Update “Latest News in Artificial Intelligence”

Engineer and specialist deep learningRodolphe Bride publishes in Flammarion editions Latest news from artificial intelligence. As the title doesn’t indicate, the author devoted himself to a brief historical, but also technical, connotation of AIs and then went beyond that by presenting its limitations.

Garry Kasparov and Lee Sedol experienced this: Over several decades of dazzling technological advancement, we’ve taught AI to learn.. In the mid-1970s, Doctor Gordon Earle Moore argued that the number of transistors in a microprocessor chip would now double every two years. What is now billed as “Moore’s Law” followed a series of theoretical, conceptual, scientific and computer innovations that pushed machines to beat the world’s best chess (with Deep Blue) and players to play (with AlphaGo). Turing test and produce deep fakei.e. malicious, streaming multimedia gimmick.

her Latest news from artificial intelligence, Rodolphe Bride first returns to the long history of artificial intelligence in a few dozen pages. Aristotle (synthesis, consistency, non-contradiction, excluded middle), British mathematician George Boole’s binary logic (preconfigures bit and computer systems based on 0 and 1), Claude Shannon converts analog signals to digital signals (this is via their Boolean algebra Automatic processing by electronics), Alan Turing’s invention of the software principle, and even the sectarian and theoretical contributions of John McCarthy and Marvin Minsky have contributed to shaping the artificial intelligences we know today and the never-ending cinema of cinema. stage in Terminator, Blade Runner, 2001, A Space Odyssey or more recently, HimIt is also discussed in the book.

After summarizing this formation, the author delves deeper into the typology of AIs and their founding principles. Concerned about knowledge and pedagogy, he reminds us that classical computer science is based on mathematics, while artificial intelligence is more based on experiences and concepts that interact with each other. And, in turn, to cite specialist software (which helps to make decisions according to pre-coded diagrams), neural networks (more complex, in layers and according to the principle of self-learning). machine learning) or even productive AIs ( deep fakes eg based on an antagonistic generator neural network, two opposing machines learning from each other). This connotation is not only used to describe different types of AI, because it also makes it possible to objectify their boundaries and underlying questions.

This is especially true for the explainability of artificial intelligence. If expert software acts on pre-coded data, a self-learning machine can make anomalous decisions before the expert can explain them. Rodolphe Gelin recalls some memorable failings, including the heavy crash of an autonomous vehicle (because he had never experienced the presence of a truck lying on the road) or the classification of a dog as a wolf (because the snowy background is associated with the snow-covered background). during AI reinforcement learning). The author hammers it home: our artificial intelligences generalize their knowledge from their experience. However, this can create all sorts of biases (specification issues, learning bias, discriminatory bias, etc.). It also touches on ethical issues, the incompetent proletariat of artificial intelligence (the famous mechanical Turks), or the inferiority of electronic circuits over human neurons in terms of energy expenditure.

Concise and accessible, but imprecise, Latest news from artificial intelligence It makes it possible to stockpile the current techniques, uses and brakes of AI.

Latest news from artificial intelligenceRodolphe Bride
Flammarion, February 2022, 176 pages

See also: