Brief introduction to machine learning

The recent growth in the volume of data created by mobile devices and social networks has dramatically impacted the need for high performance computation and new methods of analysis. Historically, large quantities of data (big data) were analyzed by statistical approaches which were based on sampling and inductive reasoning to derive knowledge from data. A more recent development of artificial intelligence, and more specifically, machine learning, enabled not only the ability to deal with large volume of data, but it brought a tremendous value to businesses and consumers by extracting valuable insights and hidden patterns.

Machine learning is not new. In 1959, Arthur Samuel defined machine learning as:

Within the field of data analytics, machine learning is a method used to devise complex models and algorithms that allow to This approach is similar to a person who increases his knowledge on a subject by reading more and more books on the subject. There are three main approaches in machine learning: supervised learning, unsupervised learning, and reinforcement learning.

Supervised learning assumes that we know what the outputs are of each data point. For example, we learn that a car that costs $80,000, which has an electric engine and acceleration of 0-100 km/h in 3 seconds, is called Tesla; another car, which costs $40,000, has a diesel engine, and acceleration of 0-100 km/h in 9.2 seconds, is called Toyota; and so on. Then, when we look for the name of a car which costs $35,000, has acceleration of 0-100 km/h in 9.8 seconds, and has a diesel engine, it is most probably Toyota and not Tesla.

Unsupervised learning is used when we do not know the outputs. In the case of cars, we only have technical specifications: acceleration, price, engine type. Then we cluster the data points into different groups (clusters) of similar cars. In our case, we will have the clusters with similar price and engine types. Then, we understand similarities and differences between the cars.

The third type of machine learning is reinforcement learning, which is used more in artificial intelligence applications. It consists of devising an algorithm that learns how to behave based on a system of rewards. This kind of learning is similar to the natural human learning process. It can be used in teaching an algorithm how to play chess. In the first step, we define the environment—the chess board and all possible moves. Then the algorithm starts by making random moves and earns positive or negative rewards. When a reward is positive, it means that the move was successful, and when it is negative, it means that it has to avoid such moves in the future. After thousands of games, it finishes by knowing all the best sequences of moves.

In real-life applications, many hybrid approaches are widely used, based on available data and the complexity of problems.