Is There a Connection Between Physics and Machine Learning?

[Gean Martins]

Hello everyone, I'm on my last undergraduation year in Physics and I've been asking myself what specific area to work with . Two months ago I've been studying Machine Learning and it amazed me so much that i push myself to come here and ask your opinions about it : There's a way to work deeply with Machine Learning and Physics ? I really apreciate your answer,thank you.

 

[DeathbyGreen]

Machine learning and physics are fairly separated. If you're looking for programming experience, I think experimental astrophysics would be a good place to look. There are a lot of large data sets to deal with. I'm in condensed matter; I've done experiment and theory, and have a decent amount of coding experience. But if youre actually looking to study things like the genetic algorithm and deep machine learning/AI you'll actually need to be involved in that field. Physics is a discipline that requires a lot of critical thinking skills, and AI isn't close to the point of being able to do "physics". There might be some overlap between biophysics (neuroscience) and neural networking type machine learning though, so maybe you could look into that.

 

[Crass_Oscillator]

Graphical models and neural networks have partial origins in the physics department. Numerous faculty in physics departments have published papers on machine learning topics, especially theoretical statistical physicists (E.g. Pankaj Mehta at Boston U, applying renormalization group methods to deep learning). Sometimes the methods don't have to be physics inspired, but have substantial theoretical overlap (e.g. Mark Newman at Michigan applying replica exchange symmetry to problems of graphical inference).

 

[Choppy]

You might want to look into medical physics.

There is a lot of interest in machine learning in medical physics right now. Machine learning can be used to assist with diagnosis, systematic segmentation of medical images into different tissues and organ, deformable image registration, bioinformatics, clinical decision making tools, workflow management, automated radiotherapy treatment planning, optimization of radiation treatment plans, etc.

 

[Gean Martins]

Thank you so much for helping me, i will look for these topics.

 

[Locrian]

One thing I will warn you about is that the term "machine learning" is vague nearly to the point of uselessness. For instance, DeathbyGreen clearly associates machine learning and deep learning (and this is common), but I would not use them interchangeably. Disagreements like this can occur even in close groups; individuals on my team have radically different opinions of how much "machine learning" there is in our code.

Having said that, I agree with the suggestions in this thread - astrophysics, medical imaging, materials science all have studies that employ machine learning. I would encourage you to branch out as much as you can. If you get comfortable with using machine learning in areas of physics, but have an opportunity to apply it to other systems, do it! Modern statistical learning is a powerful tool, and I think you'd be surprised at the range of interesting problems there are out there to apply it to.

 

[Giulio Prisco]

It turns out there might be intriguing relations between deep learning and fundamental physics:

Quantum fields as deep learning
https://arxiv.org/abs/1708.07408

"Our conjecture also implies a surprising possibility that the quantum fields, and hence matter in the universe, can memorize information and even can perform self learning to some ex tend like DNN in a way consistent with the Strong Church-Turing thesis."

 

[atyy]

There is overlap between statistical physics and machine learning. Here are some examples (related to what @Crass_Oscillator mentioned in post #3).

Yedida, Freeman and Weiss - Understanding Belief Propagation
Jordan, Gharamani, Jaakola, Saul - An Introduction to Variational Methods
Dauphin, Pascanu, Gulcehre, Cho, Ganguli, Bengio - Identifying and attacking the saddle point problem
Hinton et al - Hopfield Nets

Machine learning is mainly just statistics. Most of machine learning uses the same ideas as drawing a straight line for fitting (regression) or classification (logistic regression). In simple statistics, there are few variables and the problems are convex; in machine learning, there are many variables and the problems may be non-convex.

 

[ZapperZ]

It turns out there might be intriguing relations between deep learning and fundamental physics:

Quantum fields as deep learning
https://arxiv.org/abs/1708.07408

"Our conjecture also implies a surprising possibility that the quantum fields, and hence matter in the universe, can memorize information and even can perform self learning to some ex tend like DNN in a way consistent with the Strong Church-Turing thesis."

It's extremely risky to build a career based on someone's "conjecture".

Zz.

 

[Locrian]

Machine learning is mainly just statistics. Most of machine learning uses the same ideas as drawing a straight line for fitting (regression) or classification (logistic regression). In simple statistics, there are few variables and the problems are convex; in machine learning, there are many variables and the problems may be non-convex.

See, I don't agree with this at all. I associate machine learning with tools such as support vector machines, k-means, random forests, etc. Non-traditional predictive analytics.

But like I said, there are people I work with who would disagree with me as well, so no one should take my disagreement as argument, just as an expression of how different various individuals interpret the phrase.

 

[StatGuy2000]

See, I don't agree with this at all. I associate machine learning with tools such as support vector machines, k-means, random forests, etc. Non-traditional predictive analytics.

But like I said, there are people I work with who would disagree with me as well, so no one should take my disagreement as argument, just as an expression of how different various individuals interpret the phrase.

I think you and @atyy are in less disagreement than you may think, since machine learning tools that you mention above are extensions of either classification (k-means, random forests) or regression (support vector machines). I think where you may take issue with is the notion that regression only involves "fitting a straight line" (that really only applies to traditional linear regression).