• When you click on links to various merchants on this site and make a purchase, this can result in this site earning a commission. Affiliate programs and affiliations include, but are not limited to, the eBay Partner Network.

A discussion about using AI to grade comic books.
3 3

Started by VintageComics,

152 posts in this topic

CAHokie

CAHokie

Posted
Posted

IMG_1215.thumb.jpeg.1c6265e2349cb7a1c2ab99f7258f27a5.jpeg

Oh, Roy doesn’t think I can grade? Where is that Edgar Church - Superman # 1…. I will show him! 
 

 

* mocking you using AI brings me a weird sense of joy. :nyah:   This will be my last one though. Most likely.

Link to comment
Share on other sites
VintageComics

VintageComics

Posted
  • Author
Posted (edited)
On 12/26/2023 at 3:37 PM, VintageComics said:

What I can actually see, eliminating ALL of this complexity that we're discussing and everyone is having trouble envisioning, is having pre-graders count and grade the interiors, inspect staples and interior covers for defects, maybe do one interior AI shot for estimating page quality, and then CGC can use AI just to grade the outer covers.

 

Then a finalizer would look at the AI estimate, the pregrader's notes and come up with a finalized grade. 

THAT is how I can see AI being used very soon. 

And this idea is an evolution of this very discussion. I literally just came to that thought as I was typing up this post. 

It's quick, cost effective, easy to implement, efficient and makes grading FAR more accurate. 

I believe THAT is the route CGC will likely go soonest. 

To get back on track, does anyone else agree that this is the most likely scenario?

Edited by VintageComics
Link to comment
Share on other sites
Prince Namor

Prince Namor

Posted
Posted
On 12/27/2023 at 2:07 AM, VintageComics said:

I didn't state any absolutes within the context of this discussion. 

 

On 12/23/2023 at 6:39 AM, VintageComics said:

You might eventually be able to make it happen in the distant future with NASA type tech, but by then all human labor will have been replaced. 

You can do one thing well, but tying all of those things together is space age stuff. 

 

On 12/23/2023 at 6:54 AM, VintageComics said:

there is no way AI is grading books without human involvement any time soon. That's spaceage thinking that is just fairy tale stuff for the foreseeable future. By the time this happens, we'll only be half human. 

And THAT is the point I'm making.

 

On 12/23/2023 at 5:39 AM, VintageComics said:

There's just too much, multi-dimensional, human involvement in grading a book from what I can envision. 

Link to comment
Share on other sites
Paul © ® ⚽️💙™

Paul © ® ⚽️💙™

Posted
Posted

What did the comic nerd call his AI based gf?     Hope.

An AI Oreo is in development.   Apparently it's one smart cookie.

What is the opposite of Articial Intelligence?  Real Stupidity.

 

 

Link to comment
Share on other sites
namisgr

namisgr

Posted
Posted (edited)

You can train an amateur in comic book grading to become a professional comic book grader at CGC.  The training is iterative, and almost certainly takes months rather than years  Whoever thinks you cannot train a machine learning algorithm to become a comic book grader with accuracy comparable to many of the professional graders themselves is either unaware of what machine learning algorithms have been accomplishing over the past decade, unknowing about how iterative training refines and improves the accuracy of the algorithms over time, or both.

Start with a grading algorithm based on the systematic methodology CGC graders use to derive final structural grades.  Have it grade a learning set of books, for purposes of discussion let's say a hundred that span the full range and severity of defect types, and all of which have also been graded and thorough grader's notes recorded by five CGC graders each working independently.  Have the finalizer assign a final grade to each book.  The algorithm will then incorporate information culled from the ranges, disparities, and detailed grader's notes to derive a modified protocol that better fits the human-derived grades for the learning set.  Then the entire process will be repeated.  After each repeat, the algorithm can be expected to come closer and closer to the human-derived grades, and especially the final grade.  Finally, the algorithm and the team of professional graders will evaluate a test set of a hundred different books, so that its accuracy in a real world-relevant scenario may be assessed, and compared with each of the graders who were not the finalizer.

The image capturing component of automated grading of this type is technology that is not 'Artificial Intelligence', and is pre-existing.

This general approach has already been used to devise algorithms that can translate ensembles of nerve cell impulses from the region of the neocortex that directs leg movements in paralyzed persons with damaged spinal cords into electrical outputs that stimulate leg movement-controlling nerve cells in the spinal cord below the site of injury, allowing a person previously in a wheelchair and paralyzed for over a decade to use their thoughts and a brain-spinal cord interface machine with stimulating electrodes and an algorithm that decodes the nerve cell activity to allow them to initiate and engage in walking.  

