Energy-Efficient Photonic Neuromorphic Computing For Telecommunication .
Energy-Efficient Photonic Neuromorphic Computing forTelecommunication ApplicationsEnergie-efficiënt fotonisch neuromorf rekenen voortelecommunicatietoepassingenAndrew KatumbaPromotoren: Prof. Dr. Ir. Peter Bienstman, Prof. Dr. Ir. Joni DambreProefschrift tot het bekomen van de graad vanDoctor in de Ingenieurswetenschappen: FotonicaVakgroep InformatietechnologieVoorzitter: Prof. Bart DhoedtFaculteit Ingenieurswetenschappen en ArchitectuurAcademiejaar 2018-2019
Universiteit GentFaculteit Ingenieurswetenschappen en ArchitectuurVakgroep InformatietechnologiePromotoren:Prof. Dr. Ir. Peter BienstmanProf. Dr. Ir. Joni DambreExamencommissie:Prof. Dr. Ir. Filip de Turck (voorzitter)Prof. Dr. Ir. Peter Bienstman (promotor)Prof. Dr. Ir. Joni Dambre (promotor)Dr. Ir. Thomas Van VaerenberghProf. Dr. Ir. Guy Van der SandeProf. Dr. Ir. Xin YinProf. Dr. Ir. Geert MorthierProf. Dr. Ir. Tom DhaeneUniversiteit GentFaculteit Ingenieurswetenschappen en ArchitectuurVakgroep InformatietechnologieTechnologiepark-Zwijnaarde, 9052 Gent, BelgiëTel.: 32-9-331.49.00Fax.: 32-9-331.48.99Universiteit Gent, INTECUniversiteit Gent, INTECUniversiteit Gent, ELISHewlett Packard Enterprise, USAVrije Universiteit BrusselUniversiteit Gent, INTECUniversiteit Gent, INTECUniversiteit Gent, INTEC
4
AcknowledgementsAny Ph.D. thesis sits atop a pyramid of effort and support from many people; mineis no exception. First and foremost, I would like to thank my promoters Prof. PeterBienstman and Prof. Joni Dambre for the generous advice and guidance from theonset of my research. I am grateful for the freedom and trust you accorded meto explore and experiment extensively. I appreciate the many opportunities youhave created for me to develop my career professionally and to grow as a person.I really couldn’t have wished for better guides on this journey.To my Ph.D. defense jury members, I am grateful for your concise feedback thathas helped to shape-up this thesis to its current quality.I would like to acknowledge the steering team of the Photonics Research Group,with Prof. Roel Baets at the helm. Together with professors Wim, Geert, Bart,Gunther, Nicola and Dries, my work was positively impacted by a short conversation here and a nudge there over the years. I am also thankful to the support staff:Ilse Meersman, Ilse Van Royen, Jasper, Bert, Kristien, Mike, Michael, Peter, andJeroen, for your aid in all matters admin and technical.Turning to the PRC group, we have spent a lot of time in meetings poking andprodding at the various nuances and complexities of reservoir computing. My predecessors Thomas, Martin, and Bendix, your assistance when I was just strikingout centered and grounded the basis upon which the rest of my research was built.Martin and Thomas, thanks for bearing with my flood of Caphe questions and feature requests, even long after you had left the group. As for the current generation,Matthias and Floris, the timing of your entry could not have been more fortuitous.Our discussions and collaboration have without any doubt born numerous fruit; allthe best as you wind up. And to the ”new kids on the block”Stijn, Chonghuai, andEmmanuel, you have brought with you new energy and a fresh take on matters. Iam looking forward to seeing where you will take this burgeoning field of photonicreservoir computing.I would like to appreciate all the collaborators I had the chance to work withboth within and without Gent University. I am especially grateful to the membersof the Horizon 2020 PHRESCO project. I have benefitted tremendously from yourtechnical expertise and our discussions that have shaped my work, especially inthe last 2 years.I would also like to single-out Prof. Xin Yin and Prof. Johan Bauwelinck of
