Back on the relationship between humans and robots, on the occasion of the first edition of InnoRobo, lounge robotics, and conferences Robolift associated with it. "A robot can learn as a child?" asks Pierre-Yves Oudeyer, researcher at INRIA, and Laboratory Director Flowers, a research laboratory specializing in social robotics (presentation.
pdf). In other words, artificial intelligence being acquired conditions can it? "We imagine a future where robots would be in our homes to help: set the table, placing the dishes, motivate the elders tinker ... It involves a lot of intelligence and knowledge (including emotional). The robots must be able perform elaborate reasoning "...
Put the dishes in the dishwasher application example of mobilizing a lot of manual skills and intellectual development. It is clear that we live with robots will require intelligence. But which one? "Although the machines know to beat the best chess champions. Deep Blue can beat Kasparov, because we put all the knowledge in terms of failure in its program, but Deep Blue can not explain what are the failures.
"Asimo, to mention that it is a great robot which can be programmed to perform sophisticated behaviors like playing football, dancing, chatting or playing the bartender ... "We model these basic tasks with mathematical equations that generate the appropriate behavior. But when the robot must accomplish a new task, you have to hire a new engineering team! "Quipped the researcher.
Each new task requires a new development ... And the robots do not know that what he has learned. This robotic model is certainly useful, but it quickly reaches its limits. "Robots should be able to learn because you can not imagine in advance what they need. They must be able to catch or catch something they should be able to adapt to the behaviors of people, their habits ...
All News Tech Buzz does not make the table the same way! "They need to learn and learning to learn. "But for humans, in real life, learning is also not easy." In humans we learn through observation, trial and error, experience ... Learn in Tech News Buzz physical Buzz Real Tech News. And this learning takes time.
But there are a large number of objects and people interact and where there are endless activities and skills that can be learned. What should learn and do we not learn? We must learn from human to teach machines to learn. In children, learning is not born from nothing. Nor that humans are "universal learners.
"If children learn the language quickly and easily, they are less talented than machinery for the numbers." Children are provided with tools and constraints: they explore the synergy and muscle coordination complex rather than every muscle movement possible. What is innate, self-organizing constraints and biases is essential in learning.
Some movements are innate reflexes such as grasp a finger. We all recognize faces and moods. Finally, children born into social environments that help them set and learn about their environment ... "How can we translate the constraints of child development to the development of robots, how the innate social abilities can they afford to explore the society and learning? " How robots can they learn to speak by example? Pierre-Yves Oudeyer then refers to the research he was conducting with Kaplan on the Sony Aibo and shows us, too, how he tried to teach him new words by pointing to objects and naming them.
But as Fred Kaplan, he soon realized that in fact the robot does not see the same thing, "he was not paying attention to what I thought." "The solution for them to learn new things is to equip them with mechanisms of attention, as children naturally, to verify that they follow the same interaction that we: you point the finger, his gaze is directed in direction ...
We must imitate the natural human mechanisms to verify the attention of the robot. " But is it enough to imitate? Pierre-Yves Oudeyer shows us another video where a robot is controlled by a human who sees the cameras and the robot must meet the demands of a human to identify an object. And even with human intelligence, it is not so simple! Often, the gesture meant to draw attention is accomplished outside the visual field of the robot.
Often it is difficult to understand what object was even designated with a pointing finger ... "Even if it is" humanoid robot's sensory apparatus is very different from humans. Humans do not know what the robot is able to see. It is far from clear to implement technology to enable the machine to understand that from a visual image is meant something special ...
"You have to find other voices, recommends the researcher. Hence the idea to go look at how we teach, not to men, but animals such as chimpanzees. Often, this involves learning objects mediation, allowing for interaction with the chimp so he can learn to identify a number of objects. Can we imagine using mediation interfaces with robots? That's what we tried Oudeyer Pierre-Yves Pierre Rouanet and Fabien Danieau in a recent study (.
pdf) (video) using an iPhone (or wiimote) to refer specifically to the robot the object to which it is desired interest. What can give instructions more easily and quickly. An interface more efficient than algorithms automated learning, says its designer. Better able to identify objects in the environment and enable the robot to better imitate us is important, but children not only learn by imitating.
Much of their learning born of spontaneous explorations related to curiosity. Pierre-Yves Oudeyer and his team and tried to introduce intrinsic motivation to the robot via the "Playground Experiment", to refer to work at Sony. Here, the robot is expected to learn new tasks without having semantic knowledge on the subject (video).
