Recent Progress Towards a Fully-Integrated Hybrid Silicon Laser NeuronPlatform

Mitchell A. Nahmias†, Alex N. Tait†, Bhavin Shastri†, and Paul R. Prucnal†

†Department of Electrical Engineering, Princeton UniversityPrinceton, NJ 08542

Email: mnahmias@princeton.edu

Abstract—We combine advances in photonic integratedcircuits (PICs) with principles from neuromorphic engi-neering to create a scalable, robust, and extremely highbandwidth bio-inspired computational system. We de-scribe such a system in a wafer-bonded III-V/silicon plat-form, integrating the network (through passive silicon-on-insulator technology) and the computational elements(through active III-V laser devices) in a single substrate,and corroborate its underlying principles through prelimi-nary bench-top demonstrations.

1. Introduction

The ability to map a processing paradigm to its physi-cal implementation, rather than abstracting physical effectsaway entirely, represents an important step in streamlin-ing efficiency and performance. The marriage of photonicswith spike processing—a computational paradigm utilizedin biological neurocircuits—is fundamentally enabled bythe strong analogies of the underlying physics between thedynamics of biological neurons and lasers, both of whichcan be understood within the framework of dynamical sys-tems theory. Integrated photonic platforms offer an alter-native approach to microelectronics. The high switchingspeeds, high communication bandwidth, and low cross-talkachievable in photonics are very well suited for an ultrafastspike-based information scheme with high interconnectiondensities. In addition, the high wall-plug efficiencies ofphotonic devices may allow such implementations to matchor eclipse equivalent electronic systems in low energy us-age. Because of these advantages, photonic spike proces-sors could access a picosecond, low-power computation-ally rich domain that is inaccessible by other technologies.

The photonic spike processor is a hardware buildingblock that, like a logic gate, enables the scalability andnoise robustness necessary to construct arbitrarily complexsystems, but, unlike a logic gate, uses hybrid analog-digitalcodes to most naturally interact with the changing environ-ment of radio spectra. The incorporation of a novel, bio-logically inspired signal processing model to lightwave de-vices imports the potential of high-complexity processingto high-performance photonic hardware. Spike processingcircuits can be implemented in conventional device fabri-cation processes for integrated optical interconnects, like

High-bandwidth analog electronic interface

Reduced-information outputs Spike/optical

encoder Processing Back-end interface

High-level instructions, Data to transmit

Photonic spike processing chip

Sensory modulation feedback, “attention”

Figure 1: A diagram of a fully integrated photonic spikeprocessor. The system could handle many fast-varying(GHz) signals simultaneously and perform complex oper-ations such as classification, recognition, and adaptation.Such a unique processor could find use for spectrally awareRF systems or complex systems analysis.

silicon and CMOS photonics: manufacturing process withbillions of dollars of commercial investment. From thesedevices, our group has shown that unconventional circuitssufficient for information processing can be constructed (asshown in Figure 1). Using developing platforms such ashybrid silicon/III-V PICs [1], on-chip systems with tens ofthousands of reconfigurable elements will be possible, al-lowing the large range of possible behaviors needed for thenext generation of spectrally aware RF systems.

2. Processor Node

Recent years have seen the emergence of a new class ofoptical devices that exploit a dynamical isomorphism be-tween semiconductor photocarriers and neuron biophysics.The difference in physical timescales allows these photonicneurons to exhibit spiking behavior on picosecond (insteadof millisecond) timescales [2]. Spiking is closely related toa dynamical system property that underlies all-or-none re-sponses called excitability, which is shared by certain kindsof laser devices. Excitable laser systems have been studiedin the context of spike processing with the tools of bifurca-tion theory by [3] and experimentally by [4].

(a)

(b)

Figure 2: (a) A circuit diagram of the proposed device.Excitatory and inhibitory inputs are incident on photode-tectors, which represent synaptic variables. The inducedphotocurrent modulates carrier injection into a laser withembedded saturable absorber, which may cause the laser totransmit a pulse back into the network. (b) Analogy with abiological neuron. The soma performs integration and thenapplies a threshold to make a spike or no-spike decision.After a spike is released, the voltage is reset. The resultingspike is sent to other neurons in the network.