iiIDLAB for their important insights, both and the beginning and towards the endof my Ph.D., into how reservoir computing might fit into a telecom setting.Working in a big research group comes with opportunities for a variety of interactions, both at and outside work. Sanja, Sarah, and Yanlu, we have been officemates since the Technicum days, thanks for the convivial work environment. Thesame goes for you Nayerra, Stijn and Yuting.On the research side, I had the pleasure of collaborating with various people.Amin, Abdul, Kasper, Jochem and Joris, it was always exciting to push the limitsof both devices and equipment. I also recall numerous thought-provoking conversations with Andreas, Jesper, Paul, Jing, Koen, Utsav, Haolan, Antonio, Sören,Thijs and Eva. Thanks for the willingness to share your time and expertise.As for the De Brug and, later on, the 5th floor lunch group, lunch was alwayspackaged with some pretty interesting discussions on just about everything (mostinteresting, some bizarre and occasionally both).Outside the office, Jesper, Paul, Joan, Tommaso, Mahmoud, Sanja, Nina, Alessio,Margherita, Savvina, Ana, Camiel, Irfan, Jing, Nuria, Fabio, Floris, Grigorij, Mattias, Clemens, Hanna, Alexandros and Alejandro, those after-work and weekendmeet-ups served for quality downtime. Not forgetting Sulakshna and Saurav theimpeccable culinary duo. And for those that made the trip to Uganda in 2017, myfamily and I look back fondly to those times.No sooner had I arrived in Belgium than I was adopted by the Okumus, theNjugunas, the Ssimbas, the Mugishas, and the Asiyas. I am eternally thankful foryour hospitality and regard, and for making sure that I was properly settled-in and,in a way, like I never left home.Thanks, Michael, Gloria, Billy, Felly, Marvin, Ray, Agaba, Gina, Kaka, Tori,Ayor, Cathy L., Brenda, Malcolm, Esther, Eric, Cosmas and Onesmus for alwaysmaking the long periods between our meetings seem inconsequential. Michael,your annual visits were inexplicably timed at the moments when I needed an interlude from the rigor of research.The Nkrumah company FX, Michael, Kimera, Deno, Dero, Danny, Paulo, Bukenya and Mugagga, I am humbled by the solidarity and brotherhood. Gracias foralways coming through when it matters.Carlos, Paula, and David Owek, I am indebted to you for always challenging meto dream bigger and continually strive for excellence. Carlos and Paula, aren’t weoverdue for another ”dinner for schmucks”?A great many people have had an impact on the success of my research andpersonal well-being than I could enumerate here. The absence of your name doesnot diminish my gratitude for your generosity.Finally, I would like to thank the GBS and Kyeyune families who have beenwith me at every turn. Your love, guidance, and support have been the beacon forthis journey.Gent, February 2019Andrew Katumba
Table of ContentsAcknowledgementsNederlandse samenvatting1Fotonica . . . . . . . . . . .2Optische communicatie . . .3Fotonisch neuromorf rekenen4Conclusie . . . . . . . . . .i.xxixxiixxiixxiixxviEnglish summary5Photonics . . . . . . . . . . . . . .6Optical Communications . . . . . .7Photonic Neuromorphic Computing8Conclusion . . . . . . . . . . . . .xxviixxviiixxviiixxviiixxxiIntroduction1.1 Introduction . . . . . . . . . . . . . . . . . . . .1.2 Optical Communications . . . . . . . . . . . . .1.3 Silicon Photonics . . . . . . . . . . . . . . . . .1.4 Machine Learning . . . . . . . . . . . . . . . . .1.5 Optical Signal Processing . . . . . . . . . . . . .1.6 Objectives . . . . . . . . . . . . . . . . . . . . .1.7 Thesis outline . . . . . . . . . . . . . . . . . . .1.8 Publications . . . . . . . . . . . . . . . . . . . .1.8.1 Publications in international journals . .1.8.2 Publications in international conferences1.8.3 Patents . . . . . . . . . . . . . . . . . .1.8.4 Book Chapters . . . . . . . . . . . . . .References . . . . . . . . . . . . . . . . . . . . . . . .11466789101011131315.19192021222412.Reservoir Computing2.1 Introduction . . . . . . . . . . . . . . .2.2 Machine Learning . . . . . . . . . . . .2.2.1 ANNs . . . . . . . . . . . . . .2.2.2 Feed Forward Neural Networks2.2.3 Recurrent Neural Networks . .
