Supervised Statistical Machine-Learning

The aim of this article is to present the main principles of supervised statistical artificial learning, as well as an overview of the main algorithms and techniques in this field.

It is these statistical learning techniques that have made it possible to automate increasingly complex tasks, and thus enabled the development and deployment in everyday life of applications such as :

• automated handwriting recognition, e.g. for bank cheques and addresses on envelopes;

• voice recognition, particularly for automatic telephone systems;

• recommendation systems" capable of proposing personalized selections (i.e. adapted to individual tastes and/or interests) of music, videos or information.

Artificial learning is also at the heart of future, even more sophisticated systems such as :

• intelligent" robotics (as opposed to the robot-automats of the 1970s, which performed identical repetitive actions);

• automated driving of driverless cars;

• conversational agents" and automated translation, etc.

In addition, the increasing "digitization" of the world, and the resulting avalanche of data, has meant that over the last twenty years or so, statistical learning has become increasingly popular, covering an ever wider field. Indeed, as information from virtually all sectors (human activities such as commerce, the economy, culture, leisure, etc., but also health and the environment) is increasingly encoded, recorded and transmitted in computerized form, it is becoming possible to access, at virtually no cost, a vast quantity of data to which statistical analysis can be applied, and hence empirical modeling by artificial learning. Last but not least, learning algorithms have made significant progress over the same period, and even more so over the last five years with the development of deep learning, enabling a wide range of engineering problems to be addressed with "industrializable" performance.