Random recurrent neural networks for autonomous system

In this article, we stress the need for using dy-namical systems properties in autonomous architecture design. We rst study the dynamics of random recurrent neural networks (RRNN). Such systems are known to spontaneously exhibits various dynamical regimes, as they always tries to remain on an attractor, thus achieving stable dy-namical behaviors. Second, we try to characterize the adaptive properties of such a system in an open environment, i.e. in a system which always interacts with external signals.Under these conditions, a change in the behavior corresponds to the switch from one attractor to another one. Such bifurcation occur for very little changes in the environment signal; our system is thus unstable on its inputs. We propose a local Hebbian learning rule which tends to stabilize the response of the system for given inputs. After training, the system is able to perform recognition, i.e to produce a speciic regular cyclic attractor while the learned input is present (or even a noisy version of this learned input). Moreover, our system can make associations while learning process takes place under two \sensory" innuences. The system can indeed perform recognition, even when one sensory signal is missing. Our RRNN is then implemented on a robotic system, under visual and sensori-motor innuences. After learning periodic motor sequences in association with visual inputs, our system can now discriminate between matching and unknown visual sequences. When visual sequence matches inner sequence, the system produces regular periodic movements. On the contrary, when there is a connict between visual inputs and inner dynamics, the system tends to produce chaotic aperiodic movements. Our work nally illustrate a very general paradigm on cognitive aspects of perception : what the system perceives depends both on input signal and inner expectations on such input.