Research themes in the Uchida Lab
(3) Reservoir computing using laser dynamics

1.The concept of reservoir computing

Studies of information processing based on neural networks modelling the structure of the brain have been thriving. Recently, a novel scheme termed reservoir computing has been proposed [1] in which a fixed network, referred to as reservoir, is used for information processing. Many studies are considering this technique for machine learning and for implementing non-Von Neumann-type computers.
The choice of the word “reservoir” to describe this technique can be explained as follows. If we consider an experiment in which somebody throws rocks in a reservoir, ripples will appear on the surface of the water. The evolution of these ripples seems complicated, yet since there is a precisely defined correspondence between a given rock and the related ripples, by observing the ripples one can estimate the rock’s properties such as its shape and its speed. Reservoir computing can be seen as a process in which the information of the rocks thrown into the reservoir is estimated by observing and interpreting the information of the ripples they generated on its surface.

2.Reservoir computing and semiconductor lasers

Under given conditions on the input and output signals, fast-pulsing irregular laser oscillations can show reproducibility (this property is termed consistency) [2]. Reservoir computing can be implemented with lasers in the role of reservoirs and showing the necessary reproducibility property.

This research theme aims at performing reservoir computing by using a semiconductor laser with delayed optical feedback in place of a neural network. In this method, an input signal on which a mask signal is applied is used to modulate a first laser (Laser 1) which in turn injects its output signal into a second laser (Laser 2). The output signal of this second laser describes a feedback loop in which virtual nodes can be defined by sampling at regular time intervals. The value of the output signal is calculated by adding the states of all virtual nodes to which weights have been applied. In order to obtain a one-to-one correspondence between input and output signals, the weights of the nodes are determined by a learning process. The advantage of this scheme based on laser dynamics is that information processing speeds reaching several GHz can be achieved. This research is carried out both in simulation and experiment and finds applications in time series prediction and speech recognition. Besides, quantitative evaluation of system complexity and information processing effectiveness analysis are being carried out as well.

References

[1] H. Jaeger and H. Haas, Science, Vol. 304, pp. 78 (2004).

[2] A. Uchida, et al., Physical Review Letters, Vol. 93, pp. 244102 (2004).

[3] L. Appeltant, et al., Nature Communications, Vol. 2, pp. 468 (2011).