How should a physicist learn Machine Learning?

[Phylosopher]

TL;DR Summary

Doing my master thesis in physics using machine learning (geared toward physics not machine learning). I can not afford learning neural networks "the regular way" because lack of time. What should I start with?

Hello everyone,
I am currently a master student working on physics and neural networks. I have already started producing neural network results (I use tensorflow and keras) so I know how to program the basic things that I am required to do, the problem is that I do not understand them well.

I do not have a good solid background when it comes to the neural networks themselves, and as a master student working on his thesis, I can not/I do not want to write my thesis blindly. I want to truly understand the behavior of my program.

Problem is, I can not afford reading a 600 page book on ML! My idea was to read the papers that are mentioned in the tensorflow pages once I use a specific function or class(Example: Here). But as I said, I do not have a solid background to read them properly.
So, my question is, what resources you think I should read first before I delve deeper in these papers? I need the bare minimum so I can accelerate my learning.

Things I have bookmarked so far that I think are useful:: Quick hand on introduction, Intro book, More detailed Intro book ... Do you think these are good starting points? Do you have better suggestions?

 

[pbuk]

Quick hand on introduction - I can thoroughly recommend this, it should help you develop a deep intuition for the process and could be all you need.

Intro book - I'm not familiar with this but looking at the ToC it might be a good fit.

More detailed Intro book - this seems to have a much wider scope than you need but might be good context.

 

[Borg]

This short, four video set is the best explanation of neural networks that I've seen in my three years of study - 3Blue1Brown Neural Networks.

 

[atyy]

The minimum ideas of machine learning are that deep learning is essentially curve fitting or classification. A neural network simply parameterizes a very complicated function. A traditional algorithm for curve fitting is linear regression. A traditional algorithm for classification is logistic regression.

Compared to basic statistical methods, an odd thing about neural networks is that there are many more parameters than data points. There are tricks that prevent overfitting. However, why overfitting can be avoided is not yet well understood.

Also, methods of estimating errors in deep learning are not well understood. However, there are reasonable attempts.

 

[anorlunda]

The minimum ideas of machine learning are that deep learning is essentially curve fitting or classification

That is the essential lesson. It is possible that more than one method could lead to the same or a very similar curve fit. It's possible even if those other methods are not similar to deep learning.

So if you are just using the fit, or just using a tool to create the fit, it is not essential to learn the methods used.

 

[FactChecker]

The minimum ideas of machine learning are that deep learning is essentially curve fitting or classification. A neural network simply parameterizes a very complicated function. A traditional algorithm for curve fitting is linear regression. A traditional algorithm for classification is logistic regression.

Although I wouldn't want to make deep learning sound too ordinary. Even for someone very used to curve fitting, there are many aspects and capabilities of deep learning that are surprising, powerful, and fascinating.