iv2.32.4Reservoir Computing . . . . . . . . . . . . . . . . . . . .Photonic Reservoir Computing . . . . . . . . . . . . . . .2.4.1 Passive Photonic Reservoir Computing . . . . . .2.5 Reservoir model, training and validation . . . . . . . . . .2.5.1 Model . . . . . . . . . . . . . . . . . . . . . . . .2.5.2 Reservoir training and validation . . . . . . . . . .2.5.3 Regularization . . . . . . . . . . . . . . . . . . .2.5.4 Cross-validation . . . . . . . . . . . . . . . . . .2.5.5 Training a readout with coherently combined states2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . .References . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34Multiple Input Reservoir Computing3.1 Methods . . . . . . . . . . . . . . . . . .3.1.1 Numerical Model . . . . . . . . .3.1.2 Simulation Setup and Parameters3.2 Results and Analysis . . . . . . . . . . .3.2.1 Data Rate Studies . . . . . . . . .3.2.2 Power level analysis . . . . . . .3.2.3 Optimal design . . . . . . . . . .3.3 Conclusion . . . . . . . . . . . . . . . .References . . . . . . . . . . . . . . . . . . . . . mode Reservoir Computing4.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4.1.1 Electromagnetic simulations . . . . . . . . . . . . . . . .4.1.2 Circuit simulations and task setup . . . . . . . . . . . . .4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4.2.1 Design of a multimode Y-junction with low combiner loss4.2.1.1 Waveguide cutoff conditions . . . . . . . . . .4.2.1.2 Relevant geometrical parameters of Y-junction .4.2.1.3 Waveguide Width Optimisation . . . . . . . . .4.2.1.4 Modal Power Distribution . . . . . . . . . . . .4.2.1.5 Taper Length Optimisation . . . . . . . . . . .4.2.1.6 Wavelength dependence . . . . . . . . . . . . .4.2.2 Performance of the improved Y-junctions in photonic reservoirs . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4.2.2.1 Error rates for different data rates . . . . . . . .4.2.2.2 Energy efficiency considerations . . . . . . . .4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5961616262626263646567697070717274
v5Signal Equalization5.1 Simulation Setup and Methodology . . .5.2 Results and Discussion . . . . . . . . . .5.2.1 Metro links . . . . . . . . . . . .5.2.2 High-speed short-reach links . . .5.2.3 Long-Haul Communications link5.3 Conclusion . . . . . . . . . . . . . . . .References . . . . . . . . . . . . . . . . . . . .77798181858889906Towards a photonic reservoir computing hardware implementationwith integrated readout936.1 Optical weighting elements . . . . . . . . . . . . . . . . . . . . . 946.1.1 Vanadium Oxide weights . . . . . . . . . . . . . . . . . . 956.1.1.1 Design and characterization of a silicon photonics chip for VO2 optical weighting elements . . 966.1.2 BTO weights . . . . . . . . . . . . . . . . . . . . . . . . 996.2 Photonic reservoir computing chip with multiple input injectionand integrated readout . . . . . . . . . . . . . . . . . . . . . . . . 1006.2.1 Design of the integrated readouts . . . . . . . . . . . . . 1016.2.2 Characterization . . . . . . . . . . . . . . . . . . . . . . 1026.3 An improved architecture for passive photonic reservoir computing 1066.4 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . 1086.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1097Conclusion and future work7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7.2.1 Improvements in reservoir design . . . . . . . . . . . . .7.2.2 Further progress in telecommunication applications . . . .7.2.3 Experimental demonstration for reservoir with integratedreadouts . . . . . . . . . . . . . . . . . . . . . . . . . . .111111113113113114
List of Figures1234562.12.22.32.42.5Foutdebiet vs totaal ingangsvermogen voor verschillende injectiescenarios. De kleinst meetbare fout, gegeven het aantal bits gebruikt voor testen, is 10 3 . . . . . . . . . . . . . . . . . . . . . . xxivFoutdebiet voor monomode en multimode reservoirs voor de driebitshoofdingsherkenningstaak voor verschillende waarden voor designaalruisverhouding aan de ingang. . . . . . . . . . . . . . . . . xxivBitfoutdebiet voor de neuromorfe fotonische compensator vergelekenmet die van een FIR Feed Forward Equalizer (FFE) voor verschillende vezellengtes met een lanceringsvermogen van 15 mW. NLON - nietlineaire propagatie aan. NL OFF - nietlineaire propagatieaf. