PASV - IT школа в США

We use a navigation app to plan a route, look through the films Netflix recommends, tap "read more" on an advertisement that catches our eye, and are pleasantly surprised when unwanted email goes into spam automatically. All of it makes everyday life more comfortable. But there is something else these things have in common: they are possible thanks to machine learning.

What machine learning is

For a product to do its job, it has to be taught properly. That sounds obvious, but how is it done? There are particular algorithms and models by which a machine learns, makes predictions and takes decisions. Where in traditional programming a developer writes out the instructions the software follows, machine learning is a branch of artificial intelligence: the model is given examples, teaches itself and even improves over time. Human intervention is minimal.

How machine learning works

First of all, data. Its quantity and quality matter enormously, so it has to be prepared beforehand — cleaned in places, transformed or normalised in others. Machine learning is a field where attention is required at every stage of working with the product. Handle the data badly and the whole effort becomes ineffective.

Then the right algorithm or model has to be chosen for the task. Using the data prepared earlier, the model finds parameters and learns from them. It is then checked against a test dataset, which lets you assess its accuracy and performance.

Machine learning helps predict values, classify and analyse data, and build whole intelligent systems. How it manages that is what we look at next.

The core concepts

To get a better grasp of the subject, let's look more closely at the types, tasks, algorithms and methods of machine learning.

Types of machine learning

When the data is labelled, the records have inputs and there is a known correct answer, that is called supervised learning. It is used for recognising spam in email, for instance. Unsupervised learning is where the data is unlabelled and the algorithms have to find hidden patterns — customer segmentation or basket analysis, for example. There are other types too:

Semi-supervised — combining supervised and unsupervised learning, since it works with both labelled and unlabelled data. It is used in text processing, for instance.
Reinforcement learning — where algorithms interact with an environment and receive rewards or penalties. Online games are one example.
Transfer learning — where algorithms that have already solved one problem use what they learned to solve another. This is used when there is little data for a new task but a trained model already exists for a similar one.

Machine learning methods

Particular methods are used to analyse data and build models. The main ones:

Classification — sorting data into categories defined in advance.
Regression — predicting numerical values from the data.
Clustering — grouping data into clusters by similarity.
Ensemble methods — built on combining several models.
Dimensionality reduction — used when the number of variables in the data has to be reduced.
Deep learning — using neural networks.

Machine learning algorithms

For software to make decisions without conventional programming, it has to be able to identify patterns in data. That is what machine learning algorithms do: they analyse data and build predictive models. There are various kinds; for clarity, here are the main ones:

Linear regression. Here the dependence of one set of data on another is studied and shown as a straight line. A company's future revenue might be forecast from its current growth figures, for instance.
Decision trees. Another way to understand machine learning is to picture a tree. Every new branch is the answer to one question or a condition. Step by step we reach the leaf nodes, and those are the predictions or the decisions.
Neural networks. These models can be compared to how the human brain works. They contain a great many interconnected neurons able to recognise images, process incoming information and produce a result. That is how voice assistants work, and how a system identifies who is in a photograph.

The algorithms explain machine learning well and give a general sense of the problems it can be applied to.

Machine learning tasks

Bear in mind that behind every result lies a particular method and approach. Here are the most common tasks:

Sorting data into classes, or classification — recognising handwriting, say, or the spam in your inbox.
Predicting numerical values from existing data — forecasting demand for goods, for instance.
Grouping data by similarity, which is what makes machine learning useful for customer segmentation or sorting documents.
Reducing the number of variables without losing important information — removing noise from an image, for example.
Identifying unusual events in data, which is how network security works and how fraudulent transactions are prevented.
Analysing and interpreting visual material — medical diagnosis from a photograph or video, for instance.
Understanding and interpreting human language — chatbots and machine translation.

Having gone through the main algorithms, methods and tasks, we hope machine learning and what it is used for is clearer. But however versatile and automated it becomes, there is always a person behind it. A specialist's involvement is what keeps a model accurate, ethical and relevant.

The PASV machine learning course

At PASV you can take a machine learning course and become a professional in one of the most innovative corners of tech. In eight months you will learn the specialism from scratch to junior level and enter the field with confidence. So that students can fit it around their existing work, the course runs online.

You will learn Python from nothing, take an internship and work on a real project. Your mentors will be working programmers from Silicon Valley, and at the end of the course we will prepare you for interviewing in English so that you can be hired at an international company.