Professional Documents
Culture Documents
Neuromorphic Photonics
Rama Chaudhary
Department of physics and Computer Science, Dayalbagh Educational Institute, Agra, Uttar Pradesh, India
Abstract—Novel materials and devices in Photonics have the (ii) an external perturbation above the excitability threshold
potential to revolutionize optical information processing, beyond triggers a large excursion from this equilibrium; (iii) the
conventional binary-logic approaches. Laser systems offer a rich system then settles back to the attractor in what is called the
repertoire of useful dynamical behaviors, including the excitable
dynamics also found in the time-resolved spiking of neurons. refractory period, after which the system can be excited again.
Spiking reconciles the expressiveness and efficiency of analog pro- In context of spike processing, excitable systems have
cessing with the robustness and scalability of digital processing. been studied using tools like bifurcation theory[1]. Optical
This paper summarize various approaches for spike processing excitability in semiconductor devices is widely studied the-
using Photonics. The excitable laser systems can simultaneously oretically and experimentally. These include two-section gain
exhibit logic-level restoration, cascadability, and input-output
isolation which are fundamental challenges in optical information and saturable absorber (SA) lasers, semiconductor ring and
processing. This paper review recent exciting developments on microdisk lasers, two-dimensional photonic crystal nanocavi-
candidates excitable semiconductor lasers and device proposed ties, resonant tunneling diode photodetector and laser diode,
in literature for a Photonic spike processor. semiconductor lasers based on optical injection, semiconductor
lasers subjected to optical feedback, and polarization switching
I. I NTRODUCTION
VCSELs.
Recent years have seen much development in information
communication (dominated by optics) and in information II. T WO -S ECTION G AIN AND SA E XCITABLE L ASER
processing(dominated by electronics) in the same medium. In Prucnal and co-workers[6]-[9] and Barbay and co-
the context of information processing, nonlinear dynamical workers[4],[5] have advocated two-section gain and SA ex-
systems have been receiving considerable attention due to citable lasers as a computational primitive for spike process-
their isomorphism to biological networks. Compared to binary- ing. This system has been shown, theoretically and experi-
logic based methods implemented on standard von Neumann mentally, to be analogous to the leaky integrate-and-fire (LIF)
architectures, unconventional processing paradigms that are neuron model[2],[3], commonly employed in computational
neuroinspired are relatively more effective for solving certain neuroscience for modeling biological neural networks but time
tasks, such as pattern analysis, decision-making, optimization, scales of biological neuron and device are vastly different.
and learning. A sparse coding scheme, called spiking has Whereas biological neurons have time constants on the or-
recently been recognized by the neuroscience community as der of milliseconds, carrier lifetimes of laser gain sections
an important neural coding strategy for information process- are typically in the nanosecond range and can go down to
ing. The continued evolution of photonic technologies has picoseconds. Although it is one of the simpler spike based
reawakened interest for a relentless search in neuro-inspired models, the LIF model is capable of universal computations,
optical information processing to complement and enable new and the transmission of information through spike timings.
opportunities and potentially bridge the gap with information
communication in the same substrate. A. Principle of operation
Spiking is a sparse coding scheme with firm code theoretic The simple schematic figure two-section gain and SA ex-
justifications. Information is encoded in the temporal and citable laser is given in fig 1. The device composed of a (i)
spatial relationships between short pulses or spikes. Spike a gain section, (ii) a saturable absorber, and (iii) mirrors for
codes which are digital in amplitude but analog in time exhibit cavity feedback. In the LIF excitable model inputs selectively
the expressiveness and efficiency of analog processing with the perturb the gain optically or electrically. Arrows in fig 1(b)
robustness of digital communication. Spikes are typically re- indicate inputs θ(t) (excitatory pulses and inhibitory pulses)
ceived and generated by nonlinear dynamical systems, and can that change the gain by some amount ∆G. Enough excitatory
be represented and processed dynamically through excitability input causes the system to enter fast dynamics in which a spike
a far from equilibrium nonlinear dynamical mechanism under- is generated, followed by the fast recovery of the absorption
lying all-or-none responses to small perturbations[1]. Q(t) and the slow recovery of the gain G(t).
Excitable systems possess unique regenerative properties The Yamada model describes the behavior of lasers with
and have been employed independent gain and SA sections with an approximately con-
Excitability is behaviorally defined by three main criteria: stant intensity profile across the cavity. The inputs selectively
(i) an unperturbed system rests at a single stable equilibrium; perturb the gain. The gain medium acts as a temporal integrator
B. Results
The two-section excitable laser model with gain and an em-
bedded SA is applicable to different physical manifestations.
Nahmias et al.[8] proposed a compact VCSEL-SA exhibiting
neuron-like behavior. Shastri et al.[9] theoretically and exper-
imentally demonstrated a fiber-based graphene excitable laser.
A more recently proposed device is a two-section distributed
feedback (DFB) excitable laser neuron in a hybrid III-V/silicon
platform[1]. All these devices show behavior very much like
to a biological neuron and can be utilized for computing
Fig. 1. Two-section gain-SA excitable laser and simulation results of an SA purposes. Following figure shows simulated dynamics for
laser behaving as an LIF neuron. Variables were rescaled to fit within the VCSEL-SA based device.
desired range. Values used: A = 4.3, B = 3.52, a = 1.8, G = 0.05, L, I
0.05[8].