Curiosity-driven learning of locomotion, the robot then explores the movement, without any knowledge of the body or the environment, but by being programmed to learn specific movements (go forward) and this experiment he finds interesting is to say, explore movements that produce progress in learning, as is the case in this video where Aibo gives the impression of crawling rather than walking.
The idea is that the robot can reuse actions learned by curiosity to reach a particular point. Remains that while the general morphology of the robot has a direct impact on his explorations spontaneous ... How then can it be simplified to improve the control of learning. "Build the movement of a robot is difficult and complicated calculations, hence the idea of building a robot with a morphology simplifying the controls." Drawing on the work of Tad McGeer on passive dynamic built a machine in the 90's can generate a simple walking pattern, without electronics, a more natural motion than many robots andromorphs.
The idea here is that physics replaces the morphological computation. The Acroban Humanoid Robot Project is a flexible, highly resistant to external disturbances, because it balances out very efficiently thanks to the physics and geometry (video). "You can take the robot's hand so that you follow, without there being a line of code instructing him to follow you.
His progress is steady. Here, intelligence is self-organized by the physical paths that produces spontaneous interaction with complex human behavior. "Will we SIMULATING INTELLIGENCE? Jean-Claude Heudin (blog), Director of Research Laboratory International Institute of Multimedia University Centre Leonardo da Vinci is the author of numerous books and articles in the field of artificial intelligence and complexity science, including the latest entitled Robots and Avatars - Pygmalion's Dream.
"The history of the design of robots is that of a curse," Attack the researcher (presentation. pdf), which refers to the Bible and the prohibition of making images of a god or as men (Exodus XX, 4). This has not prevented man to try, but it certainly explains why the first impression the public on these machines is that of anxiety and fear.
Since the Eniac, the first conference of the Dartmouth Summer Research Project (which gave birth to artificial intelligence as a research discipline) and the explorations of Alan Turing on machine intelligence in the 50's where Turing, the computer has increased steadily. It was not until 1997 and the victory of Deep Blue over Kasparov in that it begins to beat the man.
Many still think that in future, the machine will turn against man. Ray Kurzweil, Singularity when he speaks of is not something else, as it provides an artificial intelligence will override the power of all human brains. "I do not share this vision", recognizes the researcher does not mean he does not believe in the curse.
The grail of artificial intelligence that we are still far surpass, in part because the intelligence of machines is radically different from human intelligence. One is digital when the other is organic. One is built (that is to say we built systems to suit the functions) when the other is evolutionary.
One makes sense when the other is emotional. One is computational (working from a succession of operations) while the other works on the inference, that is to say that it works in drawing conclusions. One is symbolic when the other is based on the senses ... The reasons why we fail to make the machine intelligences are several explanations.
Ray Kurzweil is the computing power is insufficient. The solution is simple: increase it. But this is not true, retorted Jean-Claude Heudin. "We are not facing a problem of computing capacity. There is something that escapes us in our understanding of intelligence ... And for my part, I think especially since we do not reach a sufficient level of complexity.
"The internet is like both a set of neurons as a simulation of the universe at large scale. We tend to approach the problem of artificial intelligence by a functional approach. "We try to understand the brain by cutting cell by cell, function by function. The problem with this approach is that it allows a cell to understand, but do not necessarily rise to the level of thought, conscience.
The reductionist approach is bearing fruit, but also its limitations. We lose the understanding of global properties of the whole. "Hence the idea for Jean-Claude Heunin to rely instead on a synthetic approach, a reverse process where the link is established between the agents, where they are made interact to see the properties that emerge from these interactions.
We must, he says, consider intelligence as something that brings out the cooperation of a number of agents. Jean-Claude Heudin shows what is called cellular automata, grids of cells that react according to the state of neighboring cells to create specific models, such as those based on the game of life created by the mathematician John Horton Conway, to observe rules and statements of various behaviors of force generation state.
We distinguish three types: stable states, uniform, periodic behavior or cyclical and chaotic behavior of course. But there are other strange behaviors, those which raise protocells. "This happens between order and chaos, in a thin line between the two." Such automata computational properties allowing to build a computer inside.
If there is more complex systems, as proposed Life Drop, a system that simulates the development of a microscopic ecosystem (including a new version is under development - video). With this game, we see the emergence of adaptive behaviors, with standard agents with the role of prey and other predators and evolving as a natural environment ...
Eva is a software developed by Jean-Claude Heudin, agent conversational (those famous chatterbots we have often mentioned) that there are several versions (Alicia, Hal 9000 ...). The next version should be called "Minna House Doctor" and is an artificial intelligence with a staff of thirty people who interact with each other for answers.
