Questions on the Helmholtz machine.

[JohnSSM]

In another post, asking about free energy, someone mentioned Helmholtz, and I looked into Helmholtz and I discovered the Helmholtz machines, developed from his ideas that the brain is a "statistical inference engine".  

Does Helmholtz imply that a "statistical inference engine" can create models from observed functions and their noted effects without really knowing anything about the nature of the "process" being observed?  Does the Helmholtz machine prove him right?  Is he right?   Thank you. 

Edited March 29, 2021 by JohnSSM
Grammar

[studiot]

3 hours ago, JohnSSM said:

In another post, asking about free energy, someone mentioned Helmholtz, and I looked into Helmholtz and I discovered the Helmholtz machines, developed from his ideas that the brain is a "statistical inference engine".  

Does Helmholtz imply that a "statistical inference engine" can create models from observed functions and their noted effects without really knowing anything about the nature of the "process" being observed?  Does the Helmholtz machine prove him right?  Is he right?   Thank you. 

Good morning.

I, for one, have never heard of the Helmholtz machine so I don't doubt your information.

However a reference would be helpful so that we can learn something more about it.

Thank you.

[JohnSSM]

I just started with wikipedia.

Helmholtz machine - Wikipedia

[studiot]

4 minutes ago, JohnSSM said:

I just started with wikipedia.

Helmholtz machine - Wikipedia

Thanks. +1

Having now looked at the Wiki article, which thankfully was quite short and not embellished with too much jargon that also needs looking up I can say the short answer to your question is yes but there are some serious caveats to observe.

From my applied maths side of the issue I would say that this is about what we call mathematical modelling.

That is using the known characteristics and responses of some system or structure to predict the characterisitcs and responses of another system or structure which behave similarly (hopefully identically) in some respect.

Caveats

No model is perfect except the system or subject being modelled itself.

The 'match' between system and model normally extends over some range or another.
Attempts to model outside this range are very likely to be misleading at best.
This is why interpolation, which means bracketing the output between two ( or more) known values, is considered more reliable than extrpolation, which means extending the known range of validity of the model.

In relation to the o p question, the Wiki article mentions Boltzman statistical mechanics and the Boltzman entropy formula from statistical mechanics.

This formulation rests on the principle of equal probability of all states which does not always hold good.

It is a very simplified assumption, which works in many cases but far from all and the exception have lead to much important modern Physics.

This important fact is often omitted from descriptions of statistical mechanics.

 

Before I offered you a simple, but perfectly sound, explanation of entropy, that you will not find in most treatments either.
This explanation requires about the mathematics available to an 11 year old ie the understanding that area = length time breadth.
My offer still stands.

 

 

 

[JohnSSM]

8 hours ago, studiot said:

From my applied maths side of the issue I would say that this is about what we call mathematical modelling.

Yes, the perspectives on Math play heavily into this discussion, yet my own knowledge cannot help me with mathematical understandings at this level.

May I give you another citation?  If you like, you can skip to part 12 called (A demonstration)  on page 22.
Microsoft Word - HelmholtzTutorialKoeln.doc (nku.edu)

The entire article is full of math which I can only imagine, lends some type of quantifiable proof.  As far as I can tell, the helmholtz machine does work to some high degree of objective reality.  

"The machine has clearly captured much of the world’s structure here: vertical bars appear with higher probability than horizontal bars, all patterns in the world are generated by the machine, and patterns not in the world occur with low probability and are mostly just a bit away from real patterns. Yet, as experience with these machines has shown (e.g. [6]), the machine hasn’t quite captured the world in what we might judge to be the most natural way, as shown by the unreal pattern 000010111 with probability 0.0412."

It seems to be that the machine was 88 percent correct.  And the question in my mind is..."If you had a mathematical solution that gave you correct answers 88 percent of the time, would it be a total failure, a partial success, or is it impossible for math to produce correct results 88 percent of the time?  

9 hours ago, studiot said:

Before I offered you a simple, but perfectly sound, explanation of entropy, that you will not find in most treatments either.
This explanation requires about the mathematics available to an 11 year old ie the understanding that area = length time breadth.
My offer still stands.

MY understanding of entropy really kicked in and I think I've got it nailed down fairly well, but any objective math, that I can understand, is always welcome.  A message perhaps?

[joigus]

13 hours ago, JohnSSM said:

In another post, asking about free energy, someone mentioned Helmholtz, and I looked into Helmholtz and I discovered the Helmholtz machines, developed from his ideas that the brain is a "statistical inference engine".  

Does Helmholtz imply that a "statistical inference engine" can create models from observed functions and their noted effects without really knowing anything about the nature of the "process" being observed?  Does the Helmholtz machine prove him right?  Is he right?   Thank you. 

Thanks a lot for bringing this to my attention. I'd heard about Boltzmann brains, but not about Helmholtz machines.

Edited March 29, 2021 by joigus
typo