Machine learning. Basic concepts.
1/ What is machine learning
Machine learning is a branch of artificial intelligence. In the description of artificial intelligence, terms such as neural networks, machine learning and deep learning usually converge. There is often confusion with the terms, so it is better to consider them using the example of a drawing.
Machine learning involves self-learning and the development of a computer system based on data prepared by people. The program receives a set of data. Then he learns to recognize them by certain signs and perform certain operations with them. Machine learning specialists strive to achieve from artificial intelligence the ability to find answers to questions and perform certain highly specialized tasks.
It’s worth remembering that machine learning and programming are not the same thing. The programmer needs to think through the logic of the program and prescribe all the options for the conditions of the problem and solutions to them. In machine learning, we bring input data and the corresponding result, and the computer builds logic between them. If successful, we will be able to give the machine similar tasks, and it will solve them faster for us and thereby save time and resources.
2/ How does machine learning work.
The computer does not understand what they are trying to teach it. For a computing system, the world consists of zeros and ones. And before solving a problem, the system needs to learn to understand its condition and logic. Based on the input data, the computing system develops a kind of analytical thinking. It reads data, classifies it, sees patterns and remembers.
Artificial intelligence is not so important whether it will eventually have to drive a vehicle or assess the degree of maturity of vegetables and fruits. The computer learns, but if something goes wrong, it has to be retrained and then the machine tries to avoid mistakes.
So, step by step, the machines are progressing. Sometimes the presence of a human observer is necessary not only in the learning process, but also in the process of controlling the machine. However, unlike a human, a computer is devoid of emotions, fatigue and cognitive distortions, so it often copes with tasks better.
3/ Types of machine learning
Among the main methods of computer training, there are 4 types:
A) classical training,
B) ensemble methods,
C) reinforcement learning,
D) neural networks and deep learning.
A) Classical learning
The oldest and simplest machine learning methods from the middle of the XX century, which were originally used to find patterns in numbers and calculate the proximity of points in space. Experts divide such methods into two categories.
With the teacher. The teacher program, just like at school, tells the machine what data is in front of it, how to distinguish one from the other and how they are similar. At the same time, the data is already divided by attributes, so learning is faster.
This category is characterized by two types of tasks:
Classification — division of data by a certain attribute or category. A triangle is different from a circle, but according to certain characteristics, it is possible to group a triangle and a circle into geometric bodies.
Regression is the same classification, only by numerical criteria. The machine finds the dependence of some numerical data on others and then can predict: if there is one number at the input, then there will be another at the output.
Without a teacher. The second method is slower: the computer is not informed of the signs and it looks for patterns in the data on its own. It’s as if Cinderella didn’t know how barley and millet differ, and before separating them from each other on the instructions of an evil stepmother, she would have to figure out the characteristic features of the two cereals.
Here they take the stage:
- Clustering — separation of data by features that seemed similar to the machine.
- Dimension reduction or generalization — combining several features into more general or abstract classes.
- Search for rules — the machine analyzes the data flow and finds complementary ones — those that often occur together.
B) Ensemble methods.
Ensembles are a way in which several unstable machine learning methods combine to correct each other’s errors. Depending on the input data, the result may be unpredictable. Nevertheless, the quality of this approach is head and shoulders higher than any of the methods mentioned earlier, if you assemble ensembles in one of three ways.
- Boosting. A method of sequential learning of algorithms. After processing one set of data, the machine is given the next one. Data is specially added to it, the result of which differs from the desired one and the algorithm tries to find a solution. Experts call the accuracy of the method extremely high, but there is also a minus — the lack of parallel execution of operations, which significantly slows down the speed of data processing.
- Begging. The same algorithm is trained on random samples of data, possibly repeated. The algorithm evaluates the data and gives the result to those who have scored the highest number of points. The advantage of this method over boosting in parallel operations is that the processing speed increases enormously.
- Stacking. The data is passed through several weak algorithms, after which the results are run through another, decisive one. Often the last method in the chain is regression. The purpose of the method is to teach weak students, but in practice the accuracy is still low and the approach is rarely used.
C) Reinforcement Learning
The method involves training the machine through trial and error. The computer is rewarded for “correct” actions, and fined for “incorrect” actions. Yes, it didn’t seem to you: all this is very similar to dog training, and the computer in this case is literally Pavlov’s dog.
Initially, situations are simulated in a virtual environment, after which artificial intelligence continues to learn in the real world. The more trial and error, the more rewards in situations that the computer tries on the newly occurring events. And he eventually arrives according to the scenario that will bring the maximum reward.
D) Deep learning
Deep learning is a subspecies of machine learning. It differs from the classical types by the types of data being processed and the methods of their processing. The fact is that classical algorithms required notations added to the data by a specialist before training. With the advent of deep learning, the need for a specialist has disappeared: the machine independently understands what belongs to what in the input data.
There are quite a few neural network architectures. Consider the two most popular: convolutional and recurrent.
- Convolutional. Not so long ago, the elements of objects were prescribed manually. In the example with the cat and the courier, this would mean that the courier always had with him a detailed specification for toys: what size and color they are, whether they have feathers on them and whether they can squeak and make other sounds. However, even with minimal changes in the elements of objects, the neural network stopped defining them. Soon the researchers realized that it was enough to give the neural network a large amount of data and it would find patterns on its own.
- Recurrent. At the time of the appearance of such networks, each time they were launched as the first: there was no memory in the neurons. Memory was added to the neurons of recurrent networks, so they did not make previous mistakes on each new learning cycle. But in this case, an overabundance of data was obtained. Therefore, neurons have been taught to identify information that is unnecessary for learning and forget it.
4. What applied tasks can really be solved by machine learning
Machine learning can be used to solve a wide range of tasks in a variety of fields: research, medicine and pharmacology, protein structure prediction, image processing, predicting the likely financial results of business decisions, target audience search and automation of many production and logistics processes. The intensity of the use of machine learning directly depends on what everyday tasks and new challenges humanity faces day after day.