The mother of all artificial intelligence algorithms is machine learning or machine learning: the ability to learn and execute tasks by the machine based on algorithms that iteratively learn from data. This is not a recent discipline: some machine learning algorithms have been widespread for years, but what has changed today is the large mass of data on which to apply complex mathematical calculations and given that the most important aspect of machine learning is the repetitiveness, the more models are exposed to data, the more they are able to adapt independently.
There are several machine learning techniques:
supervised learning: the algorithm is given data associated with information that interests us and, based on these data, the algorithm learns to understand how to behave; in practice, the algorithm's task is to relate known data by defining a model on the basis of which to give a correct result to the occurrence of an event not present in the previous series. A classic example is the classification of potential customers based on the profile and purchase history of other customers.
Unsupervised learning: in this case, previously classified data are not used. The algorithm must be able to derive a rule to group the cases that arise deriving the characteristics from the data themselves. As one can easily imagine, we are in the presence of a much more complicated methodology than the previous one, where the algorithm is much more complex since it is a matter of extracting from the data information not yet known. It is used for the definition of homogeneous groupings of cases: in the medical field, for example, this methodology can be used to define a new pathology based on data that until that moment had not been related. It is the methodology that has been developed with greater effectiveness thanks to the growth of big data.
Reinforcement learning: The algorithm knows the objective to be reached (for example winning a chess game) and defines how to behave based on a situation (the configuration of the board) those changes (following the opponent's moves). The learning process progresses through "rewards", defined precisely reinforcement (valid moves). Learning is continuous and, just like for human players, the more the machine "plays" the better it becomes. The Deep Blue-Kasparov duel is famous: specifically designed by IBM to play chess, Deep Blue lost the first tournament played with the chess champion in the "game 1" of February 1996 but got the better of Kasparov in the "game 6" played a year later (after losing 2 more matches and having 2 draws).
These techniques are substantiated by the application of different types of algorithms. Here are some of the most important algorithmic techniques and methods used in machine learning.
Decision trees: used especially in inductive learning processes based on the observation of the surrounding environment from which the input variables (attributes) derive. The decision-making process is represented by an inverted logical tree where each node is a conditional process that is a sequence of tests that starts from the root node and proceeds downwards choosing a direction rather than another based on the values detected. The final decision is in the terminal leaf nodes. Among the advantages, there is the simplicity and the possibility, for the man, to verify through which process the machine has reached the decision. The disadvantage is that it is a technique that is not suitable for complex problems.
Bayesian classifiers: it is based on the application of the Bayes theorem (from the name of the British mathematician who, in the 18th century, developed a new approach to statistics) that is used to calculate the probability of a cause that triggered the verified event For example: having ascertained that the high presence of cholesterol in the blood can cause thrombosis, a certain cholesterol value is detected, what is the probability that the patient is struck by thrombosis? Bayesian classifiers have different degrees of complexity.
Support vector machines(SVM, from the English Support Vector Machines): they are supervised learning methodologies for regression and classification of patterns and belong to the family of maximum margin classifiers (linear classifiers that simultaneously minimize the empirical classification error and maximize the geometric margin, i.e. the distance between a certain point x and the hyperplane which, in turn, is a linear subspace of a dimension smaller than one (n - 1) with respect to the space in which it is contained (n)).
In these machines, the learning algorithms are decoupled from the application domain which is coded exclusively in the design of the kernel function; this function maps the data based on multidimensional characteristics and allows to create an approximate model of the real world (3D) starting from two-dimensional data (2D).
The most common application of SVM is an artificial vision: in the image of a group where men and women are present (based on the kernel function that defines sex considering various parameters) the SVM manages to separate one from the other. Another thing that is important to know, in the logic of this article that intends to provide only some basic indications, is that these classifiers are opposed to the classical techniques of training of artificial neural networks.
Ensemble learning: it is the combination of different methods (starting from the Bayesian classifiers) to obtain a better predictive performance than the single methods it combines. Based on the "weight" that is given to the various methods, ensemble learning is divided into 3 fundamental techniques (bagging, boosting and stacking).
The principal component analysis (in English PCA - Principal Component Analysis): it is a data simplification technique whose purpose is to reduce the more or less large number of variables that describe a set of data to a smaller number of latent variables, limiting the loss of information as much as possible.
We hope you have learned about machine learning through this comprehensive guide. Share your thoughts with us in the comments below!
Nov 08, 2019