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvError rate vs total input power for different injection scenarios.The minimum measurable error, given the number of bits used fortesting, is 10 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxixError rates for single-mode and multimode on the 3 bit header recognition task for different values of the input SNR. . . . . . . . . xxxBit Error Rate of the neuromorphic photonic equalizer comparedto that of an FIR Feed Forward Equalizer (FFE) for different fiberlengths with a launch power 15 mW. NL ON - nonlinear propagation. NL OFF - nonlinear propagation is deactivated. . . . . . . . xxxiActivation functions commonly used in neural networks: (a) sigmoid, (b) tanh, and (c) ReLU. . . . . . . . . . . . . . . . . . . .A feed forward neural Network . . . . . . . . . . . . . . . . . . .A recurrent neural network. . . . . . . . . . . . . . . . . . . . . .Schematic representation of a reservoir computing system. Theinput signal u(t) is fed into the reservoir and the resulting reservoirstates x(t) are used to learn a linear readout that is then used togenerate the output signal y(t). . . . . . . . . . . . . . . . . . . .Signal flow in a 16-node swirl reservoir architecture. The timedependent output at each numbered node is linearly combined toresult in the answer of the computation. As for inputs, dependingon the application the input can be inserted in one or more of thenumbered nodes. . . . . . . . . . . . . . . . . . . . . . . . . . .2323232730
viii2.62.7Unfolding the passive reservoir computing system for 3 time-steps.In the left figure, U is a concatenation of all the input vectors including the bias and X is the concatenation of the reservoir statescollected at each timestep, these will then be used to train the readout of the reservoir. . . . . . . . . . . . . . . . . . . . . . . . . .Schematic for training a 16 node photonic swirl reservoir with coherently combined reservoir states. . . . . . . . . . . . . . . . . .Error rate vs. reservoir interdelay for various nodes for the input to single node case. The minimum acceptable error rate is1.0 10 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3.2 Error rate vs. reservoir interdelay for the different injection strategies. Minimum acceptable error rate is 10 3 . . . . . . . . . . . .3.3 Error rate vs reservoir interdelay for the input to all nodes case.ndelay specifies the separation,in number of bits, of the two bitsused for the XOR computation. . . . . . . . . . . . . . . . . . .3.4 Error rate vs total input power for different injection scenarios.The minimum measurable error, given the number of bits used fortesting, is 10 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . .3.5 Average power distribution over the reservoir nodes for input tonode 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3.6 Average power distribution over the reservoir nodes for input tothe central loop . . . . . . . . . . . . . . . . . . . . . . . . . . .3.7 Average power distribution over the reservoir nodes for input to toall nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3.8 Schematic of the 16-node swirl reservoir architecture highlightingthe central loop nodes that yield the optimal input configuration.This injection strategy results in the best performance - energy efficiency operation mode for this reservoir architecture. . . . . . .3.9 Error rate vs total input power for input to the central swirl loop(nodes [5, 6, 9, 10]). The solid line indicates the mean value overall repetitions while the shaded areas indicate the error boundswithin 1 standard deviation of the mean. . . . . . . . . . . . . . .3.10 Error rate vs reservoir interdelay for input to the central swirl loop(nodes [5, 6, 9, 10]). The solid line indicates the mean value overall repetitions while the shaded areas indicate the error boundswithin 1 standard deviation of the mean. . . . . . . . . . . . . . .30353.14.1Schematic of a photonic reservoir computing setup for handlingtasks involving digital optical signals. The input is a non-returnto-zero on-off-keying (NRZ-OOK) digital optical signal, the reservoir is composed of 16 nodes arranged in a swirl topology andthe output is the bitstream that results from training the readout tosolve a particular task. . . . . . . . . . . . . . . . . . . . . . . .4949505152535353545460