Minna is able to get information on the Internet and use them in the flow of conversation. Jean-Claude Heudin and his teams are already working on a third generation chatterbot that is able to learn and a fourth who should be able to modify the structure of different forms of intelligence ... 
pdf). In other words, artificial intelligence being acquired conditions can it? "We imagine a future where robots would be in our homes to help: set the table, placing the dishes, motivate the elders tinker ... It involves a lot of intelligence and knowledge (including emotional). The robots must be able perform elaborate reasoning "...
Put the dishes in the dishwasher application example of mobilizing a lot of manual skills and intellectual development. It is clear that we live with robots will require intelligence. But which one? "Although the machines know to beat the best chess champions. Deep Blue can beat Kasparov, because we put all the knowledge in terms of failure in its program, but Deep Blue can not explain what are the failures.
"Asimo, to mention that it is a great robot which can be programmed to perform sophisticated behaviors like playing football, dancing, chatting or playing the bartender ... "We model these basic tasks with mathematical equations that generate the appropriate behavior. But when the robot must accomplish a new task, you have to hire a new engineering team! "Quipped the researcher.
Each new task requires a new development ... And the robots do not know that what he has learned. This robotic model is certainly useful, but it quickly reaches its limits. "Robots should be able to learn because you can not imagine in advance what they need. They must be able to catch or catch something they should be able to adapt to the behaviors of people, their habits ...
All News Tech Buzz does not make the table the same way! "They need to learn and learning to learn. "But for humans, in real life, learning is also not easy." In humans we learn through observation, trial and error, experience ... Learn in Tech News Buzz physical Buzz Real Tech News. And this learning takes time.
But there are a large number of objects and people interact and where there are endless activities and skills that can be learned. What should learn and do we not learn? We must learn from human to teach machines to learn. In children, learning is not born from nothing. Nor that humans are "universal learners.
"If children learn the language quickly and easily, they are less talented than machinery for the numbers." Children are provided with tools and constraints: they explore the synergy and muscle coordination complex rather than every muscle movement possible. What is innate, self-organizing constraints and biases is essential in learning.
Some movements are innate reflexes such as grasp a finger. We all recognize faces and moods. Finally, children born into social environments that help them set and learn about their environment ... "How can we translate the constraints of child development to the development of robots, how the innate social abilities can they afford to explore the society and learning? " How robots can they learn to speak by example? Pierre-Yves Oudeyer then refers to the research he was conducting with Kaplan on the Sony Aibo and shows us, too, how he tried to teach him new words by pointing to objects and naming them.
But as Fred Kaplan, he soon realized that in fact the robot does not see the same thing, "he was not paying attention to what I thought." "The solution for them to learn new things is to equip them with mechanisms of attention, as children naturally, to verify that they follow the same interaction that we: you point the finger, his gaze is directed in direction ...
We must imitate the natural human mechanisms to verify the attention of the robot. " But is it enough to imitate? Pierre-Yves Oudeyer shows us another video where a robot is controlled by a human who sees the cameras and the robot must meet the demands of a human to identify an object. And even with human intelligence, it is not so simple! Often, the gesture meant to draw attention is accomplished outside the visual field of the robot.
Often it is difficult to understand what object was even designated with a pointing finger ... "Even if it is" humanoid robot's sensory apparatus is very different from humans. Humans do not know what the robot is able to see. It is far from clear to implement technology to enable the machine to understand that from a visual image is meant something special ...
"You have to find other voices, recommends the researcher. Hence the idea to go look at how we teach, not to men, but animals such as chimpanzees. Often, this involves learning objects mediation, allowing for interaction with the chimp so he can learn to identify a number of objects. Can we imagine using mediation interfaces with robots? That's what we tried Oudeyer Pierre-Yves Pierre Rouanet and Fabien Danieau in a recent study (.
pdf) (video) using an iPhone (or wiimote) to refer specifically to the robot the object to which it is desired interest. What can give instructions more easily and quickly. An interface more efficient than algorithms automated learning, says its designer. Better able to identify objects in the environment and enable the robot to better imitate us is important, but children not only learn by imitating.
Much of their learning born of spontaneous explorations related to curiosity. Pierre-Yves Oudeyer and his team and tried to introduce intrinsic motivation to the robot via the "Playground Experiment", to refer to work at Sony. Here, the robot is expected to learn new tasks without having semantic knowledge on the subject (video).
Curiosity-driven learning of locomotion, the robot then explores the movement, without any knowledge of the body or the environment, but by being programmed to learn specific movements (go forward) and this experiment he finds interesting is to say, explore movements that produce progress in learning, as is the case in this video where Aibo gives the impression of crawling rather than walking.