The LIF neuron model is a mathematical model of thespiking dynamics which pervade animal nervous systems,well-established as the most widely used model of biolog-ical neurons in theoretical neuroscience for studying com-plex computation in nervous systems. A simple model ofa single-mode laser with saturable absorber (SA) sectionhas been proven to be analogous to the equations governingan LIF neuron in certain parameter regimes (Figure 2) [2].We have recently designed excitable lasers specifically de-signed for compatibility with common photonic integratedcircuit (PIC) platforms [5], and tested the dynamical modelin fiber-based bench-top experiments [6].

2.1. Bench-Top Model

We recently demonstrated [6] a fiber-based excitablelaser as a proof of concept of excitability with an embed-

Fig. 1. (a) Graphene excitable laser. (b) Generation of excitatory inputs. (c) Different topologies of phase space that can occur as the physical parameters (pump current, length of cavity, absorption.) are varied. We desire excitability in the second phase portrait (outlined in red).

Fig. 2. Experimental results. For each plot, excitatory optical inputs on top (in blue) and response of graphene excitable laser at bottom (in red).

In conclusion, we have demonstrated a novel excitable laser employing passively Q-switching with a graphene-based SA. Such an excitable system has recently been theoretically shown to behave analogously to a spiking neuron [5][11], opening up applications to biologically inspired cortical algorithms for learning and adaptive control. Furthermore, an integrated version of this excitable laser could also be an enabler for applications of optical computing [12].

References [1] R. Sarpeshkar, “Analog versus digital: extrapolating from electronics to neurobiology,” Neural Comput. 10(7), 1601–1638 (1998). [2] B. Krauskopf, K. Schneider, J. Sieber, S. Wieczorek, and M. Wolfrum,“Excitability and self-pulsations near homoclinic bifurcations in semiconductor laser systems,” Optics Commun. 215(46), 367–379 (2003). [3] K. Vandoorne et al., “Toward optical signal processing using photonic reservoir computing,” Opt. Exp. 16(15), 11182–11192 (2008). [4] B. J. Shastri, M. A. Nahmias, A. N. Tait, Y. Tian, M. P. Fok, M. P. Chang, B. Wu, and P. R. Prucnal, “Exploring excitability in graphene for spike processing networks,” in Proc. of Numerical Simulation of Optoelectronic Devices (NUSOD), Aug. 2013, paper TuC5. pp. 83–84. [5] M. A. Nahmias, B. J. Shastri, A. N. Tait, and P. R. Prucnal, “A leaky integrate-and-fire laser neuron for ultrafast cognitive computing,” IEEE J. Sel. Topics Quantum Electron. 19(5), 1800212 (2013). [6] S. Barbay et al., “Excitability in a semiconductor laser with saturable absorber,” Opt. Lett. 36(23), 4476–4478 (2011). [7] G. J. Spuhler, R. Paschotta, R. Fluck, B. Braun, M. Moser, G. Zhang, E. Gini, and U. Keller, “Experimentally confirmed design guidelines for passively Q-switched microchip lasers using semiconductor saturable absorbers,” Opt. Soc. Am. B. 16(3), 376–388 (1999). [8] A. Yamashita, “A Tutorial on nonlinear photonic applications of carbon nanotube and graphene,” J. Lightw. Technol. 30(4), 427–447 (2012). [9] G. Sobon, J. Sotor, J. Jagiello, R. Kozinski, K. Librant, M. Zdrojek, L. Lipinska, and K. M. Abramski, “Linearly polarized, Q-switched Er-doped fiber laser based on reduced graphene oxide saturable absorber”, Appl. Phys. Lett. 101(24), 241106 (2012). [10] E. M. Izhikevich, N. S. Desai, E. C. Walcott, and F. C. Hoppensteadt, “Bursts as a unit of neural information: selective communication via resonance”, Trends in Neuroscience 26(3), 161–167 (2003). [11] M. A. Nahmias, A. N. Tait, B. J. Shastri, and P. R. Prucnal, “An evanescent hybrid silicon laser neuron,” in Proc. IEEE Photonics Conf. (IPC), Sep. 2013, paper ME3.4. [12] D. Woods and T. J. Naughton, “Optical computing: Photonic neural networks,” Nature Physics 8(4), 257-259 (2012).

Fig. 1. (a) Graphene excitable laser. (b) Generation of excitatory inputs. (c) Different topologies of phase space that can occur as the physical parameters (pump current, length of cavity, absorption.) are varied. We desire excitability in the second phase portrait (outlined in red).

