Neural networks from scratch. Overview of courses and articles in Russian, free of charge and without registration

On Habré periodically appear reviews of courses on machine learning. But such articles are more often bookmarked than the courses themselves. The reasons for this are different: courses in English, require a confident knowledge of matane or specific frameworks (or, on the contrary, the initial knowledge necessary to complete the course is not described), are on other sites and require registration, have a schedule, homework and are difficult to combine with workdays . All this prevents from now from scratch to begin to dive into the world of machine learning with its own speed, exactly to the level that is interesting and at the same time to skip uninteresting sections.

In this review basically there are only links to articles on Habré, and links to other resources as an add-on (information on them in Russian and do not need to be registered). I read all the articles and materials I recommended personally. I tried each video course to choose what I like and help the others. Most of the articles I have read before, but there are also those that I stumbled upon while writing this review.

The review consists of several sections so that everyone can choose the level from which to start.
For large sections and video courses, approximate time costs, necessary knowledge, expected results and tasks for self-examination are indicated.
')

Most of the articles were not written in a single course, so information can be duplicated. If you see that you know some part of the article, then you can safely skip it, if you did not go bankrupt with this information in the previous article, then you have a chance to read the same thing, but in other words, which should help the absorption of the material.

Introductory articles

Required level: school education, knowledge of Russian language.
Required time: a few hours.

It would seem that it is worth starting the study with the article Artificial Neural Network on Wikipedia, but I do not recommend it. The most interesting description discourages all desire to study neural networks.

Neurons in 5 minutes (too simplified description, for the humanities, but it takes only 5 minutes)
Artificial neural networks in simple words (it is better to spend 15 minutes on this article)
Basics INS (one of the four articles from the Textbook - Neural Networks )
Neural networks for beginners. Part 1 and Part 2
Neural networks, fundamental principles of operation, diversity and topology
Artificial neural networks and minicolumns of the real cortex (ninth part of the course The Logic of Consciousness )

The task

Before proceeding to the next level, create a network in the online designer . Look at all 4 examples, and in the last (spiral) train the network for no more than 100 epochs to an error level of no more than 0.1%, using the minimum number of neurons and layers.

Expand horizons

Required level: basic understanding of the work of neural networks.
Required time: a few hours.

A short course of machine learning or how to create a neural network to solve the scoring problem
The most important thing about neural networks. Lecture in Yandex (I recommend watching only the video for 1 hour, it seemed hard to read the article)
Introduction to Neural Network Architectures
What is a convolutional neural network
Convolutional neural network, part 1: structure, topology, activation functions and training set
Zoo neural network architectures. Part 1 and Part 2 (No special scrutiny, just look at the beautiful pictures and read the description diagonally)

The task

List the main ones:

types of tasks that solve neural networks
types of neural network architectures
activation functions
types of neurons / layers

We deepen knowledge

Required level: understanding of the work of neural networks, knowledge of basic architectures.
Required time: several tens of hours.

A course about Deep Learning on fingers from AFTI NGU (14 videos, 15 hours, will be informative)
OpenDataScience and Mail.Ru Group machine learning materials (10 videos, 20 hours, it will be difficult)
Lectures of the Technosphere. Neural networks in machine learning (14 videos, 25 hours, will be boring)

To decide for myself and help with the choice of the rest of the campers, I built a graph of a fall in interest in the course based on the fall in the number of views of each next video. The conclusions are disappointing - few people reach the end. The largest percentage of those that have come to an end is the course from AFTI NGU.

(The graph of the fall in the number of views was drawn up a couple of months ago and the current picture may differ slightly).

Examples of practical application

This includes basically only those articles, after which the people who read them will be able to reproduce the described results themselves (there are links to source codes or online services)

TOP30 of the most impressive machine learning projects over the past year (v.2018)
Image Enhancement with Neural Network
Detection of body parts using deep neural networks
Object classification in real time
Coloring a black and white photo using a neural network
Change sex and race on a selfie using neural networks
How to distinguish between British and American literature using machine learning
Splitting text into sentences using Tomita Parser
WaveNet: a new model for generating human speech and music
Analysis of the Quran with AI
How many neurons do you need to find out if the Alexander Nevsky Bridge is divorced?
How many cats on Habré?
Trade knows when you are expecting a baby
Stanford Neural Network determines the key text with an accuracy of 85%
Fuel for AI: a selection of open datasets for machine learning

Other materials

Articles and courses that are not included in my review, but you might like it.

Neural networks in pictures: from one neuron to deep architectures (python, numpy)
Basic principles of machine learning on the example of linear regression (python, numpy, matan)
Convolutional neural network, part 2: training in the error back-propagation algorithm (matan)
Neural networks on stepik.org (in a review two years ago, it was already called obsolete)
Course on machine learning on Coursera from Yandex and HSE (the course is available only after registration, NumPy, Pandas, Scikit-Learn)
Deep Learning For Coders (7 videos, 15 hours, English)
Google Deep Learning course at udacity (English)
Course Structuring Coursera Machine Learning Projects (paid, English)