Given the systematic nature of the professional comic book grading process, and the relative simplicity of the machine learning required, it would be much, much easier to devise an accurate grading algorithm than it would to restore limb movements, or vision, or hearing, or speech in people who've lost these functions to injury or disease, all of which have been accomplished over the past five years.

Making such an automated grading system cost effective is a separate matter, and may certainly take awhile.  But I am of the belief that the technological aspects could be readily achieved.

Edited by namisgr
Link to comment
Share on other sites
MyNameIsLegion

MyNameIsLegion

Posted
Posted
On 12/28/2023 at 7:18 AM, namisgr said:

You can train an amateur in comic book grading to become a professional comic book grader at CGC.  The training is iterative, and almost certainly takes months rather than years  Whoever thinks you cannot train a machine learning algorithm to become a comic book grader with accuracy comparable to many of the professional graders themselves is either completely unaware of what machine learning algorithms have been accomplishing over the past decade, unknowing about how iterative training refines and improves the accuracy of the algorithms over time, or both.

Start with a grading algorithm based on the systematic methodology CGC graders use to derive final structural grades.  Have it grade a learning set of books, for purposes of discussion let's say a hundred, all of which have also been graded and thorough grader's notes recorded by five CGC graders each working independently.  Have the finalizer assign a final grade to each book.  The algorithm will then incorporate information culled from the ranges, disparities, and detailed grader's notes to derive a modified protocol that better fits the human-derived grades for the learning set.  Then the entire process will be repeated.  After each repeat, the algorithm can be expected to come closer and closer to the human-derived grades, and especially the final grade.  Finally, the algorithm and the team of professional graders will evaluate a test set of a hundred different books, so that its accuracy in a real world-relevant scenario may be assessed.

The image capturing component of automated grading of this type is technology that is not Artificial Intelligence, and is pre-existing.

This general approach has already been used to devise algorithms that can translate nerve cell impulses from the region of the neocortex that directs leg movements in paralyzed persons with damaged spinal chords into electrical outputs that stimulate leg movement-controlling nerve cells in the spinal chord below the site of injury, allowing a person previously in a wheelchair and paralyzed for over a decade to use their thoughts and a brain-spinal chord interface machine and an algorithm that decodes the nerve cell activity to allow them to initiate and engage in walking.  

Given the systematic nature of the professional comic book grading process, and the relative simplicity of the machine learning required, it would be much, much easier to devise an accurate grading algorithm than it would to restore limb movements, or vision, or hearing, or speech in people who've lost these functions to injury or disease, all of which have been accomplished over the past five years.

Making such an automated grading system cost effective is a separate matter, and may certainly take awhile.

Even more importantly, CGC has the scans, grades and notes on hundreds of thousands of books. AI wouldn't need to be trained in real time so much as loaded with reference examples of books, especially SA to Moderns for all grades, with a multitude of examples of the cumulative defects that were (in theory) used in arriving at the final grade. I know Roy likes to wax poetic about the human element, the je ne sais quoi if it, the touch the feel of paper, the fabric of our lives, the smell, oooh that smell, can't you smell that smell, blah blah. Well that's the point, to remove the subjective bias of individual graders that governs the difference between a 9.4 and a 9.6m a 0.6 and a 9.8. Each grade of a specific book can be instantly weighted against the average of all previous examples of books in that grade, with similar defects, instantaneously. Humans can't do that.  Imagine bouncing every 8.0 Hulk 181 ever graded against the current copy under consideration to determine if the grade is in range.

also pointed out in this thread, once the mechanical and procedural aspects were properly set aside (goal posts!) AI could be the pre-screen agent, and the QC that is sorely lacking now as CGC incorporates AI into the grading process iteratively. You don't need 45 full time graders, you probably need less than 10 once you eliminate the need for pre-grades, multiple graders, QA etc. the current process is founded on the limitations of the human element, AI eliminates off of that. I realize that's scary for some people, people will lose their jobs or be freed up to do other things, which is sometimes euphemistic for "you're job is obsolete" THAT'S the underlying anxiety of the naysayers. Suddenly their feelings of superiority and expertise, that they are better graders than 99% of other dealers and collectors is moot. The eventual progression of this is that you could prescreen a comic with your phone from an app from someone like CGC to get a pre-screen estimated grade of your book to see if it's a good candidate for slabbing. You make it a subscription rather than a service. There are so many possibilities and business models of how this could be monetized it's crazy, but not so scary as some would want you to think.  those with their head in the sand, or elsewhere wind up in a self full-filling prophecy, because they are the first ones to be displaced, left behind, and obsolete.

