How would a supercomputer handle ASL?

Tuomo

Keeper of the Funk
Joined
Feb 10, 2003
Messages
4,702
Reaction score
5,739
Location
Rock Bottom
Country
llUnited States
One of the problems with modern Machine Learning techniques is that while they are very good at solving particular problems, it is difficult to peak under the hood and figure out what they are doing
This is in fact something I struggle with at work. I have been trained to be a good engineer by understanding my algorithms and knowing what my code is doing. If I don't understand the inner workings of an AI, then it's hard for me to feel good about my work.

AND YET. My customer doesn't give a damn whether I feel good or bad about my results. If I use an AI to generate results they like, they're happy and I get paid.

So it's complicated. I tend to sneer at the Young Punks who don't understand what their AI's are really doing. And yet their results may be better than mine. Which would YOU pay for?
 

boylermaker

Senior Member
Joined
Jan 22, 2012
Messages
582
Reaction score
531
Location
Virginia
Country
llUnited States
So it's complicated. I tend to sneer at the Young Punks who don't understand what their AI's are really doing. And yet their results may be better than mine. Which would YOU pay for?
My philosophy is that machine learning is great when all you want is prediction, and understanding isn't necessary. If understanding is important, then you probably want to go the human-learning direction of linear modelling, etc. Of course, it's possible that the thing you are trying to understand is so dominated by local interactions that trying to understand it with a human brain is a fools errand. In which case no human-learning model is going to help you understand it, so the incomprehensibility of ML stops being a downside.

Of course, if you want to understand ASL, be best thing to do would be to combine our ASL-playing AI with a GPT-3 trained on classroom lectures, then give it prompts like "The best way to use DCs is ..."
 

AdrianE

Senior Member
Joined
Feb 14, 2003
Messages
913
Reaction score
269
Location
Ottawa, Ontario
Country
llCanada

von Marwitz

Forum Guru
Joined
Nov 25, 2010
Messages
14,504
Reaction score
10,526
Location
Kraut Corner
Country
llUkraine
To program an AI for computerized ASL seems a titanic task even today, let alone in 2003...

Just for comparison:

I think it was back in 2013 when Matrix Games released a computer verison for 'World In Flames' (WiF) at 'Spiel' games fair in Essen Germany which was supposed to have an AI. While WiF is a very complex wargame on the strategic level with very complex phases, its rulebook is only a fraction of the ASLRB. I think they have one (single) programmer who is working full-time on the thing for years before the release and ever since. They are still fixing bugs the last time I took a look and have not even started with an AI.

von Marwitz
 

Philippe D.

Elder Member
Joined
Jul 1, 2016
Messages
2,154
Reaction score
1,407
Location
Bordeaux
Country
llFrance
To start a machine learning program to self-teach ASL, you'd need to write a computer program that was able to decide any conceivable rules question - as in, in any possible situation, correctly and unambiguously decide how the rules apply.

That in itself would be a pretty nifty thing to have. But I'm not holding my breath.

Comparatively, the rules of Chess, or Go, are trivial to code. They're something one could expect of an undergrad student.
 

zgrose

Elder Member
Joined
Jun 13, 2004
Messages
4,294
Reaction score
1,008
Location
Kingwood, TX
First name
Zoltan
Country
llUnited States
To start a machine learning program to self-teach ASL, you'd need to write a computer program that was able to decide any conceivable rules question - as in, in any possible situation, correctly and unambiguously decide how the rules apply.

That in itself would be a pretty nifty thing to have. But I'm not holding my breath.

Comparatively, the rules of Chess, or Go, are trivial to code. They're something one could expect of an undergrad student.
One could drop the problem space dramatically if one targets a particular scenario.
 

Philippe D.

Elder Member
Joined
Jul 1, 2016
Messages
2,154
Reaction score
1,407
Location
Bordeaux
Country
llFrance
One could drop the problem space dramatically if one targets a particular scenario.
Certainly, but even writing a program for the rules of everything that could happen in, say, a basic scenario with Infantry, some SW, one specific Gun and a few vehicles, would still be a pretty difficult task - not at all on par with such a "simple" game as Chess.
 

Actionjick

Forum Guru
Joined
Apr 23, 2020
Messages
7,877
Reaction score
5,278
Location
Kent, Ohio
First name
Darryl
Country
llUnited States
Certainly, but even writing a program for the rules of everything that could happen in, say, a basic scenario with Infantry, some SW, one specific Gun and a few vehicles, would still be a pretty difficult task - not at all on par with such a "simple" game as Chess.
Seems for the moment at least that grey matter is triumphant, at least when it comes to ASL.
 

zgrose

Elder Member
Joined
Jun 13, 2004
Messages
4,294
Reaction score
1,008
Location
Kingwood, TX
First name
Zoltan
Country
llUnited States
Probably more to do with the pocketbook than the brain. :)
 

Old Noob

Forum Guru
Joined
Jan 10, 2020
Messages
2,305
Reaction score
2,445
Country
llUnited States
If it played me, it would be muttering "Illogical, illogical. Norman, coordinate."
 
Top