Machine learning

Machine learning is an extensive subsection of artificial intelligence that studies methods for constructing algorithms capable of learning. There are two types of training. Case study, or inductive learning, is based on the identification of patterns in empirical data. Deductive training involves the formalization of expert knowledge and their transfer to a computer as a knowledge base . Deductive learning is usually referred to the field of expert systems, therefore the terms machine learning and learning from precedents can be considered synonymous.

There are many objects (situations) and many possible answers (responses, reactions). There is some correlation between answers and objects, but it is not known. Only a finite set of precedents is known — the “object, response” pairs, called the training set. On the basis of these data, it is required to restore the dependence, that is, to build an algorithm capable of producing a sufficiently accurate answer for any object. In order to measure the accuracy of the answers, the quality functional is introduced in a certain way.

This formulation is a generalization of the classical problems of approximation of functions. In classical problems of approximation, objects are real numbers or vectors. In real applications, input data about objects can be incomplete, inaccurate, non-numeric, heterogeneous. These features lead to a wide variety of machine learning methods.

Since the machine learning section, on the one hand, was formed as a result of the separation of the science of neural networks into network training methods and types of network architecture topologies, and on the other, it incorporates mathematical statistics methods, the following machine learning methods are based on neural networks. That is, the basic types of neural networks, such as the perceptron and multilayer perceptron (as well as their modifications) can be trained as with a teacher, without a teacher, with reinforcement, and actively. But some neural networks and most statistical methods can be attributed to only one of the ways of learning. Therefore, if it is necessary to classify machine learning methods depending on the learning method, then, regarding neural networks, it is not correct to refer them to a particular type, but to classify neural network learning algorithms more correctly.

• Active learning is different in that the learning algorithm has the ability to independently assign the following situation to be studied, in which the correct answer will be known:
• Teaching with partial involvement of a teacher - for a part of precedents, a pair of “situation, required solution” is set, and for part, only “situation”
• Transductive training - training with partial involvement of the teacher, when the forecast is supposed to be done only for the precedents from the test sample
• Multitasking training - simultaneous training of a group of interrelated tasks, each of which is assigned its own pairs of “situation, required solution”
• Multivariate training - training, when precedents can be combined into groups, in each of which there is a “situation” for all precedents, but only for one of them (and it is not known what) there is a pair of “situation, the required solution”

Types of input data

• Characteristic description of objects is the most common case.
• Description of relationships between objects, most often of pairwise similarity, expressed using a distance matrix, kernels or data graph
• Time series or signal.
• Image or video series.

Types of quality functionals

• When training with a teacher - the quality functional can be defined as the average error of the answers. It is assumed that the desired algorithm should minimize it. To prevent retraining, a regularizer is often added explicitly or implicitly to the quality functional to be minimized.
• When training without a teacher, the quality functionals can be defined differently, for example, as the ratio of the average intercluster and intracluster distances.
• When training with reinforcements, quality functionals are determined by the physical environment, which indicates the quality of the agent's adaptation.

Spheres of application

The goal of machine learning is the partial or complete automation of solving complex professional problems in various fields of human activity. The scope of machine learning applications is constantly expanding. Widespread informatization leads to the accumulation of huge amounts of data in science, industry, business, transport, and health care. The problems of forecasting, management and decision-making that arise in this case are often reduced to learning from precedents. Earlier, when there was no such data, these tasks were either not set at all or were solved by completely different methods.