Optoelectronic technology for implementation of pulsed neural networks
Alexandre Ricardo Soares Romariz, Kelvin H. Wagner · 2003
In this work, a new optoelectronic implementation of nonlinear oscillators representing features of neural dynamics for computational purposes is proposed and evaluated. In the area of artificial neural networks, the incorporation of dynamic effects in neural models is a matter of growing interest, as new possibilities arise when the pulsed character of the communication between neurons is taken into account. Features of different neural dynamic models are discussed. A particular two-dimensional model (FitzHugh-Nagumo) was chosen for the implementation, mainly because it depends on a single nonlinearity. The optical implementation of the required nonlinearity is done with the aid of Vertical-Cavity Surface-Emitting Lasers (VCSELs). The fact that the emitted wavelength of a VCSEL varies with driving current is used, in combination with a birefringence-based filter, to produce a nonlinear map from driving voltage to detected power without the use of nonlinear devices. Linear electronic feedback completes the implementation, in which optical and electronic input signals can be easily combined. This nonlinear mapping is different from the third-degree polynomial used in the original neural model, but the conditions for preservation of the main dynamical features are found through first-order stability analysis and concepts from bifurcation theory. A simplified model of the dynamics of the thermal effects in the VCSEL correctly predicts the qualitative changes in the nonlinear mapping implementation from low frequencies to a few MHz. Experiments and simulations of an isolated optoelectronic implementation are shown, as well as results from the optical coupling between two similar pulsing artificial neurons. Simulations of whole networks illustrate interesting dynamical features which are not readily available in traditional static neural networks. Another possible analogy between spike-processing and optical signal processing is identified in the case where individual photons are detected from a low-intensity coherent field. In such a regime, the discrete character of photodetection events presents itself naturally as a series of pulses with high variance. A preliminary investigation on the amount of information that individual photons can convey about an optical signal is made through analysis and simulation of Poisson stream processing in a simplified digit-recognition task.