Machine learning is a multidisciplinary field created at the intersection of, and by the synergy between, computer science, statistics, neurobiology, and control theory. Its emergence has played a key role in several fields and has fundamentally changed the vision of software programming. If the question before was, how to program a computer, now the question becomes is how computers will program themselves. Thus, it is clear that machine learning is a basic method that allows a computer to have its own intelligence.

As might be expected, machine learning interconnects and coexists with the study of, and research on, human learning. Like humans, whose brain and neurons are the foundation of insight, Artificial Neural Networks (ANNs) are the basis of any decision-making activity of the computer.

Machine learning refers to the ability to learn from experience without any outside help, which is what we humans do, in most cases. Why should it not be the same for machines?

From a set of data, we can find a model that approximates the set by the use of machine learning. For example, we can identify a correspondence between input variables and output variables for a given system. One way to do this is to postulate the existence of some kind of mechanism for the parametric generation of data, which, however, does not know the exact values of the parameters. This process typically makes reference to statistical techniques, such as the following:

• Induction
• Deduction
• Abduction

The extraction of general laws from a set of observed data is called induction; it is opposed to deduction, in which, starting from general laws, we want to predict the value of a set of variables. Induction is the fundamental mechanism underlying the scientific method in which we want to derive general laws, typically described in a mathematical language, starting from the observation of phenomena.

This observation includes the measurement of a set of variables and, therefore, the acquisition of data that describes the observed phenomena. Then, the resultant model can be used to make predictions on additional data. The overall process in which, starting from a set of observations, we want to make predictions for new situations, is called inference. Therefore, inductive learning starts from observations arising from the surrounding environment and generalizes obtaining knowledge that will be valid for not-yet-observed cases; at least we hope so.

Inductive learning is based on learning by example: knowledge gained by starting from a set of positive examples that are instances of the concept to be learned, and negative examples that are non-instances of the concept. In this regard, Galileo Galilei's (1564-1642) phrase is particularly clear:

The following diagram consists of a flowchart showing inductive and deductive learning:

A question arises spontaneously: why do machine learning systems work where traditional algorithms fail? The reasons for the failure of traditional algorithms are numerous and typically include the following:

• Difficulty in problem formalization: For example, each of us can recognize our friends from their voices. But probably none can describe a sequence of computational steps enabling them to recognize the speaker from the recorded sound.
• High number of variables at play: When considering the problem of recognizing characters from a document, specifying all parameters that are thought to be involved can be particularly complex. In addition, the same formalization applied in the same context but on a different idiom could prove inadequate.
• Lack of theory: Imagine you have to predict exactly the performance of financial markets in the absence of specific mathematical laws.
• Need for customization: The distinction between interesting and uninteresting features depends significantly on the perception of the individual user.

A quick analysis of these issues highlights the lack of experience in all cases.

The power of machine learning is due to the quality of its algorithms, which have been improved and updated over the years; these are divided into several main types depending on the nature of the signal used for learning or the type of feedback adopted by the system.

They are as follows:

• Supervised learning: The algorithm generates a function that links input values to a desired output through the observation of a set of examples in which each data input has its relative output data, which is used to construct predictive models.
• Unsupervised learning: The algorithm tries to derive knowledge from a general input without the help of a set of pre-classified examples that are used to build descriptive models. A typical example of the application of these algorithms is found in search engines.
• Reinforcement learning: The algorithm is able to learn depending on the changes that occur in the environment in which it is performed. In fact, since every action has some effect on the environment concerned, the algorithm is driven by the same feedback environment. Some of these algorithms are used in speech or text recognition.

The subdivision that we have just proposed does not prohibit the use of hybrid approaches between some or all of these different areas, which have often recorded good results.

Supervised learning is a machine learning technique that aims to program a computer system so that it can resolve the relevant tasks automatically. To do this, the input data is included in a set I (typically vectors). Then, the set of output data is fixed as set O, and finally, it defines a function f that associates each input with the correct answer. Such information is called a training set. This workflow is presented in the following diagram:

All supervised learning algorithms are based on the following thesis: if an algorithm provides an adequate number of examples, it will be able to create a derived function B that will approximate the desired function A.

If the approximation of the desired function is adequate, then when the input data is offered to the derived function, this function should be able to provide output responses similar to those provided by the desired function and then acceptable. These algorithms are based on the following concept: similar inputs correspond to similar outputs.

