1
From Human to Artificial Intelligence

1.1. The different forms of intelligence

1.1.1. Human intelligence typologies

DEFINITION.- Etymology of the word intelligence: Borrowed from the Latin intelligentia, "faculty of perceiving, understanding, intelligence", derived from intellegere ("to discern, grasp, understand"), composed of the prefix inter- ("between") and the verb legere ("to gather, choose, read"). Etymologically, intelligence consists of making a choice, a selection.

1.1.2. Artificial intelligence (AI) and human intelligence

One of the main challenges of artificial intelligence is that it can only simulate one or a limited number of forms of human intelligence! There are many forms of intelligence, as shown in Figure 1.1.

1.1.3. Object intelligence and human assistance

The human mind can use conventional statistical modeling to plot graphs with dozens of variables, but the decision resulting from these calculations remains more complex. A computer with artificial intelligence software can test and then sift through millions of variables with incredible speed, helping human intelligence to increase its analytical capacity.

Figure 1.1. The different forms of intelligence

1.2. History of "artificial" intelligence1

1.2.1. Mechanical forms

The first phase of this process was the desire to "mechanize" man and their thought processes. The idea was to represent man as a machine, in order to design learning machines.

The first works are attributed to the Catalan Ramon Lulle (1232-1315). Theological predicates, subjects and theories were organized in geometrical figures. These were considered perfect: circles, squares and triangles. By activating a mechanism consisting of dials, levers and cranks, a wheel could be turned, and the propositions of the thesis moved on guides to position themselves on points. These points were used to affirm that the proposition was positive, or true, or, conversely, negative, and therefore false. With the attributes of God and nine questions, the machine proved the existence of God!

Giordano Bruno2, Nicolas de Cues3, Athanasius Kircher4 and Gottfried Leibniz5 were influenced by the work of Ramon Lulle. Descartes (1596-1650) drew a parallel between the animal-machine of this period and the man-machine. The animal is nothing more than a perfected machine, a clockwork of metal parts and springs.

This era is represented by Maillard's6 famous "swan", capable of swimming (1733), and Jacques Vaucanson7 "duck", which simulated digestion (1738). Julien Offray de la Mettrie (1709-1751) published a work entitled: L'Homme-machine8 (The Man-Machine).

Figure 1.2. Vaucanson's duck, 1738 (model by Riskin Jessica). Illustration by a 19th-century inventor of his own version of a mechanical digestive duck. An arrow indicates where the main action takes place. Based on Chapuis and Édouard Gélis, Le Monde des Automates (The World of Automata).

1.2.2. The desire to model neurons and cybernetics

DEFINITION.- Physiologist: The physiologist conducts research into the functions of the body's physiological systems: cardiovascular, digestive, musculoskeletal, nervous, respiratory, reproductive and urinary. He examines how these systems function, under both normal and pathological conditions. After identifying the various systems, he integrates them and builds models ranging from the molecular and cellular scales to those of tissues, organs and the whole organism.

DEFINITION.- Logician: specialist in logic as a discipline.

Warren Sturgis McCulloch, physiologist (1898-1969), joined forces with William Pitts, logician (1923-1969), to carry out work in the field of intelligence9. These two actors wanted to define a way of calculating nervous activity. McCulloch refers to Descartes in these works, as well as to Leibniz, whom he considers to be their precursors. They speak of artificial neurons and the perceptron. There was no learning mechanism. They just wanted to perceive "things". Both authors refer to Turing and Church.

This work influenced the cyberneticists, whose leading figures were Norbert Wiener10 (1894-1964) and John von Neumann. They used an old Greek word to designate a new science. The relationship between Pitts and Weiner began as that of a teacher and a student before becoming tumultuous. It was for this reason that Pitts sank into alcohol and burned all his thesis documents11.

DEFINITION.- Cybernetics is the unification of the nascent fields of automation, electronics and mathematical information theory, as an entire theory of control and communication, both in animals and in machines.

The term cybernetics, from the Greek ??ße???t?? (kubernêtês) meaning pilot or governor, was proposed in 1947 by American mathematician Wiener to promote this new discipline. The work of Wiener and von Neumann12 (1903-1957) was published in English and French. Von Neumann is considered the father of computer architecture and was also an economist, mathematician and physicist. In von Neumann's 1945 report on building an intelligent machine, the only article cited is by McCulloch and Pitts. Today, the word cybernetics is considered obsolete. It was considered "satanic" by the Russians during the Stalin era, before Soviet researchers took up the discipline in the 1950s. Today, this root has resurfaced in other words such as cybersecurity, cyber spy or cyber physical system.

1.2.3. The arrival of computers

The Electronic Discrete Variable Automatic Computer (EDVAC) in 1949, proposed by von Neumann, was one of the very first electronic computers to work only in binary, while the Electronic Numerical Integrator and Computer (ENIAC), based on the German Z3, was less intelligent, but calculated with decimals.

The EDVAC consists of 6,000 vacuum tubes and 12,000 diodes, consumes 56 kW, occupies a surface area of 45.5 m² and weighs 7,850 kg. The ENIAC is a machine made up of 17,468 vacuum tubes, 7,200 crystal diodes, 1,500 relays, 70,000 resistors, 10,000 capacitors and around 5 million hand-welded connections. It weighs 30 tons and has a surface area of 167 m². Its power consumption is 150 kW.

The strength of these early systems was their speed of calculation, which is shown in Table 1.1 for a multiplication of two 10-digit numbers. These first calculators were built for military purposes, for instance, to calculate a ballistic trajectory. Physicians and pharmacists saw a use for these machines, notably in drug galenics and drug dosage calculations.

Table 1.1. Evolution of calculation speeds

Turing (1912-1954) is the father of connectivism. His report Intelligent Machinery dates from 194813. His aim was to organize a disorganized machine.

DEFINITION.- Connectivism is Turing's idea that knowledge is distributed through a network of connections and, therefore, that learning consists of the ability to build and navigate these networks, and that this can be implemented in a machine.

In the 2000s, educationalists resurrected connectivism, which was subsequently reused by artificial intelligence. For pedagogues, connectivism is a theory of learning in the digital age, a response to the limitations of behaviorism, cognitivism and constructivism. It sought to explain the effects that technology has on the way we live, communicate and learn. This more contemporary approach was developed by George Siemens and Stephen Downes14. There is a kind of reflexivity between the way machines learn and the way educators want us to learn with machines.

Learning was particularly highlighted by the work of Frank Rosenblatt (1928-1971) and the creation of the real machine capable of recognition: the perceptron15. The approach was that of a probabilistic model. The MARK1 Perceptron was built on the principle of multi-layer linear threshold response arrays. It comprised a grid of 400 photocells. This matrix simulated a retina, and was used to discover objects presented to it. Rosenblatt's algorithm is quite simple. The first step is to create a binary supervised classification of a variable as a function of a predictor. The second step is a stochastic search for a linear operator to act as a separator. Each time a new pair, individual and observation, is misclassified, the model is modified, which corresponds to the evolution of a frontier. Several runs are made on the database until convergence is reached.

This approach poses several problems:

• It is possible not to have convergence if the data are not separable.
• If the...