Fig. 2. Experimental results. For each plot, excitatory optical inputs on top (in blue) and response of graphene excitable laser at bottom (in red).

In conclusion, we have demonstrated a novel excitable laser employing passively Q-switching with a graphene-based SA. Such an excitable system has recently been theoretically shown to behave analogously to a spiking neuron [5][11], opening up applications to biologically inspired cortical algorithms for learning and adaptive control. Furthermore, an integrated version of this excitable laser could also be an enabler for applications of optical computing [12].

References [1] R. Sarpeshkar, “Analog versus digital: extrapolating from electronics to neurobiology,” Neural Comput. 10(7), 1601–1638 (1998). [2] B. Krauskopf, K. Schneider, J. Sieber, S. Wieczorek, and M. Wolfrum,“Excitability and self-pulsations near homoclinic bifurcations in semiconductor laser systems,” Optics Commun. 215(46), 367–379 (2003). [3] K. Vandoorne et al., “Toward optical signal processing using photonic reservoir computing,” Opt. Exp. 16(15), 11182–11192 (2008). [4] B. J. Shastri, M. A. Nahmias, A. N. Tait, Y. Tian, M. P. Fok, M. P. Chang, B. Wu, and P. R. Prucnal, “Exploring excitability in graphene for spike processing networks,” in Proc. of Numerical Simulation of Optoelectronic Devices (NUSOD), Aug. 2013, paper TuC5. pp. 83–84. [5] M. A. Nahmias, B. J. Shastri, A. N. Tait, and P. R. Prucnal, “A leaky integrate-and-fire laser neuron for ultrafast cognitive computing,” IEEE J. Sel. Topics Quantum Electron. 19(5), 1800212 (2013). [6] S. Barbay et al., “Excitability in a semiconductor laser with saturable absorber,” Opt. Lett. 36(23), 4476–4478 (2011). [7] G. J. Spuhler, R. Paschotta, R. Fluck, B. Braun, M. Moser, G. Zhang, E. Gini, and U. Keller, “Experimentally confirmed design guidelines for passively Q-switched microchip lasers using semiconductor saturable absorbers,” Opt. Soc. Am. B. 16(3), 376–388 (1999). [8] A. Yamashita, “A Tutorial on nonlinear photonic applications of carbon nanotube and graphene,” J. Lightw. Technol. 30(4), 427–447 (2012). [9] G. Sobon, J. Sotor, J. Jagiello, R. Kozinski, K. Librant, M. Zdrojek, L. Lipinska, and K. M. Abramski, “Linearly polarized, Q-switched Er-doped fiber laser based on reduced graphene oxide saturable absorber”, Appl. Phys. Lett. 101(24), 241106 (2012). [10] E. M. Izhikevich, N. S. Desai, E. C. Walcott, and F. C. Hoppensteadt, “Bursts as a unit of neural information: selective communication via resonance”, Trends in Neuroscience 26(3), 161–167 (2003). [11] M. A. Nahmias, A. N. Tait, B. J. Shastri, and P. R. Prucnal, “An evanescent hybrid silicon laser neuron,” in Proc. IEEE Photonics Conf. (IPC), Sep. 2013, paper ME3.4. [12] D. Woods and T. J. Naughton, “Optical computing: Photonic neural networks,” Nature Physics 8(4), 257-259 (2012).

Figure 3: Experimental data of excitable fiber laser. (a)Traces showing a response to multiple integrated stimulior one strong stimulus. (b) For too weak of an input, nochange in output is detectable.

ded SA. While the performance of this benchtop prototypeis much less than that of an integrated version (bandwidth:100 kHz, energy per spike: 10 nJ), it experimentally con-firms the possibility of using laser systems to emulate thespike processing capabilities required for cortical process-ing: temporal integration of multiple inputs, threshold de-tection, and all-or-nothing pulse generation. Figure 3 is ademonstration of the system’s ability to exhibit excitabil-ity when a series of spikes are incident upon it. Excitatorypulses increase the gain carrier concentration, which per-forms temporal integration. Enough excitation results in anexcursion from equilibrium causing the laser to fire a pulsedue to the saturation of the absorber to transparency. Thisis followed by a relative refractory period during which anexcitatory pulse is unable to cause the laser to fire. Thephase-space excursion resulting from an excitable responseis stereotyped and repeatable while subthreshold activationresults in no output: key all-or- nothing properties for pulseregeneration, reshaping, and signal integrity.

