Reading time: 15 minutes
Artificial intelligence is no longer science fiction, it is very real and very present among us today. As research progresses, its definition is becoming clearer.
A moribund definition
Artificial Intelligence (AI) is defined by " l 'set of theories and techniques implemented to create machines capable of simulating intelligence ". However, this definition is not unanimous and gives AI the nickname " grand myth of our temps ", explains Wikipedia.
The problem comes from the term " intelligence " which is not clearly defined. Let's take mathematical calculation as an example. Do you need intelligence to perform a mathematical calculation? Although algorithms are known and mechanizable today, they do not take away the complexity that has put off generations of mathematicians and even today. If you are asked the square root of 10, very few of us will be able to find the answer with just a pencil and paper. And yet calculation is excluded today from the realm of AI, because the algorithms for resolution are well known and established.
Will we always be able to solve everything with an algorithm? Is intelligence all algorithmic? Possibly yes... However, the complexity of the problems to be solved can easily exceed our capacity as humans to design them. Advances in neural networks currently allow AI to learn and create its own algorithms by trial and error, just as a human would, but without the constraints of time and power.
The game, a favorite field for AI!
Deep Blue
" Le May 11, 1997, Deep Blue beat Garry Kasparov, Russian chess player crowned best player in the world in 1984 and remained number one in the world for more than twenty years. 1.8 meters, 1.4 tons and 256 powerful microprocessors were needed to achieve this historic masterstroke. But then, who is this powerful new competitor to Garry Kasparov? Developed by IBM, Deep Blue is a computer specialized in the game of chess.
Some kind of artificial intelligence before its time? Or almost? Deep Blue is an AI "under infusion" of human intelligence. In its memory (as in that of a great master) are recorded the evaluations of many positions, all "rated" by the consultants of the great masters. Deep Blue performs advanced calculations only for the positions selected by humans. Without humans, this computer could never have beaten Garry Kasparov alone. »
AlphaZero
AlphaZero: one of the first AI programmed to learn chess in 2017
" 20 years later, in 2017, AlphaZero, instead of learning from a database of games played by humans, this version of AI software was simply programmed with the rules of the game and algorithms. From there, self-taught, the AI simulated games and mastered the game.
In just 4 hours, AlphaZero has assimilated all known knowledge about the game and has become a true master. Of the 100 games played against Stockfish, the best chess program ever developed, the AI won 25 by playing with white, 3 with black, while the rest of the games ended in draws. A real feat! ", according to an article by Fabernovel.
AlphaZero has created better algorithms. In only 4 hours, this software became world chess champion. In this case, we can certainly speak of artificial algorithms, created from scratch from its experimentation. This AI does what no human can do.
The field of games is one where learning is relatively simple. To play chess, you need to know some basic rules that even a 5 year old child can learn. An AI can easily assimilate these rules and simulate a few billion fictitious games in a few hours to find the best ways to win. However, this is not the case for human activities. Life is not a finite activity like a game of chess. In activities where there is no winner or the game never ends, AI can hardly determine whether it is a winning or a losing algorithm. Its learning curve is much slower...
The short history of artificial intelligence
Here are some memorable dates in the development of AI around the world:
1950
- Alan Turing wondered if a machine could be intelligent and developed his fameux test to find out.
- John McCarthy used the term artificial intelligence for the first time. The first AI departments appear in universities.
1960
- Computer power was not sufficient at the time, so developments were more theoretical and algorithmic than practical. The algorithms of some games (draughts, chess) were designed.
- One of the first artificial intelligences for Othello is IAGO, developed in 1976 by Caltech University in Pasadena, California, which defeated Japanese champion Fumio Fujita without difficulty.
1980
- Automatic machine learning is developing. The computer starts to deduce from " règles to suivre " just by analyzing data.
- In 1988 Hans Berliner's HiTech computer was the first program to beat a chess grandmaster, Arnold Denker (74) in a match (3.5-1.5). Later strong players were beaten, like grandmaster Bent Larsen (then at 2560 Elo points), who was defeated in 1988 by Deep Thought in a tournament in California.
1990
- In 1997, the victory of Deep Blue (IBM) (nicknamed Deeper Blue in this revenge match) over Garry Kasparov (3.5-2.5) marked a turning point. For the first time the best human chess player was beaten in a match (and not in a single game) by a machine.
2000
- Web 2.0, big data and new computing power and infrastructure are enabling some computers to explore unprecedented amounts of data. This is deep learning.
- Arrival of the personal assistants " intelligents", such as the Apple Siri in 2007, Google Now in 2012 and Microsoft Cortana in 2014.
