Computer grading for ultra-moderns ... this could actually happen.
291fifth
Posts: 24,710 ✭✭✭✭✭
Since most ultra-moderns are submitted soon after they are issued and are untoned, why not? Think of the eye-strain that could be avoided.
All glory is fleeting.
1
Comments
Ya got to find a way to say, "this one is better than that one" so yes, if they're otherwise identical, someone will computer image and map and score MS70 and PF70 coins so that there can be some "qualifiers" that get them some extra money.
Real old coins have absolute rarity, and die varieties, and surface quality (i.e. "not messed with"), intermediate coins have degree of strike fullness and luster fullness and colorful toning (and degree of "believability" of the colors) so why not?
Absolutely, I expect to live to see someones capitalize on "Actual firsty strikes" and "the better of the dies" for otherwise perfect, common, and identical to the naked eye coins, complete with enlarged images with tiny arrows pointing to the money features.
Liberty: Parent of Science & Industry
For an objective grade maybe a ball park guestimate could be ascertained. But nothing beats the human brain for a subjective grade taking in all sorts of aesthetic variables today's AI computers can't equal.
Strides are made continuously in AI for game play (chess, checkers, go etc.) and self driving cars but are based on processor power, sensor inputs and possibility analysis of all, in game play, moves. None of these are applicable to coin grading. Even comparing a prospective coin to a data base of photos (like facial recognition) of previously graded coins of the same date, denomination and mint would fall short.
Besides, when the Zombie Apocalypse brings down modern society and computer networks on which we have become dependent upon, humans will not know how to grade that Morgan Dollar you want slabbed. Whatca gonna do? Oh the humanity !
Computer grading is in our future.... especially with the advances in AI..... right now, it is not a profitable endeavor...just not enough return to invest the considerable time and effort required to develop the programming and data base. Cheers, RickO
Determining whether a coin is ms63-ms65 or pr64-pr66 based on originality, hairlines, marks, eye appeal etc... is completely subjective and therefor grades can fluctuate.
With moderns, I can easily see a computer taking this over today.
If no marks at 5x (or some other magnification) then ms/pr70
if 1-2 marks 69
3-5 68 etc
Except the computer also needs to judge the severity and location of the marks. It will need to be programed for all the prim areas too. Sometimes even one can knock a coin down to 67 or 68.
You're absolutely correct.
Not exactly -- by showing the software lots of coins in lots of human assigned grades it will figure out where the prime areas are for each sort of coin. If done right, I doubt the software will even need to 'know' what series it is looking at.
it will figure out where the prime areas are for each sort of coin. If done right, I doubt the software will even need to 'know' what series it is looking at.
The effective UIS coin evaluation algorithm is coming. Sooner or later, your opinion won't matter. Inevitable.
If PCGS and NGC are not looking at algorithms, I'd be surprised.
I could see computer grading system capping a grade (the coin shouldn't grade higher than X), but not determining a grade.
Coin grading has always been like a beauty contest, with the contestants parading by the eyes of the judges, who, that day at that time with their criteria in mind, will select the winners. Technical grading has aspirationally provided the baseline of objectivity, which I presume can be definitively done with computers some day. Still and all, there was always the gamble and hope that some human expert out there at the services would really like your coin because of aesthetic "eye appeal."
They'll always be room, then, for the visceral admiring reaction of a respected human judge or group of judges. A superimposed designation from such a source that "This here coin's right purty!!" will always add market value over and above the computer's technical grade.
Here's a warning parable for coin collectors...
I totally agree RIcko they are making some great headway into getting computers to train themselves using pictures.
PCGS already has all the data needed or links to it, plus a huge database of counterfeit dies.
So many coins just are the grade they are and always will be unless damaged, AI could do great with those, even now. But every so often there is a special one, perhaps the AI could flag that for human review because of toning, extreme luster or strike or whatever.
Haha just kidding to make you humans feel better, a properly configured AI could grade a coin in an unreal ways, AI could break it down into all kinds of categories. Plot and measure the scratches, % of original strike left, luster percentile, mark attributions and on and on and on and on.
It could search auction results matching a given coin to other results for that same type in the same grade, In some auctions there can be $100's or $1,000 in difference even in the same grade. AI could probably figure out if your submitted coin was a low or high dollar coin for the grade, or had some make collector happy anomaly that made it shoot the moon at auction.
Opps sorry I'll stop.
Wmwood
Ai needs input to learn with a good algorithm for analyzing. Does PCGS have photos of every coin they have graded with diagrammed key points?
Yes, auction catalogs would add to the database. But they still need classification and diagramming. Lots and lots of man hours.
Is it coming? Eventually. But I think it's a lot more involved than can be imagined.
Artificial intelligence as you've described it is actually a subset of the AI field: reinforcement learning. It allows for complex decisions to be made with regards to some notion of reward and cost for every action taken.
As you said, this is a very poor fit for the task of grading coins.
Luckily, machine learning is perfect for this. There are many case studies in which computers perform comparably to humans in classification tasks, the simplest of which is identifying an object.
The grading lines between 68 and 70 are pretty simple and not really an art - most all moderns have equivalent characteristics with regards to luster and strike, so what it really comes down to is the number of ticks.
I haven't tried it before, but it does not strike me as a very difficult problem for neural networks to solve given today's computing power. Just treat the coin's grade as it's corresponding "class."
To address your concern of "diagrammed key points," neural networks are particularly useful precisely because they are incredibly gifted in picking out their own features, unlike old school computer vision where researchers had to go in and hand code features for a network to detect.
