Jürgen Schmidhuber (1963, Munich) became interested in the artificial intelligence when I was a teenager He dreamed of building a robot that was more intelligent than himself. Some 40 years later he chairs the NNAISENSE company -That designs robots that learn as children do- and pursues the same dream. Artificial intelligence is advancing by leaps and bounds, and with it, the possibility of fulfilling its desire: "In two decades there will be more intelligent robots than humans".
"My intention is to build an automaton that learns to be smarter than me and solve problems that I can not solve. When I get it I will retire, "says Schmidhuber with a smile, who is also co-director of the Dalle Molle Artificial Intelligence Research Institute in Manno (Switzerland). In 1997, this German was the architect of an algorithm called Long Short Term Memory (LSTM), which has revolutionized the processing of natural language by machines and is used by companies such as Google, Facebook or Amazon.
Now his team at NNAISENSE tries to manufacture robots that are autonomous and learn as children do: "Most of the time they do not try to imitate their parents, but they learn by themselves. They see processes, interpret them and reproduce them. " The youngest ones, explains Schmidhuber, invent "their own experiments" to learn how the world works. "For example, when they release their toys they see how they fall to the ground faster and learn more about gravity," he says. Minors also make predictions about what actions will cause pain or satisfaction: "If a baby is given a table, he will feel pain, and next time he will try to avoid the blow by surrounding the object or without touching it."
The robots manufactured by this expert "learn as children to do experiments that teach them how the world works". At the beginning, they are given a series of orders but then, by means of systems of machine learning, they themselves learn new skills on their own, according to this German participant in the IROS 2018 -The largest robotics congress in the world-.
Schmidhuber gives the example of a robot that learns to run "without any teacher to teach" in the same way that a child does: "A baby may need a year to control their muscles, learn to get up, walk and run. Neural networks of this robot also need weeks to do it. " While this humanoid at first fell to the ground continuously, the expert in artificial intelligence says that "came a point where he was able to move one foot after another and even to run."
But this ability to learn as children is not only important for humanoid robots. In another project, Schmidhuber and his team worked with small Audi cars that reached 120 kilometers per hour to learn on their own to park: "They were equipped with cameras and sensors. They collided with other vehicles and had to try again and again, little by little they avoided hitting themselves and finally they could park in difficult situations ".
This expert in artificial intelligence explains that the key is to endow robots with "the curiosity of a baby that tries to make experiments and understand how the world works and the consequences of its actions". This way of learning, he argues, does not differ much from that of adults: "A child is a little scientist who experiments with his toys. Maybe 20 years later that child is for example a scientist who does experiments in the European particle physics laboratory CERN"
Although automata of the future can learn in a similar way to people, "they will be very different from the humanoid robots that are manufactured today." "The robots of the future, for example those that go to the Moon or to Mercury, will not physically resemble us, they will be composed of systems and sensors that humans do not have and they will be much smarter", concludes Schmidhuber.