I like some coins and not others due to the design on them or the color of them. A computer could never measure that. But if you plug in a large enough set of variables and a complex enough system of measurement yes a program probably could be written that made thousands of different measurements and could devise a grading program. In fact, although probably not economically feasible, you could probably reverse engineer the process by scanning in enough subjects with known numerical grades that a decent analytical program could self derive the algorithms necessary to develop the grading system. In others words is you scanned in 100,000 Morgans with known grades it could correlate all the factors making up those grades and then derive a fairly accurate grade for future ungraded ones.

""You're comparing apples and duck eggs here your example is so far from a comparable""

Yes and no. What you are essentially saying is numerical grading of coins is a bit like a beauty contest with judges holding up cards. It also goes back to this comment:

"the best AI's in the world can recognize an individual's face in a crowd of thousands, but can't tell you if the face is ugly, beautiful, or in-between"

which is also subjective. Given that there are some recognizable standards of beauty, derived through countless hours of blind testing of body shape silhouettes and the like a program could be written to determine where a given person would, on average, fall. Individual choice may vary, beauty is after all in the eye of the beholder, but average standards could be determined based on such things as shape of the eyes, nose, lips, mouth, cheekbones, fat content, hair style, color, etc., etc., etc. The more variables included the more accuracy would develop.

Again, I asked of this was feasible or being worked on not that I would find it preferable or recommend it.

It is not feasible and to the best of my knowledge no work is currently underway in the field.

The costs in manpower, equipment, and time would far, far outweigh any possible benefits. An intelligent, self-learning AI-driven coin analysis and grading system would likely run tens of millions of dollars in R&D and production costs; meaning that any company wishing to market it as part of a business would need to charge ridiculous fees for slabs.

In systems analysis and engineering it's a matter of cost-benefit. In other words, on one hand I can spend $10 million to develop a machine that does a job 80% as good as an equivalent human worker. It will cost about $10,000 a year in maintenance and repairs. On the other hand, a human worker can perform the same job at 100% with only $10,000 in training and onboarding, but costs me $40,000 a year in salary. How do me and my company know which is the better path, short-term and long-term? There are lots of factors to consider, such as market trends (it's easier to remove humans from my staff if business slows down, not so easy to remove my $10 million coin-grading machine), ROI (how many coins do I have to grade to recoup my investment and turn a profit over yearly expenses, and what price point do I need to set my product at to both remain competitive and make margins?), so on and so forth.

As of 2019 the market, the task, and the product price level all favor human graders. For this to change any time soon, computer grading systems would need massive reductions of scale in cost and massive gains in accuracy and usefulness. If the best human grader gets a technical grade "correct" 90% of the time, for a $35 grading fee, and the best computer grader gets a technical grade "correct" 96% of the time, but it costs $350, can the 6% accuracy improvement (just a hypothetical) justify the price point? I think not, and that's the reason why this is currently not feasible. The computer grading system would need to be both price-competitive AND better/more accurate than human graders, and the technology just isn't there to hit those goals.

Not much of a demand for computer grading, I doubt it will be used anytime soon.




They dropped it.

I agree with that thread that 100% accuracy is expensive and unlikely. It does seem that "good enough," where there is enough data, is achievable. Machine learning is great when there are distinct categories with a fair number of entries in each, but not so good if the sample size is small.

Consider something that can take a standard photo and let you know whether it will grade "high enough to be worth sending in" or "will definitely grade cleaned". This does not replace a human grader, but does let human graders focus on the subjective part of grading.