How to get into the field of machine learning?

Question

IT education

No matter how much time I try to enter the field of machine learning, more and more often I come across the fact that I don’t know where exactly to start. I started reading several books, at the beginning of one - an explanation of the theory of graphs and elements of the theory of entropy, in another book brief data on the theory of probability and mathematical statistics are given, but the problem is that both books provide this material (entropy theory, mathematical statistics, etc.). etc.) very briefly, and it is difficult to understand the essence of the material and from what it follows. From what I read, I realized that you definitely need to know the theory of probability and the mat. statistics, but how much theory really needs to be studied in order to simply train a neural network to recognize text and understand what I did (and not just copy-paste an example). So the question is - can someone from this area explain to me, what exactly needs to be studied and in which direction to dig? That is, do I really need to know well graph theory, entropy theory, mat. statistics (I understood nothing without it), optimization methods and mountains of theory that are given in the preface of these books? How does it really happen?

Answer 1

To understand how a neuron that recognizes symbols works, neither entropy theory nor mathematical statistics are needed. It is enough to understand how the code of this network works. There is nothing more difficult than multiplication.
To start diving, you can take the same playground.tensorflow.org and its code on github.

Answer 2

New things must always be learned on a specific task.
After - we begin to analyze and look for a solution based on current knowledge and experience.
Then, we supplement our assumptions on the solution with information from the Internet.
Again we analyze, search, read.
Sooner or later, we reach the concept of an artificial neuron and read / look: what is it for, etc.
When it is clear: how many neurons need to be connected, exactly how, and most importantly: “WHY?!”, the task will be solved by 50%.
Then - you need to understand: how to do learning / memorization and train.
Fix something (errors - in 100% of cases everyone has them!).
And, at the end of all the work, get a working AI bot with NS and with a complete solution to the task.

Answer 3

the neural network comes down to recalculating the weight (significance) of each neuron,
everything interesting is in the weight recalculation algorithms,
based on mathematical analysis and probability theory,
to start with, google gradient descent, common but not the only one,
simplified - for a specific task we are looking for a method that balances between learning speed and return (reset weights) when correcting errors