The human brain performs a lot of complex analytical tasks: we learn to find patterns, unite objects and phenomena by common features, recognize and understand languages, both living and programming, collect data and make forecasts: for example, whether the shares of the company that we invest in will grow and fall.
Machines learn to do the same, but on a different level of complexity, using mathematical logic.
What is ML and where is it used?
Based on the downloaded data and algorithms, computers carry out analytical work, getting results even faster than people. It is called machine learning (ML).
Where machine learning is used:
- recognition of natural speech (Siri, Alice and other virtual assistants);
- recognition of handwritten letters and numbers;
- identification of language in the document;
- recommendations on websites (similar articles, products, music);
- spam filters;
- search for similar documents;
- determination of suspicious transactions;
- forecast of the securities value;
- analysis of demand, sales volume;
- detectors of abnormal behavior (for example, online casinos visitors);
- risk management;
- self-driving cars and smart devices: for example, robotic vacuum cleaners;
- computer games.
The list can be extended by another 20-30 items, since computers become more complex, and methods of machine learning are being improved. Therefore, let’s clarify another issue: how it works.
Data is a key
Datasets, or structured databases, are important in the machine learning process. Computer needs much more information to learn in order to recognize objects correctly. For example, a programmer from the Netherlands used artificial intelligence to teach the entrance door to let the cat in the house. To make the program differentiate the pet from the rest, he downloaded about a hundred photos.
If the task is more complicated, for example, to teach the machine to distinguish the image of any cat from photos of other animals, the required number of images will be tens of thousands. After processing 10 000 pictures marked as “a cat”, computer with a high degree of probability will recognize the 10 001st unsigned image as “a cat”.
Great, if we already have a good selection of pictures with cats. But this is not always the case: large and structured databases are expensive. The good news is that machine learning can also be carried out using unstructured databases involving relevant algorithms.
Algorithms: choose the path
One and the same problem almost always has different ways of solving. The method or algorithm for ML is chosen depending on the type of task and database that is available to developers.
The main types of machine learning:
- deep learning.
Traditional learning is usually divided into two categories – “with a teacher” and “without a teacher”. In the first case, the data is marked out, and in the second one, the machine analyzes them on its own and looks for regularities. The result “with the teacher” is faster and more precise, but it can be difficult to find or create a suitable database.
Environmental learning is used where the task of the robot is to survive in a certain environment (real or virtual). In such a way characters (NPCs) in computer games, robotic vacuums and self-driving cars are trained. The machine doesn’t operate on data, it generalizes the situation and tries to get out of it with real benefits.
The method of ensembles involves several machines with different learning methods, and they are taught to correct each other's mistakes.
Deep learning is used for neural networks. These are structures from connected and interacting simple processors that mimic the work of brain neurons. Today, this is the most promising area of ML: it is neural networks that are used to created services for speech and image recognition.
Machine learning in the future
The development of ML limited by the available technology level. The first self-organizing network for image recognition was created in 1975. But until powerful processors appeared, we could not upload a photo to the network and find out who it is.
The next breakthrough in machine learning is associated with quantum technologies, as well as the further development of Data Mining (data collection and organization): as we know, machine learning is impossible without data. Improvement of cloud technologies will also give a new space for the ML development.
Probably, by 2025, neural networks will become more complex and will form a basis for new devices and technologies. Voice assistants will become more useful, smarter and more accessible; robots will master new professions. We will see unmanned vehicles and get more convenient Internet services.