ix4.24.34.44.54.64.74.84.95.15.25.35.4Dispersion diagram for 220 nm SOI waveguide for the TE polarization and a center wavelength of 1300 nm. nef f is the complexeffective index of the mode. [1] . . . . . . . . . . . . . . . . . . .Sketch of the Y-junction indicating the sections critical to its performance [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . .Total transmission (summed over all output modes) in the Y-junctioncombiner for different taper lengths. Results are shown for inputto the upper arm of the junction consisting of the fundamental (0),1st , 2nd and 3rd order modes as well as the average transmissionacross all input modes. The baseline transmission of 50%, for theadiabatic single-mode Y-junction, is also indicated. . . . . . . . .Error rate vs reservoir interdelay for the single-mode reservoir forthe 3-bit header recognition task. . . . . . . . . . . . . . . . . . .Error rate vs reservoir interdelay for the multimode reservoir forthe 3-bit header recognition task. . . . . . . . . . . . . . . . . . .Error rates for single-mode and multimode on the 3 bit headerrecognition task for different values of the input SNR. . . . . . . .Comparison of single-mode and multimode 16 node reservoir average power per node for input to node 0. . . . . . . . . . . . . .Comparison of single-mode and multimode 36 node reservoir average power per node for input to node 0. . . . . . . . . . . . . .68Schematic representation of the photonic reservoir computing setupas is used in this dissertation for impairment removal. The inputsignal is a distorted non-return-to-zero on-off-keying (NRZ-OOK)signal that has traversed a fiber optic link, the integrated photonics reservoir is composed of 16 nodes arranged in a swirl topologyand the reservoir states are recorded at each node with aid of aphotodetector. These states are then recorded and used to train aset of weights that represent the readout function that generatesthe final output signal. Note that in the reservoir, the numberedlightlight blue circles (ni ) are the nodes in the reservoir and in thisarchitecture are the locations where states are combined and split,and also serve as the input and detection points. . . . . . . . . . .Schematic representation of the simulation setup to generate datafor the signal equalization task. The input pseudo-random bit sequence (PRBS) signal is modulated onto a laser signal, transmittedover a fiber link, amplified and filtered, after which the field of theoptical signal is saved to file to be used as input to the photonicreservoir simulation model. OBPF - Optical band pass filter. . . .Error rates vs reservoir interconnection delay and fiber length for16 node PhRC equalizer for an NRZ-OOK link at 10 Gb/s. . . .Error rate vs latency and fiber length for a reservoir with interconnection delay time equal to half the bit duration. . . . . . . . . .7963647171727373828384
x5.5Error rate vs data rate for a reservoir with interconnection delaytime equal to half the bit duration and latency 1 bit. A Soft Decision Forward Error Correction limit (SD-FEC limit) of 0.2 10 2is also shown. Error free operation is possible for all error ratevalues below this limit. . . . . . . . . . . . . . . . . . . . . . . .5.6 BER of the PhRC equalizer as compared to that of an FIR FeedForward Equalizer (FFE) trained on the same amount of data fordifferent fiber lengths. The launch power is set to 15 mW. NLON- nonlinear propagation. NL OFF - nonlinear propagation isdeactivated (Nonlinear Index of the fiber in Table 5.1 is set to 0). .5.7 Schematic representation of the setup used for short reach simulations. The Mach-Zehnder Modulator of Figure 5.2 is replaced byan Electro-Absorptive Modulator. The EDFA at the end of the linkhas also been removed. . . . . . . . . . . . . . . . . . . . . . . .5.8 Voltage-dependent transmission of the EAM used for the high speedshort reach link at 1550nm. . . . . . . . . . . . . . . . . . . . .5.9 BER vs fiber length for 40 Gb/s short reach access link with aDFB and an EAM. The blue dashed line indicates the HD-FEClimit The filled section demarcates the error margin. . . . . . . . .5.10 Schematic representation of the setup used for the long haul simulations. The setup is a modification of Figure 5.2 with a recirculating fiber loop used to reach large propagation distances. Eachspan was fixed to a length Lspan of 100km and the amplifier gainwas set to undo the attenuation in each span. . . . . . . . . . . . .5.11 BER vs number of spans for a long-haul communications link.Blue dashed line indicates the HD-FEC limit. . . . . . . . . . . .6.16.26.36.4A section of the VO2 test chip built for the development of VO2based non-volatile optical elements. The layout fit in an area of4 4 mm2 . Enlarged versions of some of the key components inthe design are given in Figure 6.2 and Figure 6.3 . . . . . . . . . .Example of a race-track resonator device with VO2 and metal pads.Red areas correspond to the