par Smerieri, Anteo ;Duport, Francois ;Haelterman, Marc ;Massar, Serge
Référence 2013 International Workshop on Soft Robotics and Morphological Computation(July 14-19, 2013: Centro Stefano Franscini Ascona, Switzerland), International Workshop on Soft Robotics and Morphological Computation 2013, Bio-Inspired Robotics Lab, ETH Zürich, Zürich, Switzerland, page (0-28)
Publication Publié, 2013-07-18
Publication dans des actes
Résumé : Reservoir Computing is a recently introduced computational tool that overcomes someof the weaknesses of recurrent neural networks[1]. A reservoir computer is composed bya recurrent nonlinear network, which performs feature recognition on the inputs, and alinear output layer. Learning only occurs in the linear layer; a reservoir computer cantherefore use many more nonlinear units in the recurrent layer without becomingunmaneageable. Reservoir computers excel in processing time-dependent signals, withperformances at the state-of-the-art level for various tasks; at the same time, they aregood candidates for hardware implementations.Here we present an overview of our lab‘s activities in the field of physical reservoircomputers. We have built an optoelectronic[2] and a full-optical[3] reservoir computer,both based on time multiplexing, i.e. the idea of using a single nonlinear node and a delayline rather than several physically distinct nonlinear nodes.For the optoelectronic setup we use a Mach-Zehnder modulator as the nonlinear element,while we use a semiconductor optical amplifier (SOA) for the all-optical setup; a fiberspool is used in both cases as the delay line. We show in both cases the performance onbenchmark tasks, which is on par with the one from software reservoirs of comparablesizes.We also show an hardware implementation of the analogh readout layer[4]. For this, weuse a Mach-Zehnder modulator to multiply the reservoir states, encoded as lightintensities, by arbitrary coefficients, and a capacitor to integrate the multiplicated statesand produce the output of the linear layer.Preliminary results show that while we sacrifice some of the performance of our reservoirusing this readout, we also gain a factor of 20 in its operating speed, by removing theneed for postprocessing the reservoir outputs. The main advantage of the analog readouthowever is that it opens the possibility of online training, which we hope that will lead toa further improvement on the reservoir performance and eventually to the first allhardware,self-operating reservoir.References[1]M. Lukosevicius and H. Jaeger, Computer Science Review, 3127–149, 2009.[2]Y. Paquot, F. Duport, A. Smerieri, J. Dambre, B. Schrauwen, M. Haelterman, and S.Massar, Scientific reports, 2, 287, 2012.[3]F. Duport, B. Schneider, A. Smerieri, M. Haelterman, and S. Massar, Optics express,vol. 20, 20,22783–95, 2012.[4]A. Smerieri, F. Duport, Y. Paquot, B. Schrauwen, M. Haelterman, and S. Massar,Proceedings of NIPS, 2012.International Workshop on Soft Robotics and Morphological Computation 2013O