Researchers use dynamic systems and machine learning to give AI spontaneity.
Autonomous functions of robots, such as spontaneity, are highly sought after. Many of the control mechanisms of autonomous robots are inspired by the functions of animals, including humans. Robotists often design robot behaviors using predefined modules and control methodologies, which makes them task-specific, limiting flexibility. Researchers offer an alternative method based on machine learning to design spontaneous behaviors using complex temporal patterns, such as neural activities in the brain of animals. They hope to see their design implemented on robotic platforms to improve their autonomous capabilities.
Robots and their control software can be classified as a dynamic system, a mathematical model that describes the constantly changing internal states of something. There is a class of dynamic system called high-dimensional chaos that has attracted many researchers because it is a powerful way to model animal brains. However, it is generally difficult to achieve control of high-dimensional chaos due to the complexity of the system parameters and its sensitivity to different initial conditions, a phenomenon known by the term “butterfly effect”. Researchers at the University of Tokyo Laboratory of Intelligent Systems and Intelligence and the Next-Generation Artificial Intelligence Research Center are exploring new ways to exploit the dynamics of high-dimensional chaos to establish human-like cognitive functions.
“There is a high-dimensional chaos trajectory called chaotic walking (CI) that can explain brain activity when remembering and associating memory,” said PhD student Katsuma Inoue. “In robotics, CI has been a key tool in establishing its own behavioral patterns. In this study, we propose a recipe for implementing CI in a simple and systematic way, using only intricate time series models created by high-dimensional chaos. We believed that our approach could lead to more robust and adaptable applications when it comes to designing cognitive architectures. It allows us to design spontaneous behaviors without an explicit predefined structure in the controller, which would otherwise serve as an impediment. ”
Reservoir computing (RC) is a technique for automatic learning based on the theory of dynamic systems and providing the basis for a group approach. RC is used to control the type of neural network called the repetitive neural network (RNN). Unlike other machine learning approaches that tune all the neural connections in a neural network, RC only adjusts a few parameters while keeping all other connections in an RNN fixed, which allows the system to train faster. When the researchers applied the principles of RC in a chaotic RNN, they showed the spontaneous patterns of behavior they expected. For a long time now, it has been a difficult task in the field of challenge and robotics and artificial intelligence. In addition, network training is carried out before and in a short time.
“Animal brains provide high-dimensional chaos in their activities, but it is impossible to explain how and why they use chaos. The proposed model may provide insight into how chaos helps our brain process information,” said associate professor Kohei Nakajima. “Furthermore, our recipe would have a wider impact outside the realm of neuroscience, as it can also be applied to other chaotic systems. For example, next-generation neuromorphic devices inspired by biological neurons can have high-dimensional chaos and would be great candidates to implement our recipe. I hope to see them before we pass. “
Reference: Katsuma Inoue, Kohei Nakajima and Yasuo Kuniyoshik, 11 November 2020, “Designing Spontaneous Behavior Change Through Chaotic Walking”. Advances in Science.
DOI: 10.1126 / sciadv.abb3989
Funding: This work was based on the results of a project commissioned by the New Energy and Industrial Technology Development Organization (NEDO). It was sponsored by KI JSPS KAKENHI (scholarship number JP20J12815). KN JSPS has had the support of KAKENHI (grant JP18H05472) and MEXT Quantum Leap Flagship Program (MEXT Q-LEAP) (grant JPMXS0120319794). This work was supported by NEDO [serial numbers 15101156-0 (dated 24 June 2016) and 18101806-0 (dated 5 September 2018)] and a chair at Frontier AI Education, Information Science Science and Technology, and the Next Generation AI Research Center [serial number not applicable (dated 1 June 2016)].