Link to comment
Share on other sites
namisgr

namisgr

Posted
Posted (edited)

I just don't think the grader's notes that are recorded and archived are of sufficient detail and reliability for training a machine learning algorithm to grade.  Such training would depend on the assessment by each grader of every defect and a structured outline of how they combined for the derivation of their assigned overall grade for each book.  I would imagine this is the process by which new graders are trained, not off the often sketchy and incomplete notes routinely recorded, but from one on one discussion of the collection of defects and systematic nature of the grading method between grading trainer and grading trainee.

Edited by namisgr
Link to comment
Share on other sites
MyNameIsLegion

MyNameIsLegion

Posted
Posted
On 12/28/2023 at 7:53 AM, namisgr said:

I just don't think the grader's notes that are recorded and archived are of sufficient detail and reliability for training a machine learning algorithm to grade.  Such training would depend on the assessment by each grader of every defect and a structured outline of how they combined for the derivation of their assigned overall grade for each book.  I would imagine this is the process by which new graders are trained, not off the often sketchy and incomplete notes routinely recorded, but from one on one discussion of the collection of defects and systematic nature of the grading method between grading trainer and grading trainee.

Respectfully, I think we underestimate machine learning. Once you establish the lingo, (CF, LRC, spike tick, spin stress, etc etc) It's going ot learn pretty fast. Having such a large data set is a huge advantage. Eventually it would be telling you how many times the notes were incongruent with the images. 

Link to comment
Share on other sites
CAHokie

CAHokie

Posted
Posted

They may not have adequate information saved to help AI grade yet, but with the amount of grading they do, they could start capturing it today. Within 6 months to a year, they would have a massive amount of data for it to learn from. 

Link to comment
Share on other sites
VintageComics

VintageComics

Posted
  • Author
Posted
On 12/28/2023 at 8:18 AM, namisgr said:

You can train an amateur in comic book grading to become a professional comic book grader at CGC.  The training is iterative, and almost certainly takes months rather than years  Whoever thinks you cannot train a machine learning algorithm to become a comic book grader with accuracy comparable to many of the professional graders themselves is either unaware of what machine learning algorithms have been accomplishing over the past decade, unknowing about how iterative training refines and improves the accuracy of the algorithms over time, or both.

As JC25427N repeatedly stated, training the software is not as easy as most people think. and I can understand why.

The "speck of dust" analogy is a perfect one, because a human has the experience to identify and move past it much quicker than a machine program for the same reason that a human and machine don't think the same way. As someone with extensive experience in neuroscience, you should know this. 

So while consistency is an easier goal, adaptability is probably the tougher one and this conversation needs to address both. 

It's not impossible, it's just a road filled with many obstacles, and how soon those obstacles are overcome is what this entire discussion has morphed into. 

On 12/28/2023 at 9:01 AM, MyNameIsLegion said:

Respectfully, I think we underestimate machine learning. Once you establish the lingo, (CF, LRC, spike tick, spin stress, etc etc) It's going ot learn pretty fast. Having such a large data set is a huge advantage. Eventually it would be telling you how many times the notes were incongruent with the images. 

This is where @JC25427N would be really beneficial in the conversation. 

On 12/28/2023 at 8:18 AM, namisgr said:

Making such an automated grading system cost effective is a separate matter, and may certainly take awhile.  But I am of the belief that the technological aspects could be readily achieved.

Thank you for FINALLY agreeing with on on something I've been stating from the beginning. lol

No, seriously. Thank you. (worship)

Link to comment
Share on other sites
namisgr

namisgr

Posted
Posted (edited)
On 12/28/2023 at 10:47 AM, VintageComics said:

As JC25427N repeatedly stated, training the software is not as easy as most people think. and I can understand why.

A key word search of PubMed for the term 'machine learning' finds 136,885 peer reviewed publications in professional scientific journals.   There are over 8,000 published studies archived for the subfield of 'nervous system' alone.  So despite not being easy, it has become routine for those versed and skilled in the requisite programming and computer science aspects.

https://pubmed.ncbi.nlm.nih.gov/?term=machine+learning&sort=date&size=20

To boot, the machine learning for grading of comic books is a straightforward application, and made even easier by the codified and systematic process by which numerical grades are derived by professional graders, and the way the codified and systematic processes are used to rapidly train new graders.

Edited by namisgr
Link to comment
Share on other sites
Guest
This topic is now closed to further replies.
3 3
Go to topic listing