# Machine learning for novices Part 3

## The limitations of neural networks

However, in neural networks there is nothing magical and in most cases, concerns about the script of “Terminator” is not baseless. For example, scientists have trained a neural network to recognize handwritten digits (the application might use, say, mail). How can run this app and why is there nothing to worry about?

For example, we work with images of 20×20 pixels, where each pixel is represented by a shade of grey (256 possible values). In response, we have one of the numbers from 0 to 9. A neural network structure is the following: in the first layer of 400 neurons, where each neuron will be equal to the intensity of the corresponding pixel. In the last layer of 10 neurons where each neuron is the probability that the original image drawn in the corresponding figure. Between them there is a certain number of layers (such layers are called hidden) with the same number of neurons, where each neuron is connected to neuron from the previous layer and with any more.

The edges of the neural network (in the picture they are shown as arrows) will correspond to some number. Where the value of the neuron will be considered as the following sum: value of the neuron of the previous layer * the value of the edges connecting the neurons. Then from this amount takes a certain function (e.g. sigmoid function, which we discussed earlier).

Ultimately, the goal of neural network training is to find such values in the edges to the first layer giving the neural network pixel intensity, at the last layer we have obtained the probability that the image was painted some figure.

In simpler words, in this case a neural network is a computation of a mathematical function, where the arguments are the other mathematical functions that depend on other mathematical functions and so on. Of course, such computation of mathematical functions were fitted to some of the arguments, neither of which existential risk of the question.

### Interesting facts and life hacks

We present several interesting and not entirely obvious examples of the use of machine learning in real life.

For example, the second campaign of Barack Obama was actually won by the best team in the area of data analysis. Of course, we are not talking about what they told him to lie about something, have taken a much more clever way: they chose, in any state, before any audience, on what day and on what topic he should speak. And each time they are measured, and how this affects the polls of the “who would you vote for if the election were held next Sunday?”. In other words, such decisions were not made by spin doctors, and professionals exclusively for data analysis. This becomes especially interesting in light of the fact that, according to experts, this gave him the advantage of 8-10%.

In addition, the modern Internet is difficult to imagine without retargeting, or personalized advertising. Remember: you choose some product online, and after buying a for two weeks you show it in different kind of advertising.

Give life skill. Once you bought the axe, not to admire the axes for another two weeks, go to the fashion store. Further, depending on your gender preference, visit either the men’s or women’s undergarments. Throw a few items in the basket, close the basket and walk away. These guys have very large budgets on retargeting: the next two weeks, depending on your preference, will be pursued half-dressed or beautiful men, or beautiful half-dressed women. It’s better than the axe.

**Recommendations for training**

If you have the desire to learn machine learning technology, it should be noted that for in-depth study of the subject understanding of all of the Foundation necessary to understand at least in mathematical analysis, linear algebra and optimization techniques (with emphasis on the last two). In addition, it is desirable to know the basics of programming and some programming language. In machine learning, typically used R, Python or Matlab.

For the initial study is perfect classic course on Coursera on machine learning. Its main feature is the exceptional practical orientation and strong background of mathematics knowledge in this course is not required.

To motivate the study in the following way: view curriculum. If you review all the lectures and solve all the tasks, then you will be able to apply these things in practice. In particular, you will master algorithms of linear regression (regression task), logistic regression, support vector machine, neural network (classification problem), K – means (clustering), PCA (reduction of dimension), anomaly detection and learn how to build a recommendation system (recommendation of the products according to your estimates, for example, movies or songs). In the last week of the course you will also get advice on how to build a system that will detect and decipher the inscriptions on the images or to identify people.