Artificial intelligence: fantasies and realities


Anthony Rousseau

Head of R&D, Partner




back to the summary

Artificial intelligence has always nourished the wildest fantasies. From "2001, the Odyssey of Space" to "Terminator", we are predicted of futures all as different as each other, but with the same constant: the machines will be endowed with intelligence. But where are we today? Are machines truly capable of intelligence? How can this change our jobs?


Low IA and strong AI


Even if the term "artificial intelligence" is extremely fashionable at the moment, we are still very far from the fantasies it arouses in our minds. First of all it is necessary to distinguish two ways of "giving intelligence" to a machine.

Historically, the first approach has been to teach human behavior through a modeling of the world and the rules that govern it. In this approach, sometimes found under the name "weak AI", it is the human who defines and programs the knowledge in the machine. He endeavors to put in black and white what he knows and the way in which he reasons on it. This approach has occupied the researchers for a few decades, but with a major problem: there are things we are not able to explain. Try to explain by "a + b" to a human how to fit on a bike. Not easy? So imagine a machine ...


Another approach was then developed. Instead of being based on "rules" taught by a human machine, it is based on a totally inverse postulate: the machine will be asked to observe the "world" and draw conclusions from it itself . This approach is entirely based on statistical mathematics. Depending on a certain number of input data presented to the machine (these represent the "world" that it knows), it will be asked to deduce from the probabilities: what is the probability that such an event will take place of any other observed or observable events? So that the more events it sees, the more it will be able to deduce others from them effectively.

For example, if a machine observes that a customer arriving on the page of a product X by a specific Facebook advertisement buys the same product and repeatedly observes it (different customers, different contexts, etc.) it will be able to "learn" that: such campaign Facebook + product page X => product purchase X.



The importance of data


Knowing this, it becomes essential to have as much data as possible to learn these systems. Raw data collected by the web, marketing signals, calls from call centers represent sets, the "world" from which artificial intelligence can feed. This is the new Eldorado of companies that want to get the job done.

A well-designed and well-trained AI will be able to model and "understand" the environment of a company and thus exploit all the hidden and untractable data from a human to derive potentially phenomenal competitive advantages.


Supervised and unsupervised learning


Another tremor has shaken for several months already the growing world of artificial intelligence. Indeed, at present, most of the learning is done in a "supervised" way, that is to say that one presents to the machine data to be treated as input, while indicating for each entry the "answer" expected output. Concretely, this approach allows the AI ​​to create relations between input and output (these mathematical links are also called "neural networks"), but they remain human-oriented. For it is the human who defines and associates upstream given input and expected output, then instructs the machine to deduce invisible correlations, in order to be able subsequently to predict unpublished events.

However, producing large amounts of "annotated" data is a long, tedious and expensive job. It was therefore quickly interesting to be able to begin to free oneself from these steps of preparing the data. For this, new approaches, called "unsupervised", appear. The principle is simple: it is always a matter of presenting data to the machine as input, but this time in a raw, unstructured way. With the help of different algorithms, it will be asked, in addition to learning the "world" that is placed at its disposal, to structure the data itself and to deduce relationships and sets of completely automatic. In doing so, in addition to costly human intervention, this approach represents a first step towards the conceptualization and elaboration of "real" artificial intelligences, which will nevertheless remain so-called "specialized" intelligences.


An intelligent machine?


But for all that, can we say that machines are intelligent? They are currently "simply" able to reproduce what they observe and to generalize concepts in order to specialize in a given task. Despite the alarmist articles that regularly flourish in the press, we are still far from the concept of a "general" artificial intelligence, which would be able to compete with human beings and even humanity as a whole. On the other hand, they can now be used very effectively in the treatment of masses of data that can not be exploited by humans. These masses of data, which are increasingly imposing, represent today an untapped potential for many companies.



back to the summary

GET your pass today!


register now!

AI Europe 2017  by

AI EUROPE 2017 is organized by Corp Events, a corporate event management agency that specializes in the organization of first-class trade shows in the field of new technologies