> Here is an interesting article. 
> New Machine Learning Theory Raises Questions 
> About the Very Nature of Science

Naw, don't buy that stuff - we are in my favorite territory here, as I worked the day shift at the idea factory for decades.
Models are just tools to evaluate comically oversimplified versions of reality, and our goal is for ever-sharper tools, so the tools have never been in charge.
Neural networks replace "programmed rules and procedure" with "learned stimulus-response" from a pile of data, but Prolog still remains far more popular an "AI Platform" than any of the neural nets, as it has since the early 1970s.

The article says:

“Hong taught the program the underlying principle used by nature to determine the dynamics of any physical system,”

Ah, but does Hong really know the underlying principle(s) himself?
He apparently THINKS he does, but we came up with those underlying principals how, exactly?
We observed "a" and "b", and did some math, and that math worked well for both, and also for another thing ("c") we had not seen before the equation was made about "a" and "b".
So we told ourselves a harmless fairy tale about how the equation had "predictive power", and we had uncovered a "secret of the universe".

Ha!  Not so fast!

Just for fun, Google "unresolved issues in dynamics  -email"  (the "-email" is to eliminate results about the massive number of problems with "Microsoft Dynamics 365 Customer Relationship Management [CRM]" software).

There are 5.4 million results. Even if only some of them are "on target" for this issue, you don't have to take Course 8 to see that there clearly are many, many unresolved issues in many aspects of "dynamics".

Yes, neural networks have been around for quite some time, and can be impressive in narrow practical applications.  But such "expert systems" occasionally do unexpected things as a result of their "training" that some call "emergent behavior" until the "result" is that your Tesla auto-drives itself straight into the back of a stopped fire truck at 65 mph on the 405 freeway, and this is the THIRD TIME it happened:

https://wired.com/story/tesla-autopilot-why-crash-radar/

I have a little neural network running on a raspberry pi board that looks at chlorine level and pH in a pool, and turns on the saltwater chlorine generator as needed, and dispenses Borax to keep the pH from getting too high.  It was "trained" with an understanding of pool chemistry as represented by tables of pH and chlorine level, but any error would be simple to correct, as normal (non-neural network) code will report any wide divergence from normal conditions, as pumps can fail, and so on.  But any dressing up of a neural network with tinsel and glitter and endowing it with prescience is a fool's error, as it never shows its work, and you never get any equations out of the deal, so you just can't be certain of what it was "thinking" at any one time.  So, it keeps a pool "in spec" when I am half a world away, but it is not an attempt to be perfect, and merely reduces the "pool guy" need to a once-a-month visit.


So, TL;DR - One black swan, and you are slamming at 65mph into the back of that big red fire truck with the flashing lights and ear-splitting siren that any idiot could see and avoid.

             ***********************************************
The BEE-L mailing list is powered by L-Soft's renowned
LISTSERV(R) list management software.  For more information, go to:
http://www.lsoft.com/LISTSERV-powered.html