2.2. Integrated Excitable Laser

As illustrated in Figure 2, our device consists of threeprimary components: two photodetectors and an excitablelaser. The photodetectors receive optical pulses from a net-work and produce a push-pull current signal which modu-lates the laser carrier injection. The excitable laser acts asa threshold decision maker and clean pulse generator anal-

An Evanescent Hybrid Silicon Laser Neuron

Mitchell A. Nahmias, Alexander N. Tait, Bhavin J. Shastri, and Paul R. PrucnalLightwave Communications Laboratory, Department of Electrical Engineering, Princeton University, Princeton, NJ 08544, USA

mnahmias@princeton.edu

Abstract—We propose a hybrid silicon laser and photodetectorsystem that can emulate the electro-physiological behavior of areal neuron at ultrafast time-scales. Networks of lasers wouldscale up easily using a silicon III-V wafer-bonding platform.

I. INTRODUCTION

Neuromorphic networks have experienced a surge of popu-larity over the last twenty years, motivated in part by bringingcomputation closer to its underlying physics [1]. Scaling tolarger numbers of neurons, however, requires prohibitivelycomplex networks in which there is a fundamental density-bandwidth trade-off. Alternatively, photonic systems can po-tentially offer much higher bandwidths and lower energyusage than electronics, making the spike-based approach toinformation processing a perfect fit for the technology. Today,there are a growing number of applications that require higherspeeds and lower latencies that may be outside the abilitiesof the fastest electronic circuits, including processing of theRF spectrum or ultrafast control. Taking advantage of theephemeral dynamics in photonic systems such as those inlasers could lead to processors that operate on picosecondtime scales. In addition, there is a close analogy between thedynamics of lasers and those of biological systems, both ofwhich can exhibit excitability [2].

Here, we propose a hybrid silicon distributed feedback(DFB) laser and photodetector system that can emulate botha Leaky Integrate-and-Fire (L&F) neuron and a synapticvariable, completing a computational paradigm that can beused to emulate a wide variety of functional cortical algorithms[3]. Lasers offer a speed increase of about seven orders ofmagnitude, outpacing both biological and electronic systems.Networks of such devices are easily scalable using a siliconIII-V wafer-bonding platform [4].

II. DEVICE DESCRIPTION

As illustrated in Fig. 1, the device consists of three pri-mary components: two photodetectors and an excitable laser.The photodetectors receive optical pulses from a networkand produce a push-pull current signal which modulates thelaser carrier injection. The excitable laser acts as a thresholddecision maker and clean pulse generator analogous to theneural axon hillock. A simple model of a single-mode laserwith saturable absorber (SA) section has been proven to beanalogous to the equations governing an L&F neuron in certainparameter regimes [2]. We chose to implement this modelusing a hybrid silicon evanescent DFB laser [5], the deviceon the left in Fig. 2. The photodetector front-end proposed in

Gain SA

input current excitatory input

inhibitory input output

Fig. 1. A circuit diagram of the proposed device. Excitatory and inhibitoryinputs are incident on photodetectors, which represent synaptic variables. Theinduced photocurrent modulates carrier injection into a laser with embeddedsaturable absorber, which may cause the laser to transmit a pulse back into thenetwork. The resulting output travels to other devices in the network. Dottedlines represent low-frequency pumping and bias wires.

gain

SA

photodetector

RF connection

output input(s)

+V voltage bias

grou

nd

Fig. 2. Cross-section of a silicon hybrid evanescent laser neuron design.Optically active III-V components are bonded to an SOI substrate with ribwaveguides. Optical spikes of potentially multiple wavelengths are incidentonto the photodetector (right) from a passive SOI network. The resulting RFcurrent pulse modulates the gain section of a two-section laser (left). (In-hibitory photodetector and pumping current source not shown for simplicity.)Legend : Si (gray), dielectric insulator (blue), III-V lower etch level (orange),III-V upper etch level (yellow), metal (white), proton implanted III-V (black).

this paper adds full synaptic conduction dynamics that greatlyimprove the capabilities and biological accuracy of the laserneuron.

