Learning the non-physics part of Statistical Mechanics

[ANewPope23]

Hello, this is my first question on PhysicsForum. I am primarily interested in statistics/machine learning. I have recently discovered that many of the ideas used in machine learning came from statistical physics/ statistical mechanics.

I am just wondering if it's a bad idea to attempt to learn the computational/mathematical aspects of statistical mechanics with zero physics background? Maybe it's easier to learn some physics before attempting this?

 

[FactChecker]

I associate machine learning with neural networks, pattern recognition, etc. Can you be more specific about how machine learning is related to statistical physics/ statistical mechanics?

 

[ANewPope23]

Here are some examples:

"According to Watkin et al. (1993), statistical physics tools are not only well suited to analyze existing learning algorithms but also they may suggest new approaches. In the paradigm of learning from examples (the paradigm considered in this book), examples are drawn from some unknown but fixed probability distribution and, once chosen, constitute a static quenched disorder (Watkin et al., 1993)."

"Statistical Physics and Representations in Real and Artificial Neural Networks" https://arxiv.org/abs/1709.02470

https://arxiv.org/pdf/1706.09779.pdf Ising Models and Spin Glass Models are used in machine learning.

 

[FactChecker]

I can't tell if these are applications of statistical mechanics in the field of neural networks or the reverse -- applications of deep neural networks in the field of statistical mechanics. It looks more like the latter to me, which I could imagine and Google shows several articles along those lines.

If it is true that they are applications of deep neural networks in the field of statistical mechanics, then you may be making a mistake in trying study statistical mechanics to help you understand machine learning.

 

[SchroedingersLion]

No one can learn statistical mechanics without a proper physics background. Besides, the amount of machine learning in a typical statistical mechanics introductory course is close to zero.
If you are interested in machine learning, you should learn machine learning.

 

[boneh3ad]

Isn't the non-physics part of statistical mechanics just... statistics?

 

[gleem]

This might be a place to start. A physicists introduction to the mathematics of statistical mechanics, only about 50 minutes. Since it applies to physics you will have to change the physics examples to neural networks.i.e., see the analogies between the two.

http://theoreticalminimum.com/courses/statistical-mechanics/2009/spring/lecture-2

 

[gleem]

Instead of going to a statistical mechanics book for physicists there seems to be a plethora of material on statistical mechanics of neural networks which would seem more appropriate. eg. https://nms.kcl.ac.uk/reimer.kuehn/CS03/Sompolinsky_PhysicsToday.pdf

 

[Stephen Tashi]

Isn't the non-physics part of statistical mechanics just... statistics?

If you find a text where it is, I'd like to know about it. Random variables "sample spaces" (or "probability spaces"), estimators, probability models - all familiar things when problems are presented in the context of statistics don't appear in the expositions of statistical physics that I've seen. There is traditional terminology in statistical physics that predates (and overcomes) the terminology of modern probability theory and statistics.