While PCGS won't need images of every modern they've ever graded, what is true is that they will need a fairly large database of them - probably on the order of thousands or tens of thousands for each design and for each side of the coin.
Since grading is subjective...
Since grading standards have "evolved"...LOL.
Since the TPGS have over graded and under graded a very large number of coins among the millions graded (most under graded have been corrected)...
Since coins in past auctions are just images...
Since eye-appeal varies for each of us...
Since lighting and orientation change what can be seen on a coin...
Good luck programing the computer. I sure would not wish to be one of the "Experts" chosen to educate the machine either. However, I'd like to have all the $$$$ that went down the rat hole with Comp-U-Grade and the "Expert" while chasing that dream.
Modern coin grading is definitely governed by one of the more complex energy functions in the universe.
Of course, the eventual (but unavoidable) advent of computer grading will also be the end of gradeflation. That has both positive and negative connotations....Cheers, RickO
A deep convolutional neural network could be trained to spit out a grade given a training dataset comprising images sufficient for grading and the ground truth grade corresponding to the images. This is called "supervised machine learning." The accuracy of such a system is limited by the ground truth given to the system. One of the big issues with grading today is that there is no ground truth, as grades are demonstrated to be potentially variable over time. If there were a ground truth, the concept of the crack-out game would be foreign to us. If we accept that grading varies over time, then the network would need periodic retraining, which means we don't care about repeatability. Another problem is determining what constitutes sufficient imaging for grading, and then executing it effectively and repeatably.
The prospect of having this is interesting and makes for good discussion, but as you can see, the biggest obstacles are not technological.
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars • Variety Attribution
You would need to train the computer by allowing it to learn by reviewing 1,000's of coins and manually entering into the computer the coin'a grade. Which means theoretically you would need 1,000's of accurately graded coins...which in itself is a challenge. Since computers would always be objective, presuming you can get past the issue above, many would cry/whine when they receive an MS66 on their (today's approach) MS69.
The challenge is for moderns the sclae is really 65-70, not 1-70. That is something that most would need to accept in order for computer graded coins to be acceptable by the market.
Think of all the re-submissions that will generate. And with computer technicality I'm thinking most would be headed for a downgrade.
Great transactions with oih82w8, JasonGaming, Moose1913.
When a TPG can do that with computer grading and get the submission cost to between $5-$10 that will be a game changer.
Modern coin collecting has been gaining momentum due to affordability especially in relation to melt. Furthermore many people are collecting favorite dates like anniversary or birthday of mod issues from different countries. Most people walking into a coin show have between only $100 - $300 to spend. At this level risk is reduced also vs $4500 in just one coin.
Which means something else has to be created to force resubmissions. Getting the grade right the first time, every time, is not nearly as profitable....maybe not even sustainable. And if grading can be put on computer...why not an AP available by iphone at any coin show? Who needs coin dealers then other than as conduits to move the coins around at a fixed 2-5%? If I were the TPG, I'd be moving real slow in this regard....just like the automakers when it comes to electric cars. Once all the previously graded coins get redone, that's about it. But, figure 50-75% of collectors might just get up and walk away if they have to regrade everything.
There is no modern coin collecting gaining momentum. The sheer number of new issues every year is a weight that can't continue to be supported. Stamps went through that stage and collapsed. Sports cards too.
i used to work for a major software company in the 1990's. we had some of the best programmers in the world. not only that, we worked with top programmers at MIT, Stanford, and Harvard to develop software that can spot infinitesimally small differences comparing two similar objects, with amazing accuracy!
without boring you with all the particulars, i can say that based on what i know about what we could do then: i am positive that computer coin grading is not only possible, but could be done now.
this does not mean the computer is autonomous. it just means that computer easily spots differences between reference coins and new examples...modern or not. human interaction is always required for some type of analysis, even if only for anomalies.
the software, if developed correctly, is much more accurate at grading than a human. however, it is human nature to be skeptical of it anyway!
coin grading has always been a point of contention with collectors. we argue now about pcgs, ngc, and cac. we grade these coins anyway, even though they are already graded.
i don't believe that computer grading ever eliminates people's opinion (grade) of a coin, even if it becomes the best system we have.
Of course, it can be done. The question is whether it can be done effectively.
Accuracy can only be measured against ground truth. We have no ground truth in grading, so accuracy cannot be effectively measured. Repeatability, on the other hand, can. And when the grading models change, the repeatability measurements could be used to estimate how many coins have become misgraded by the current standards.
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars • Variety Attribution
I think we are on the cusp of an era where purists like many of us will place a premium of Human Graded Coin.
I like it. A new acronym to use in eBay searches: "HGC"
...its all over once computers actually start collecting coins boys
Yes AI can and WILL be able to grade coins BUT will the AI be able to judge eye appeal? Toning, luster(maybe with refraction). AI will become so advanced (already is) with chemical analysis to determine if your "blast white" coin was dipped or cleaned, or that your wildly toned coin was gassed.....what then? Maybe the natural toned "unadulterated" and non dipped coins will be the wave and highly collectible coins of the future? ....just a thought. I will never let a computer determine what kind of art I enjoy,or what kind of woman I am attracted to,or what I want on my hamburger and definetly not what kind of coin I purchase. Buy what you like... not the holder or grade and you will always be happy. Unless your trying to make a living at this...let the dealers and market makers deal with rhe rest of the BS.
Cheers!
Coinbits instead of Bitcoins?
Lemme call the press !!!