Much of the energy cost of electronic conversion in opticalsystems comes from the need for high-speed clocked transistorcircuitry and the need to demultiplex WDM channels beforeconversion. In our case, electronic conversion does not havethe goal of signal regeneration, but instead of exploitingelectronic physics for intermediate analog processing. Theuse of passive integrated electrical wires does not sacrificebandwidth or sensitivity in this device. Pulse reshaping is notrequired due to the clean pulse generation in the excitablelaser. The conversion between optical and electronic domainsalso cutails the propagation of optical phase noise and the needfor direct wavelength conversion, thus eliminating two major

Figure 4: A schematic of an integrated excitable laser fullyintegrated into a hybrid silicon/III-V platform. This devicecan interface with a passive SOI network. Only one pho-todetector (PD) is shown.

ogous to the neural axon hillock. We chose to implementthis model using a hybrid silicon evanescent DFB laser [5],shown in Figure 4.

Much of the energy cost of electronic conversion in op-tical systems comes from the need for high-speed clockedtransistor circuitry and the need to demultiplex wavelength-division multiplexed (WDM) channels before conversion.In our case, electronic conversion does not have the goalof signal regeneration, but instead of exploiting electronicphysics for intermediate analog processing. The use ofpassive integrated electrical wires does not sacrifice band-width or sensitivity in this device. Pulse reshaping is notrequired due to the clean pulse generation in the excitablelaser. The conversion between optical and electronic do-mains also curtails the propagation of optical phase noiseand the need for direct wavelength conversion, thus elim-inating two major barriers facing scalable optical comput-ing. Every device in the primary signal pathway performsboth physical and computational roles, resulting in a ro-bust, ultrafast, expressive, and extremely efficient signalprocessing primitive.

3. Network

CMOS Memory"

Excitatory Weight "Bank {λi ; i≠j}"

Inhibitory Weight "Bank {λi ; i≠j}"

Current"Pump"

IN#

Output"Cross-"

Coupler"(λj)"

THROUGH#

Laser Neuron (λj)"

Broadcast Waveguide"

optical link" electrical link" control wire"

Label'photodiodes'

PDs"

Figure 5: An example of a network node, complete withpassive filters for neural weights.

The communication potential of optical interconnects

(bandwidth, energy use, electrical isolation) have receivedattention for neural networking in the past; however, at-tempts to realize holographic or matrix-vector multiplica-tion systems have encountered practical barriers, largelybecause they cannot be integrated. Techniques in siliconPIC fabrication is driven by a tremendous demand for op-tical communication links within conventional supercom-puting systems [7]. The first platforms for systems inte-gration of active photonics are becoming commercial real-ity [1], and promise to bring the economies of integratedcircuit manufacturing to optical systems. Our work investi-gates the potential of modern PIC platforms to enable large-scale all-optical systems for unconventional and/or analogcomputing, using a standard device set designed for digitalcommunication (waveguides, filters, detectors, etc.).

3.1. Broadcast-and-Weight

λ1"

λ2"

λ3"

λ4"

Waveguide"WDM"

Figure 6: Diagram of a broadcast loop. Each unit has fullspectrum access to the outputs of every other unit withinthe loop.

Our scheme leverages recent advances in PIC technol-ogy to address interconnect challenges faced by distributedprocessing. It has similarities with the fiber networkingtechnique broadcast-and-select, which channelizes usablebandwidth using WDM; however, the protocol flattens thetraditional layered hierarchy of optical networks, accom-plishing physical, logical, and processing tasks in a com-pact network protocol. Although its processing circuits areunconventional, the required device set is compatible withmainstream PIC platforms. WDM effectively channelizesavailable bandwidth without spatial or holographic multi-plexing and avoids coherent interference effects during fan-in. High-bandwidth optical channels are compatible withour proposed laser neuron devices, which could access a

picosecond computational domain that impacts applicationareas where both complexity and speed are paramount.

In this scheme, a group of nodes (a node is illustratedin Figure 4) shares a common medium in which the out-put of every node is assigned a unique transmission wave-length and made available to every other node (Figure 6).Each node has a tunable spectral filter bank at its front-end. By tuning continuously between the ON/OFF reso-nant states, each filter drops a portion of its correspond-ing wavelength channel, thereby applying a coefficient oftransmission analogous to a neural weight. The filters ofa given receiver operate in parallel, allowing it to receivemultiple inputs simultaneously. An interconnectivity pat-tern is determined by the local states of filters and not astate of the transmission medium between nodes. Rout-ing in this network is transparent, massively parallel, andswitchless, making it ideal to support asynchronous signalsof a neural character.