2010
- Watson, the IBM-designed supercomputer, wins two of the three rounds of the game show Jeopardy, a general knowledge game show, beating his two human competitors by a wide margin in cumulative wins.
- In 2015, AI will make significant progress in the practice of go, which is more complex to apprehend than failures (among other things because of the greater number of positions. 10 170 to go, against 1 050 for chess and plausible games: 10 600 to go, against 10 120 for chess.
- In October 2015, AlphaGo, an AI software designed by Google subsidiary DeepMind, beats for the first time Fan Hui, the three-time European go champion and thus takes up what was considered one of the greatest challenges for artificial intelligence.
- In 2017, during the poker tournament " Cerveau vs IA" organized in a Pennsylvania casino, the IA Libratus, developed by researchers at Carnegie-Mellon University in Pittsburgh, is confronted with human opponents. The human players opposed to Libratus, all poker professionals, successively face the machine in a face-to-face game. After more than 120 000 hands have been played, Libratus accumulates 1 766 250$ (virtual). The human player who lost the least amount of money in his duel with the machine, Dong Kim, still has a deficit of more than 85 000$. The machine's clean and clean victory marks a new stage in the development of artificial intelligence. It illustrates the progress made in AI's handling of "imperfect information", where thinking must take into account incomplete or hidden data. Estimates of the number of possibilities in a face-to-face no limit poker game are approximately 10 160.
- Deep Mind, a Google subsidiary, is beginning work on artificial general intelligence (AGI) to create a general intelligence similar to human intelligence.
- AlphaZero has been developed to play any game knowing only the rules and learning to play against itself. After 9 hours of training, AlphaZero beat the Stockfish chess program with 28 wins, 72 draws and no defeats .
- Artificial intelligence models developed by Microsoft and Alibaba each beat humans in a Stanford University reading and comprehension test.
- The research institute OpenAI announces that it has created an artificial intelligence program capable of generating texts so realistic that this technology could be dangerous.
2020
- Google's AIT can now detect breast cancer more accurately than doctors can.
- The first patents on inventions made by AI are filed.
2030
- Singularity: AI will reach and surpass human intelligence as a whole.
As with all sciences, discoveries and advances in AI are occurring at an ever-increasing rate. We have reached a period of exponential development and this is leading us into an uncertain future.
The development of artificial intelligence blocked by technology
Algorithms initially developed by artificial intelligence researchers are becoming part of larger systems. AI has solved many very complex problems and its solutions have been used throughout the technology sector. Examples include data mining, industrial robotics, logistics, voice recognition, banking applications, medical diagnostics, pattern recognition, Google search engine, chatbots, and more.
The development of this new science has been continually blocked by technological limitations and this is still the case today. For Ray Kurzweil, futurologist and research director at Google, the problem of AI's advancement lies in the lack of computing power. Based on Moore's law, he predicts that machines with intelligence comparable to humans will arrive around 2030.
Research areas in AI
Strong Artificial Intelligence
The concept of strong artificial intelligence refers to a machine capable of producing significant intelligent behaviour and modeling abstract ideas. But also to experience an impression of real consciousness, "real feelings" (whatever one might put behind those words), and "an understanding of one's own reasoning".
Weak artificial intelligence
The notion of weak artificial intelligence is a pragmatic engineering approach: to try to build more and more autonomous systems (to reduce the cost of their supervision), algorithms able to solve problems of a certain class, etc. The notion of weak artificial intelligence is a pragmatic engineering approach: to try to build more and more autonomous systems (to reduce the cost of their supervision), algorithms able to solve problems of a certain class, etc. The notion of weak artificial intelligence is a pragmatic engineering approach. But, this time, the machine simulates intelligence. It seems to act as if it were intelligent. Concrete examples of this can be seen with conversational programs that try to pass the Turing test, such as ELIZA. These programs manage to roughly mimic the behaviour of humans in a dialogue with other humans.
AI, present and inescapable
Artificial intelligence is already with us. Its now daily advances make it an increasingly present force that is here to stay. More and more, we see the emergence of powerful tools integrating AI to assist doctors, insurers, meteorologists...
In 2019, Yoshua Bengio, professor of computer science at the Université de Montréal and winner of the annual A. M. Turing Award, said that in 2019 AI had the intelligence of a 2-year-old child. Thus, power and computing capacity is the main obstacle to its explosive growth. But Moore's Law predicts that by 2030, power will be the key to AI surpassing human intelligence. This is what some call singularity, because beyond this date no one can predict what will happen.
