ThA3 (Invited)
Introduction
The photonic computational primitive presented in this paper is the first all optical implementation of a
spiking leaky integrate and fire (LIF) neuron. Due to its hybrid analog/digital nature it is capable of overcoming
both the noise problems of purely analog optical devices and the limited computational capabilities of individual
digital optical devices. It is possible to chain together large numbers of these devices, each individually capable
of significant computation, to implement far more complex computations on high bandwidth signals than is
currently possible. This device represents both a technology capable of scaling the complexity of computations
that can be performed on high-bandwidth signals, and a contribution to the nacent field of photonic
neuromorphic engineering. This technology applies to a wide range of avionics applications in which data
bandwidth is too high or the response delay too short for electronic processing.
Analog computation can be more efficient than digital computation when there is a straightforward
mapping between the operations needed in the computation and the analog device properties of the technology
used. Analog optical computation has found widespread application, for example, in filtering microwave
signals. However, the complexity of these computations is limited by noise accumulation. In this paper we
present a hybrid analog/digital spike-processing primitive that elegantly implements the functionality of an
integrate-and-fire neuron. An optical signal processing system founded upon these spike-processing devices has
the potential to be scalable, computationally powerful, and have high bandwidth.
In the spike-processing paradigm, analog processing is performed within the device (weighting, delaying,
spatial integration, temporal integration), and information is transmitted between devices only through the
timing of output spikes generated through thresholding. This processing paradigm evolved in nervous systems
as a means to overcome the problem of noise accumulation inherent in purely analog computation [4]. The
integrate-and-fire neuron is one of the most widely used abstract models of spiking neurons in computational
neuroscience [2] because it retains much of the richness and computational power of real neurons despite its
simplicity. Interest in spike processing for various engineering applications has been growing [3]. Spike
processing algorithms are understood in a number of important biological sensory processing systems and are
growing use in signal processing applications [3]. From the standpoint of computability and complexity theory,
integrate-and-fire neurons have been proven to be powerful computational primitives capable of simulating both
Turing machines and traditional neural networks [3].
The photonic implementation presented here exploits many of the technologies already developed for
telecommunications systems (diode lasers, EO modulators, inline fiber optic amplifiers, semiconductor optical
amplifiers, non-linear fibers based on doped glasses, and photonic crystals) that have useful non-linear switching
properties. Previous fiber based implementations of neural computation have been limited to analog approaches
using Silicon Optical Amplifiers that are prone to noise accumulation and instability when cascaded [3]. Our
approach takes advantage of recently developed passive non-linear waveguide optical devices that are noise
free, high bandwidth, and capable of performing non-linear operations.
50