The idea is that the robot can reuse actions learned by curiosity to reach a particular point. Remains that while the general morphology of the robot has a direct impact on his explorations spontaneous ... How then can it be simplified to improve the control of learning. "Build the movement of a robot is difficult and complicated calculations, hence the idea of building a robot with a morphology simplifying the controls." Drawing on the work of Tad McGeer on passive dynamic built a machine in the 90's can generate a simple walking pattern, without electronics, a more natural motion than many robots andromorphs.
The idea here is that physics replaces the morphological computation. The Acroban Humanoid Robot Project is a flexible, highly resistant to external disturbances, because it balances out very efficiently thanks to the physics and geometry (video). "You can take the robot's hand so that you follow, without there being a line of code instructing him to follow you.
His progress is steady. Here, intelligence is self-organized by the physical paths that produces spontaneous interaction with complex human behavior. "Will we SIMULATING INTELLIGENCE? Jean-Claude Heudin (blog), Director of Research Laboratory International Institute of Multimedia University Centre Leonardo da Vinci is the author of numerous books and articles in the field of artificial intelligence and complexity science, including the latest entitled Robots and Avatars - Pygmalion's Dream.
"The history of the design of robots is that of a curse," Attack the researcher (presentation. pdf), which refers to the Bible and the prohibition of making images of a god or as men (Exodus XX, 4). This has not prevented man to try, but it certainly explains why the first impression the public on these machines is that of anxiety and fear.
Since the Eniac, the first conference of the Dartmouth Summer Research Project (which gave birth to artificial intelligence as a research discipline) and the explorations of Alan Turing on machine intelligence in the 50's where Turing, the computer has increased steadily. It was not until 1997 and the victory of Deep Blue over Kasparov in that it begins to beat the man.
Many still think that in future, the machine will turn against man. Ray Kurzweil, Singularity when he speaks of is not something else, as it provides an artificial intelligence will override the power of all human brains. "I do not share this vision", recognizes the researcher does not mean he does not believe in the curse.
The grail of artificial intelligence that we are still far surpass, in part because the intelligence of machines is radically different from human intelligence. One is digital when the other is organic. One is built (that is to say we built systems to suit the functions) when the other is evolutionary.
One makes sense when the other is emotional. One is computational (working from a succession of operations) while the other works on the inference, that is to say that it works in drawing conclusions. One is symbolic when the other is based on the senses ... The reasons why we fail to make the machine intelligences are several explanations.
Ray Kurzweil is the computing power is insufficient. The solution is simple: increase it. But this is not true, retorted Jean-Claude Heudin. "We are not facing a problem of computing capacity. There is something that escapes us in our understanding of intelligence ... And for my part, I think especially since we do not reach a sufficient level of complexity.
"The internet is like both a set of neurons as a simulation of the universe at large scale. We tend to approach the problem of artificial intelligence by a functional approach. "We try to understand the brain by cutting cell by cell, function by function. The problem with this approach is that it allows a cell to understand, but do not necessarily rise to the level of thought, conscience.
The reductionist approach is bearing fruit, but also its limitations. We lose the understanding of global properties of the whole. "Hence the idea for Jean-Claude Heunin to rely instead on a synthetic approach, a reverse process where the link is established between the agents, where they are made interact to see the properties that emerge from these interactions.
We must, he says, consider intelligence as something that brings out the cooperation of a number of agents. Jean-Claude Heudin shows what is called cellular automata, grids of cells that react according to the state of neighboring cells to create specific models, such as those based on the game of life created by the mathematician John Horton Conway, to observe rules and statements of various behaviors of force generation state.
We distinguish three types: stable states, uniform, periodic behavior or cyclical and chaotic behavior of course. But there are other strange behaviors, those which raise protocells. "This happens between order and chaos, in a thin line between the two." Such automata computational properties allowing to build a computer inside.
If there is more complex systems, as proposed Life Drop, a system that simulates the development of a microscopic ecosystem (including a new version is under development - video). With this game, we see the emergence of adaptive behaviors, with standard agents with the role of prey and other predators and evolving as a natural environment ...
Eva is a software developed by Jean-Claude Heudin, agent conversational (those famous chatterbots we have often mentioned) that there are several versions (Alicia, Hal 9000 ...). The next version should be called "Minna House Doctor" and is an artificial intelligence with a staff of thirty people who interact with each other for answers.
Minna is able to get information on the Internet and use them in the flow of conversation. Jean-Claude Heudin and his teams are already working on a third generation chatterbot that is able to learn and a fourth who should be able to modify the structure of different forms of intelligence ...
No comments:
Post a Comment