Getting started with neural networks?

Question

Neural networks

Hello! Recently, I have been increasingly surrounded by thoughts about neural networks. Somewhere I saw a video on YouTube about the possibilities, somewhere I heard it out of the corner of my ear, etc.
So far, in my understanding, there is no way at all to make a network that trains itself
.
2) What language? Does it matter? Good knowledge of the language must be?
3) Basic knowledge, how good knowledge of mathematics should be or depends on the task of creating a network?
4) How can I start to get acquainted with this: books, materials? Please advise,
thanks in advance for your replies!

Answer 1

• in JavaScript
  
• Play with TensorFlow
  
• some other toy with NN
  

Answer 2

First you need to install conda, then jupyter notebok and set up all this shit, this is a separate quest. The simplest example is at the beginning of all books, in my opinion it is connected with some kind of plant. It all starts with scikit-learn. And as for the language, it’s only python, other languages ​​​​are only suitable for loading models and getting a forecast from them, and then, for example, models from pytorch you load where the hell, and plus the data must be prepared, and often in a rather specific way, therefore again, it is best to cook in python, because zae ... do it in other languages.

Answer 3

Neural networks are just algorithms, the most resource-intensive part of the process, but not the most important.
The main thing is the preparation of the training sample, i.e. data to feed to networks. There are ready-made implementations, where for typical tasks like recognition, you actually feed a finished image as input and get some kind of result, but you probably have some kind of your own task.
The second problem of neural networks is that they are not cheap, more precisely, the learning process itself, for interesting and not simple academic tasks like 2 + 2 = 4, requires computing resources, and a lot of money.
At the same time, nothing guarantees that after spending a month of computer hours on training (it’s really a month on your computers or you rent it for thousands of dollars on Amazon and calculate everything in a day), you will get a result. And from here the legs of another task grow - how to speed up learning, t .e. can throw out unnecessary data or reduce the network or pre-train it in advance (for example, it is recommended to train the first layer of the network in advance based on the classifier, and then slip this layer into your network - it will speed up training and increase the chance of a positive result at times).