Objetivo
Robots are of increasing importance and will become an integral part of European society. Despite continuing advances, current robots cannot approach the abilities of even the simplest mammals. Achieving continuous and real-time learning without interference between learnt tasks remains a difficult problem. To attain the learning ability and finesse of movement that animals display, information from a large number of sensorimotor and cognitive signals must be integrated. However, few principles of integration have been proposed. Our inter-disciplinary consortium of physicists, neuroscientists and engineers will investigate the principles of integration and representation in the brain. Real-time spiking neural networks, based in part on the cerebellum, a major site of sensorimotor integration and motor learning in the brain, will be developed. Both hardware and software approaches will be employed. The control abilities of these spiking networks will be evaluated with robots in real environments and compared with state-of-the-art robotic control. Robots are of increasing importance and will become an integral part of European society. Despite continuing advances, current robots cannot approach the abilities of even the simplest mammals. Achieving continuous and real-time learning without interference between learnt tasks remains a difficult problem. To attain the learning ability and finesse of movement that animals display, information from a large number of sensorimotor and cognitive signals must be integrated. However, few principles of integration have been proposed. Our inter-disciplinary consortium of physicists, neuroscientists and engineers will investigate the principles of integration and representation in the brain. Real-time spiking neural networks, based in part on the cerebellum, a major site of sensorimotor integration and motor learning in the brain, will be developed. Both hardware and software approaches will be employed. The control abilities of these spiking networks will be evaluated with robots in real environments and compared with state-of-the-art robotic control.
OBJECTIVES
To understand the neural principles that give an organism the ability to learn multiple tasks in real-time with minimal destructive interference between tasks, and to recreate this ability in real-time spiking neural networks:
1) Perform experiments yielding data suitable for an explicit model of the cerebellum;
2) Use the data to simulate and analyse theoretically cerebellar function;
3) Design real-time spiking neural networks that implement beneficial principles identified by the experiments, analysis or simulations;
4) Construct both hardware and simulated networks;
5) Evaluate the control abilities of the real-time spiking networks.
DESCRIPTION OF WORK
In order to achieve our scientific objectives the following will be performed. Learning rules will be derived using information and learning theory with the aim of developing spiking representations that improve learning and coordination abilities. Analytical models will be developed to understand the information processing, the representations and population codes in the cerebellum. The storage capacity and the importance of the learning rules for storage performance will be evaluated. Electrophysiological experiments in vitro (in slices) and in vivo will be used to characterise cerebellar neurones, synapses, and synaptic plasticity. Recordings with multielectrode arrays from cerebellar slices will be performed to investigate the population dynamics of excitation and plasticity in the cerebellum. Cells will be filled and reconstructed in order to construct compartmental models. Detailed computational models of cerebellar neurones will be improved, constructed and analysed. With the aid of neuronal modelling, simplified descriptions of these elements will be prepared that greatly reduce the computing costs of simulating large numbers of neurones.
A neural network simulation environment will be developed to simulate in real-time networks of the order of 1 million of spiking neurones. Neural hardware to speed up simulations will be constructed and combined with existing aVLSI hardware. The performance of large spiking cerebellar models running in the simulation environment will be tested in experiments both with a simulated and a real humanoid robot. The performance of the real-time spiking control mechanisms will be compared with state-of-the-art robotic control systems.
Ámbito científico
- natural sciencescomputer and information sciencessoftware
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringroboticsautonomous robots
- natural sciencesbiological scienceszoologymammalogy
- natural sciencescomputer and information sciencesdata sciencedata processing
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Convocatoria de propuestas
Data not availableRégimen de financiación
CSC - Cost-sharing contractsCoordinador
75230 PARIS CEDEX 05
Francia