Generally, in the real world, this assumption is not valid; however, some situations exist in which it is acceptable. Clearly, the proper functioning of such algorithms depends significantly on the input data. If there are only a few training inputs, the algorithm might not have enough experience to provide a correct output. Conversely, many inputs may make it excessively slow, since the derivative function generated by a large number of inputs increases the training time. Hence the slowness.

Moreover, experience shows that this type of algorithm is very sensitive to noise; even a few pieces of incorrect data can make the entire system unreliable and lead to wrong decisions.

In supervised learning, it's possible to split problems based on the nature of the data. If the output value is categorical, such as membership/non-membership of a certain class, then it is a classification problem. If the output is a continuous real value in a certain range, then it is a regression problem.

The aim of unsupervised learning is to automatically extract information from databases. This process occurs without a priori knowledge of the contents to be analyzed. Unlike supervised learning, there is no information on the membership classes of examples, or more generally on the output corresponding to a certain input. The goal is to get a model that is able to discover interesting properties: groups with similar characteristics (clustering), for instance. Search engines are an example of an application of these algorithms. Given one or more keywords, they are able to create a list of links related to our search.

The validity of these algorithms depends on the usefulness of the information they can extract from the databases. These algorithms work by comparing data and looking for similarities or differences. Available data concerns only the set of features that describe each example.

The following diagram shows supervised learning (on the left) and unsupervised learning examples (on the right):

They show great efficiency with elements of numeric type, but are much less accurate with non-numeric data. Generally, they work properly in the presence of data that is clearly identifiable and contains an order or a clear grouping.

Reinforcement learning aims to create algorithms that can learn and adapt to environmental changes. This programming technique is based on the concept of receiving external stimuli, the nature of which depends on the algorithm choices. A correct choice will involve a reward, while an incorrect choice will lead to a penalty. The goal of the system is to achieve the best possible result, of course.

In supervised learning, there is a teacher that tells the system the correct output (learning with a teacher). This is not always possible. Often, we have only qualitative information (sometimes binary, right/wrong, or success/failure).

The information available is called reinforcement signals. But the system does not give any information on how to update the agent's behavior (that is, weights). You cannot define a cost function or a gradient. The goal of the system is to create smart agents that have machinery able to learn from their experience.

When developing an application that uses machine learning, we will follow a procedure characterized by the following steps:

• Collecting the data: Everything starts from the data, no doubt about it; but one might wonder from where so much data comes. In practice, it is collected through lengthy procedures that may, for example, derive from measurement campaigns or face-to-face interviews. In all cases, the data is collected in a database so that it can then be analyzed to derive knowledge.
• Preparing the data: We have collected the data; now, we have to prepare it for the next step. Once we have this data, we must make sure it is in a format usable by the algorithm we want to use. To do this, you may need to do some formatting. Recall that some algorithms need data in an integer format, whereas others require data in the form of strings, and finally others need it to be in a special format. We will get to this later, but the specific formatting is usually simple compared to the data collection.
• Exploring the data: At this point, we can look at data to verify that it is actually working and that we do not have a bunch of empty values. In this step, through the use of plots, we can recognize patterns and whether or not there are some data points that are vastly different from the rest of the set. Plotting data in one, two, or three dimensions can also help.
• Training the algorithm: Now, let's get serious. In this step, the machine learning algorithm works on the definition of the model and therefore deals with the training. The model starts to extract knowledge from the large amounts of data that we had available, and from which nothing has been explained so far. For unsupervised learning, there's no training step because you don't have a target value.
• Testing the algorithm: In this step, we use the information learned in the previous step to see if the model actually works. The evaluation of an algorithm is for seeing how well the model approximates the real system. In the case of supervised learning, we have some known values that we can use to evaluate the algorithm. In unsupervised learning, we may need to use some other metrics to evaluate success. In both cases, if we are not satisfied, we can return to the previous steps, change some things, and retry the test.

• Evaluating the algorithm: We have reached the point where we can apply what has been done so far. We can assess the approximation ability of the model by applying it to real data. The model, previously trained and tested, is then valued in this phase.
• Improving algorithm performance: Finally, we can focus on the finishing steps. We've verified that the model works, we have evaluated the performance, and now we are ready to analyze the whole process to identify any possible room for improvement.