Artificial Intelligence, self-taught and soon to be autonomous ?
Let's be clear, Artificial Intelligence is not the reserved domain of Google ! But when it comes to measuring the progress of AI, it is clear that the American mastodon, which invests astronomical sums of money, is always on the front line...
HAL 9000 or Skynet may not be so far from being born... The first is the fruit of the fertile imagination of the British writer Arthur C. Clarke and created by Dr. Chandra. HAL 9000 is a supercomputer with Artificial Intelligence that is so autonomous and independent of the human being that it finally wants to eliminate its teammates in the Odyssey of space saga. The second is also Artificial Intelligence, initially created to automate the American nuclear response but which will eventually produce Terminators to exterminate the human species.
Reality has certainly not yet reached the point of fiction, and today's Artificial Intelligences are still far from claiming their independence or from wanting at all costs the heads of human beings to rule over the planet. But, while waiting for such a catastrophic scenario to emerge, Google nevertheless hit hard in terms of announcements early this fall. Capable of creating a self-taught Artificial Intelligence called AlphaGo Zero for Go game learning, the Silicon Valley company has also published other major advances on its AutoML project, an Artificial Intelligence designed to help engineers code other AIs.
AutoML, an Artificial Intelligence more efficient than its developer fathers
Thus, AutoML is no longer only able to assist but can simply demonstrate a bit of autonomy to create its own Machine Learning software. Launched in May 2017 during the Google I/O conference, AutoML can now create a "child" model of Machine Learning, study its behavior and analyze the results to make corrections and design a new model even more powerful. AutoML repeats this process thousands of times to, in fine, always develop an ever more powerful and efficient AI algorithm, even surpassing those created by its founding fathers and human beings.
As an example and to illustrate such an evolution, AutoML's auto-produced Artificial Intelligence algorithm presents a better balance than that created by human developers in the recognition of multiple objects on an image. At 43% with AutoML's auto-generated system, this rate drops to 39% with the system developed by the engineers and developers working on this project. This is a small difference on a sample of more than 80 million images, but it is a good proof of the advances made in Artificial Intelligence.
AlphaGo Zero, a self-taught Artificial Intelligence
At the same time, AlphaGo Zero has replaced his elder AlphaGo, an developed by Google DeepMind that defeated South Korean Go game champion Lee Sedol in March 2016 ! Where chess is even relegated to the rank of being a major player in terms of complexity, AlphaGo had outperformed the Asian champion thanks to the millions of games that its developers had put at his disposal to train him. A year and a half after his victory, AlphaGo has already retired, leaving the field open to AlphaGo Zero, discovered on October 17th thanks to a publication in the scientific journal Nature.
AlphaGo Zero is a simplified architecture that manages to do without human knowledge to learn how to play the game of Go. Only the rules of the game are enough for him to embark on the adventure, train and improve. and David Silver, his two creators at DeepMind, are just talking about "the strongest go player in history". They necessarily preach a little for their parish, but "Zero" has won 100 games against his ancestor. To train and achieve such a feat, AlphaGo Zero therefore only starts from the rules of the game to launch a first game against himself, study the movements, draw conclusions, correct and relaunch a new game with these new data. In three days, Artificial Intelligence played nearly five million games and underwent intensive training without the slightest human intervention, except for the initial task of giving it the rules of the game.
Other fields of application for Artificial Intelligence ?
After 40 days of reinforcement learning, and only by reinforcement when AlphaGo added supervised learning, AlphaGo Zero became the best player in the world. And all this with a single network of deep neurons when his predecessor, once again, used two of them to first indicate the next movement to play and then predict the winner according to the current positions on the Goban (the apron on which we play go) !
Such progress in such a short space of time on the scale of mankind on Artificial Intelligence makes it possible to envisage new fields of application in fields such as health, energy or communication. On DeepMind's blog, its creators still think so. “These moments of creativity give us the assurance that AI will be a multiplier of human ingenuity, helping us in our mission to solve some of the most important challenges facing humanity", explain the two authors Demis Hassabis and David Silver. For example, can we imagine Artificial Intelligence surpassing human beings in certain research ? Likely from the moment the machine no longer needs human knowledge and data to progress...
If Artificial Intelligence can therefore design a learning machine algorithm better than a human being or do without its services to learn a task, the human being remains at the origin of his immense progress and therefore keeps his hand on his good development and the direction to follow. HAL 9000 and Skynet will therefore wait a little while before conquering our world.