3.2. Silicon Photonic Integration

Since routing is already performed by filters at the frontend of each laser neuron described in the previous sec-tion, the broadcast medium must simply implement an all-to-all interconnection, supporting all N2 potential connec-tions between participating units. To satisfy these require-ments, we use a loop-based architecture: a single integratedwaveguide with topology of a loop (i.e. ring). A broadcast-and-weight cell thus consists of several laser neuron prim-itives coupled to a BL medium, as illustrated in Figure 6.This architecture allows each node in the network full ac-cess to signals from every other network node. Filtering ateach node (Figure 5) allows neural weights to be applied.

The ability to control each connection, each weight, in-dependently is critical for creating differentiation amongstthe processing elements. A great variety of possible weightprofiles allows a group of functionally similar units to com-pute a tremendous variety of functions despite sharing acommon set of available input signals. Reconfiguration ofthe filters’ drop states, corresponding to weight adaptationor learning, intentionally occurs on timescales much slowerthan spike signaling. Reconfigurable filters can be imple-mented by a micro-ring resonator whose resonance is tunedthermally or electronically. The broadcast medium couldbe a silicon-on-insulator (SOI) waveguide, which is fullycompatible with the laser neuron structure described in theprevious section.

4. Conclusion

We have described a potential platform—enabled byunique optoelectronic physics and the recent emergence ofscalable photonic integrated circuits (PICs)—that emulatesbiological networks of neurons at ultrafast speeds. Sili-con photonic platform development has revolved aroundpoint-to-point links for multi-core computing systems. We

have examined an opportunity for this technology to extendto unconventional architectures that rely heavily on inter-connect performance. Broadcast-and-weight is a new ap-proach for joining neuron-inspired processing and opticalinterconnect physics. The LIF model with a synaptic vari-able, coupled with tunable routing in a passive SOI networkon a scalable platform, could open computational domainsthat demand unprecedented temporal precision, power effi-ciency, and functional complexity. Its enormous bandwidthcapabilities would allow for a system to emulate corticalfunctions at the time scale of radio frequency (RF) waves,creating a system fully cognizant of all aspects of the RFspectrum in real-time. In the long term, the platform couldprovide wide-band, reconfigurable, and robust communi-cation with flexible spectrum access in the RF spectrum ona single chip.

Acknowledgments

The authors would like to thank Yue Tian, Mable Fok,and David Rosenbluth for their fruitful suggestions andsupport.

References

[1] Di Liang and John E Bowers. Recent progress in laserson silicon. Nature Photonics, 4(8):511–517, 2010.

[2] Mitchell A Nahmias, Bhavin J Shastri, Alexander NTait, and Paul R Prucnal. A leaky integrate-and-fire laser neuron for ultrafast cognitive computing.19(5):1800212, Sep./Oct. 2013.

[3] B. Krauskopf, K. Schneider, J. Sieber, S. Wieczorek,and M. Wolfrum. Excitability and self-pulsationsnear homoclinic bifurcations in semiconductor lasersystems. Opt. Commun., 215(4-6):367–379, January2003.

[4] S. Barbay, R. Kuszelewicz, and A. M. Yacomotti. Ex-citability in a semiconductor laser with saturable ab-sorber. Opt. Lett., 36(23):4476–4478, December 2011.

[5] M. A. Nahmias, A. N. Tait, B. J. Shastri, and P. R.Prucnal. An evanescent hybrid silicon laser neuron. InProc. IEEE Photonics Conf. (IPC), pages 93–44, Seat-tle, WA, USA, September 2013.

[6] B. J. Shastri, M. A. Nahmias, A. N. Tait, Y. Tian,B. Wu, and P. R. Prucnal. Graphene excitable laser forphotonic spike processing. In Proc. IEEE PhotonicsConf. (IPC), Seattle, WA, USA, September 2013.

[7] Bahram Jalali and Sasan Fathpour. Silicon photonics.Lightwave Technology, Journal of, 24(12):4600–4615,2006.