Blog: Artificial Intelligence
This is an article that takes part in a whole work related with Quantum Machine Learning. If you have not checked the previous articles you can read them here: https://firstname.lastname@example.org/quantum-algorithms-2b6c9d7f80b9
As in the previous section, in this one we will introduce the second basic topic of this paper, artificial intelligence. Specifically, we will focus on one of its most worked and exploited branches by both the industry and the research sector: machine learning .
It is a branch of computing in which a series of algorithms are applied so that the computer “learns” about data that is introduced. The vast majority of this data comes in large groups, that is the reason why we call it Big Data. The algorithms this discipline handles can be of various kinds, since it depends on the problem that one wishes to solve or attack (we will encounter three types of learning algorithms).
In turn, this field of AI relias on both statistics and probability since many learning and decision-making techniques are based on techniques from these disciplines.
In the field of machine learning we usually find three types of learning: supervised, unsupervised and reinforcement.
In this type of learning, correct examples of what you want the computer to learn are introduced to it and, with the appropriate algorithm, it must learn to find the solution on its own. Within this type of learning we find different problems, specifically: regression and classification. We are going to analyze them more in detail.
The first, regression, is used when you want to make predictions. We have two types of regressions: linear (a) and non-linear (b)
Both linear and non-linear reflect the relationship between the variables x and y. Where x is the information that we introduce to the algorithm, in other words, with which it “trains”.
In contrast, y is the value that the algorithm returns, that is, what it predicts. In the formulas (1) the coefficients w are the “weights”, these are updated during the training to the point we want. So, we have a good model that describes and reproduces our data. Updating weights is the most common form of learning within machine learning and AI. A method called gradient descent is used, which is nothing more than an optimization algorithm based on the error that is made between the known value and the predicted value. For reasons of space we will not go into detail. Currently, the regression is being used a lot in real estate and the stock market.
On the other hand we have the classification. Similar to the case of regression, we can distinguish between binary or multi-class classification. As in the regression, gradient descent is also used here to update the weights and so that the computer can learn to recognize which class the information belongs to. For example, abinary classification could be to distinguish between cats or dogs in images. This way our classes would be “cat” and “dog” and in the training one would introduce a considerable number of correct images to create a learning rule. Later this would allow us to recognize dogs and cats in images that the algorithm has never seen before. Not only in images, it can also be used to recognize voice patterns, numerical data of various kinds, etc.
Unlike supervised learning, now we do not previously know the correct answer. That is, we only have data that one introduces in the computer. So the main objective of this type of algorithm is looking for patterns, regularities, irregularities or other things in the data. In other words, shred and analyze the data. This is very useful in data science where a person or a team of people must analyze large amounts of data.
Two major problems within unsupervised learning are the detection of anomalies and clustering.
The latter is used when we want to find relationships within the data we enter. An example could be the information a supermarket has about its customers and their tastes. They could get to know what kind of customers buy in their stores as a result of the best selling products. If a person would have to analyze all of these he could spend days or months to reach to the same conclusions. In contrast, a well-applied algorithm in a computer with good computing power could do it in a matter of hours.
As for the detection of anomalies, as its name suggests, it is about finding irregularities in data. That is, applying an algorithm to learn a specific task or to detect a specific pattern and then focus on what does not recognize, ignores or does not fit the learning rule established by the algorithm. This helps us understand the input data. Nowadays a very useful application is in the detection of frauds of all kinds.
This type of learning is the one that is booming the most in the past few years. In some applications the result of a system is a series of actions, for example a game. One must make decisions based on what happens in the game. That is, the information that one receives helps you make future decisions and based on your decisions you can win or lose. So you can say that if you make good decisions you will win the game but if you take bad you will lose. But one can learn from bad decisions in order to avoid the same mistakes in the future. This example of a game is what reinforcement learning is based on.
In reinforcement learning we can find four basic definitions: a state, an action, an agent and what could be called a policy taken to the example of the game. Let’s try to shed light on these definitions by returning to the example of the game. The agent would be the player. The state would be the box in which the agent is or the moment of the game in which it is. The action is what the agent can do based on the information that comes to him, that is, the movement he can make, the possibilities he has, etc. Finally, the policy is the succession of actions carried out by the agent and the state to which he arrives by making those decisions. The goal is to win the game, in this case. The basis of this type of behavior are the Markov chains. They are used to make decisions motivated by a reward or punishment given to the agent by the decision made.
Other applications, apart from games, in which reinforcement learning is taking part in are robotics and autonomous driving. In the first one, this type of learning is used to make decisions and to learn different tasks. In the second, vehicles that drive themselves are getting better and better. It can be seen in companies like Tesla. This will take us to a future in which we will not need a driver’s license.
In this article we have introduced the types of learning we can face within machine learning. This will help us understand the two following articles in which we are going to talk about four different algorithms of machine learning that we are going to use in the final section of this work.
 Ethem Alpaydin, Introduction to Machine Learning, 2nd Edition. The MIT Press, 2010.