Log in or Sign up
Coin Talk
Home
Forums
>
Coin Forums
>
Coin Chat
>
Coin-Recognition Technology?
>
Reply to Thread
Message:
<p>[QUOTE="messydesk, post: 3007137, member: 1765"]Busy thread, and I'm going to make it busier. A couple things I want to touch on here:</p><p><br /></p><p><br /></p><p><br /></p><p>Yes and no. As stated earlier, PCGS has this with their Secure Plus/Gold Shield service. They can easily recognize if a coin has already been through for grading if they choose to scan it when resubmitted, and the grading history becomes an influencing factor in the grading process. It's not a uniform technology or database, since each TPG wants to maintain a competitive edge over the others.</p><p><br /></p><p><br /></p><p><br /></p><p>Yes, but this number would probably be easily removed from a coin without affecting value. Not so for a diamond.</p><p><br /></p><p><br /></p><p><br /></p><p>This is correct. The reason that a computer or AI can't evaluate eye appeal is that we can't arrive at a ground truth evaluation of it. Ask 9 different experts about the eye appeal of a coin and you'll get 10 different answers. A computer will do no better than that, and that's if the training and test data (lots of it) are all good. </p><p><br /></p><p>One interesting things about AI is that if you're able to think about how you'd develop the skill and execute the task, the AI version is probably similar. So for assessing eye appeal, I look at a coin under a consistent light, both sides, and make a judgment based on my internal "eye appeal rules," which are guided by my tastes in coins. I could look at 10000 coins and judge their eye appeal, and perhaps do so consistently (I have to if it is to be useful training data). Now think of what has to be fed to a computer to train it. We need dynamic (not static) pictures of the coins appearance to a pair of eyes (not a single eye) under consistent light that gives the same views I am privy to when I'm assessing eye appeal. Fortunately, for motion and eye appeal, I don't need huge images. 1K x 1K will do, but I need color, so 3 MB per frame. The dynamic requirement now means that for each coin I need to have many frames. If I want to be consistent from coin to coin, then I need the same views, which I don't really know in advance of seeing the coin, so figure 100 views of the coin to simulate moving the coin around. Double that, since I need both sides, and double it again, since I need to use two eyes at once to view the coin. For those keeping score, that's 1.2 GB of data for each coin. While compressed storage will shrink that, the amount of data that we're training with is still 1.2 GB once it's read from storage. If I viewed 10000 coins, we want our AI system to have that many for a combination of training, testing, and validation. 12 Terabytes, and that's just to get my opinion automated, assuming my tastes don't change, since if they do, we have to start the training step over again. And if my tastes aren't the same as someone else's the results will be worthless. It should be apparent by now that attempting to assess eye appeal, or other predominantly subjective tasks for which we can't determine the ground truth, with AI is not the hill to die on.</p><p><br /></p><p>One thing I really think AI could do well is detect acceptable toning. I'm avoiding the word "artificial" because I don't want to revisit the debate of exactly what artificial means. PCGS already has a ton of images of acceptable and questionable coins, and they usually take these pictures so as to show a complete picture of the toning, suppressing the luster. Assuming it isn't necessary to reassess some of them for correctness, an AI system could be trained to tell the difference between the two classifications, then classify images of test coins acquired in the same manner (suppressed luster, complete toning).[/QUOTE]</p><p><br /></p>
[QUOTE="messydesk, post: 3007137, member: 1765"]Busy thread, and I'm going to make it busier. A couple things I want to touch on here: Yes and no. As stated earlier, PCGS has this with their Secure Plus/Gold Shield service. They can easily recognize if a coin has already been through for grading if they choose to scan it when resubmitted, and the grading history becomes an influencing factor in the grading process. It's not a uniform technology or database, since each TPG wants to maintain a competitive edge over the others. Yes, but this number would probably be easily removed from a coin without affecting value. Not so for a diamond. This is correct. The reason that a computer or AI can't evaluate eye appeal is that we can't arrive at a ground truth evaluation of it. Ask 9 different experts about the eye appeal of a coin and you'll get 10 different answers. A computer will do no better than that, and that's if the training and test data (lots of it) are all good. One interesting things about AI is that if you're able to think about how you'd develop the skill and execute the task, the AI version is probably similar. So for assessing eye appeal, I look at a coin under a consistent light, both sides, and make a judgment based on my internal "eye appeal rules," which are guided by my tastes in coins. I could look at 10000 coins and judge their eye appeal, and perhaps do so consistently (I have to if it is to be useful training data). Now think of what has to be fed to a computer to train it. We need dynamic (not static) pictures of the coins appearance to a pair of eyes (not a single eye) under consistent light that gives the same views I am privy to when I'm assessing eye appeal. Fortunately, for motion and eye appeal, I don't need huge images. 1K x 1K will do, but I need color, so 3 MB per frame. The dynamic requirement now means that for each coin I need to have many frames. If I want to be consistent from coin to coin, then I need the same views, which I don't really know in advance of seeing the coin, so figure 100 views of the coin to simulate moving the coin around. Double that, since I need both sides, and double it again, since I need to use two eyes at once to view the coin. For those keeping score, that's 1.2 GB of data for each coin. While compressed storage will shrink that, the amount of data that we're training with is still 1.2 GB once it's read from storage. If I viewed 10000 coins, we want our AI system to have that many for a combination of training, testing, and validation. 12 Terabytes, and that's just to get my opinion automated, assuming my tastes don't change, since if they do, we have to start the training step over again. And if my tastes aren't the same as someone else's the results will be worthless. It should be apparent by now that attempting to assess eye appeal, or other predominantly subjective tasks for which we can't determine the ground truth, with AI is not the hill to die on. One thing I really think AI could do well is detect acceptable toning. I'm avoiding the word "artificial" because I don't want to revisit the debate of exactly what artificial means. PCGS already has a ton of images of acceptable and questionable coins, and they usually take these pictures so as to show a complete picture of the toning, suppressing the luster. Assuming it isn't necessary to reassess some of them for correctness, an AI system could be trained to tell the difference between the two classifications, then classify images of test coins acquired in the same manner (suppressed luster, complete toning).[/QUOTE]
Your name or email address:
Do you already have an account?
No, create an account now.
Yes, my password is:
Forgot your password?
Stay logged in
Coin Talk
Home
Forums
>
Coin Forums
>
Coin Chat
>
Coin-Recognition Technology?
>
Home
Home
Quick Links
Search Forums
Recent Activity
Recent Posts
Forums
Forums
Quick Links
Search Forums
Recent Posts
Competitions
Competitions
Quick Links
Competition Index
Rules, Terms & Conditions
Gallery
Gallery
Quick Links
Search Media
New Media
Showcase
Showcase
Quick Links
Search Items
Most Active Members
New Items
Directory
Directory
Quick Links
Directory Home
New Listings
Members
Members
Quick Links
Notable Members
Current Visitors
Recent Activity
New Profile Posts
Sponsors
Menu
Search
Search titles only
Posted by Member:
Separate names with a comma.
Newer Than:
Search this thread only
Search this forum only
Display results as threads
Useful Searches
Recent Posts
More...