Can "AI" discover new Physics

[Ouroboros]

I’m often amazed both how stupid and how intelligent our Jack Russell can be at times. Mind you, his maths is dire. 

[saac]

Maths is overrated, he'll turn out fine don't worry  

Jim

[andrew s]

13 minutes ago, Ouroboros said:

I’m often amazed both how stupid and how intelligent our Jack Russell can be at times. Mind you, his maths is dire. 

Lot of people are like that too.

Regards Andrew 

[DaveL59]

15 minutes ago, Ouroboros said:

I’m often amazed both how stupid and how intelligent our Jack Russell can be at times. Mind you, his maths is dire. 

I guess then if the dog is the relative equivalent of these directed/focused AI is a cat then the equivalent of an unsupervised one? After all they rarely do what you tell them or expect, well other than sleep and eat that is 😉 

[saac]

1 hour ago, DaveL59 said:

I guess then if the dog is the relative equivalent of these directed/focused AI is a cat then the equivalent of an unsupervised one? After all they rarely do what you tell them or expect, well other than sleep and eat that is 😉 

The cat has reached a state of enlightenment - or so it thinks anyway!

Jim 

[Ouroboros]

3 hours ago, iantaylor2uk said:

I don't think a dog would come up with new mathematical ideas: https://www.ox.ac.uk/news/2021-12-01-machine-learning-helps-mathematicians-make-new-connections

No, indeed. I wonder how far away we are from AI coming up with a mathematical proof that mathematicians can’t understand? 

[Alien 13]

2 hours ago, Ouroboros said:

I’m often amazed both how stupid and how intelligent our Jack Russell can be at times. Mind you, his maths is dire. 

Throw a ball at any speed or angle or distance and most dogs will catch it proving that their "Maths" is very good indeed..

Alan

[Macavity]

The remanent interest (for me anyway) is the mystery re. the nature of the state variables!
It seems the number of variables can be quantified with confidence (by rounding down!).
I casually wondered if, repeating the "experiment", would give *different* variables? 😉

Edited August 1, 2022 by Macavity

[Macavity]

35 minutes ago, Alien 13 said:

Throw a ball at any speed or angle or distance and most dogs will catch it proving that their "Maths" is very good indeed..

Certainly food for thought - Perhaps an internal brain-model of a parabola? Shapes (of skylines) seem
to remain in my memory - Recently cyber-revisiting (awful?) 60/70's "College Campus" architecture! lol

https://magazine.scienceconnected.org/2015/05/biologist-explains-how-we-learn-to-catch/
The idea of "biological timing clocks"? Nice to see a ref. to Cricket! T20 spectating is fairly hazardous? 😅

Edited August 1, 2022 by Macavity

[Ratlet]

7 hours ago, Macavity said:

The remanent interest (for me anyway) is the mystery re. the nature of the state variables!
It seems the number of variables can be quantified with confidence (by rounding down!).
I casually wondered if, repeating the "experiment", would give *different* variables? 😉

I can see a decent sci-fi short story where they set one of these machine algorithm things up, feed it the results from dozens of experiments from particle physics, cosmology, etc and let it run and watch it chew through the data and update them on the accuracy of the model and the number of variables required.  Over time it gets ridiculed and forgotten about.  In the end it gets shutdown, with the technician monitoring it taking a last look at the screen as it displays the output:

Variables: 1.0

Model fit: 99.99999999999......%

Edited August 2, 2022 by Ratlet
Added a %

[billhinge]

On 31/07/2022 at 22:57, Macavity said:

Maybe we're not that into "talking computers"! But: Can "AI" discover new Physics? 🤔
For a change not the Youtube link! (But there is a FUN video embedded!)
https://www.creativemachineslab.com/hidden-variables.html
Interesting stuff, if you're into Robotics? 🥳

Sounds pretty cool to me, there were similar projects for financial markets. If you believe in technical analysis, which TA indicators are were more significant impacts on price . eg https://deepai.org/publication/bayesian-regression-and-bitcoin

If you are familiar with AI and github its fairly easy to try yourself

When I was a first year undergraduate one of the first questions discussed in tutorials was working out factors that influence a system and how they are proportional to the result. Sounds like an AI version of that

Lots of food for thought

 

 

[DaveL59]

7 minutes ago, billhinge said:

Sounds pretty cool to me, there were similar projects for financial markets. If you believe in technical analysis, which TA indicators are were more significant impacts on price . eg https://deepai.org/publication/bayesian-regression-and-bitcoin

If you are familiar with AI and github its fairly easy to try yourself

When I was a first year undergraduate one of the first questions discussed in tutorials was working out factors that influence a system and how they are proportional to the result. Sounds like an AI version of that

Lots of food for thought

 

 

hehe yeah technical indicators, support and resistance levels and so on, learned some of that when I dabbled with forex trading and writing my own bots. Thing is they rely on a behaviour within a sort of norm that doesn't really exist and is influenced by external events that can blow those away. Factoring in those events and their effect tho isn't trivial and typically needs some human interpolation based on experience and hunches. Easier to write a system to try ride a trend than to make quick coin on short fast moves but needs deep pockets to ride the waves along the way.

[billhinge]

27 minutes ago, DaveL59 said:

hehe yeah technical indicators, support and resistance levels and so on, learned some of that when I dabbled with forex trading and writing my own bots. Thing is they rely on a behaviour within a sort of norm that doesn't really exist and is influenced by external events that can blow those away. Factoring in those events and their effect tho isn't trivial and typically needs some human interpolation based on experience and hunches. Easier to write a system to try ride a trend than to make quick coin on short fast moves but needs deep pockets to ride the waves along the way.

Maybe, but the point I was making was that it is possible to use computational tools to see what are the key factors that an AI learnt and by what percentage each factor matters

It's not a leap of faith to me to use the above analogy with physical models and parameters. Problem I see is that if we try to understand gravity, would it learn Newton or Einstein or GR hybrids if we teach it with a falling apple vs satellite vs black hole? The AI's conclusions would be influenced by its observations and accuracy (and experimenter's bias) surely - as would ours? 

Edited August 2, 2022 by billhinge

[Ouroboros]

18 hours ago, Alien 13 said:

Throw a ball at any speed or angle or distance and most dogs will catch it proving that their "Maths" is very good indeed..

Alan

Or maybe it’s just a good ol’ feedback control mechanism in action? There’s probably a bit of learning to develop eye/mouth coordination and muscle memory …. on top of a lot of instinctive behaviour.