silicon photonic structures while theblue ones correspond to the metallic contacts for the VO2 stripes(marked in black). . . . . . . . . . . . . . . . . . . . . . . . . . .An example of a straight waveguide structure used to characterize the loss and characteristics of the metal-insulator transition.Red areas correspond to the silicon photonic structures while theblue ones correspond to the metallic contacts for the VO2 stripes(marked in black). . . . . . . . . . . . . . . . . . . . . . . . . . .Extinction ratio for various gaps between the bus waveguide andthe ring resonator. Racetrack resonators are deposited with VO2strips with width 5 µm. . . . . . . . . . . . . . . . . . . . . . . .8485868787888896979798
xi6.5Demonstration of non-volatile tuning (change in transmission spectrum) of a hybrid Si-BTO race-track device using voltage pulses(100 ns and between 10 and 20 V). The black line indicatesthe range over which the transmission state can be changed (nonvolatile memory window) when operating at a fixed wavelength.With this window, 10 unique levels of transmission can be resolved. Details in and Figure from [7] . . . . . . . . . . . . . . .6.6 Mask plan for the silicon photonics chip consisting of a 2 9 passive swirl photonic reservoir, and three readouts : a silicon readout,a VO2 readout and a BTO readout as demarcated in the figure. Therest of the chip has different test structures to test passives components used in the reservoir and components to develop the BTOand VO2 optical weighting elements. . . . . . . . . . . . . . . . .6.7 The three subsections of the reservoir each connected to one of thethree different readouts: blue - Si readout, red - BTO readout andgreen - VO2 readout . . . . . . . . . . . . . . . . . . . . . . . . .6.8 Layout of the VO2 readout. Dashed red lines indicate connectionsfrom reservoir nodes. (a) is the zoom-in of the pipeline for a singlereadout unit which consists of a ring as the nonlinear element anda phase shifter followed by a VO2 deposited racetrack resonatorthat together constitute a weighting component. (b) shows the fullreadout consisting of all 6 readout units. Wires routing to metalpads for tuning the rings and driving the phase shifter are indicated.Note that the implementation of the Si readout is the same butwithout the deposition of the VO2 . . . . . . . . . . . . . . . . . .6.9 Layout of the BTO Readout. Dashed red lines indicate connections from reservoir nodes. (a) represents the pipeline for a singlereadout unit which consists of a a phase shifter followed by a pairof grating couplers. The grating couplers will be used to couplelight off the SOI chip into the BTO layer bonded on top of thechip. The nonlinear element will be provided by BTO racetrackresonator and the weighting element will be a BTO MZI. (b) showsthe full readout consisting of all 6 readout units. Wires routing tometal pads for tuning the rings and driving the phase shifter areindicated. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6.10 Eye diagram of integrated high speed photodetector on demonstrator chip measured at 4 Gb/s. An open eye (an indicator of errorfree operation) was obtained for input signals upto 4 Gb/s. For themeasurements a G-S-G RF probe was used to contact the PD andthe resulting signal amplified with an RF amplifier before capturewith a real-time oscilloscope. . . . . . . . . . . . . . . . . . . . .6.11 Schematic of the setup used for system level characterization of thephotonic RC chip. CW - Continuous Wave, RF - Radio Frequency.6.12 A slice of states from the chip measured at 4Gb/s at node 16 inFigure 6.7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .99100101103104104105106
xii6.13 Schematic of the new photonic RC architecture. In data flow terms,it is the swirl architecture augmented with loop connections. . . . 107
List of Tables4.14.24.35.15.25.3Transmission for Y-junction combiners of different input and output waveguide widths w1 when the input is a particular modal excitation. The results here correspond to the case of excitation tothe upper arm only. Missing values indicate that that particularmode is not guided for that waveguide width. . . . . . . . . . . . 66Modal decomposition of Y-junction combiner with 1 µm wide waveguides and 0.1 µm taper length . . . . . . . . . . . . . . . . . . . . 67Modal decomposition of Y-junction splitter with 1 µm wide waveguides and 0.1 µm taper length . . . . . . . . . . . . . . . . . . . . 67Key parameters for SSMF used in transmission simulations. . . .Transmitter parameters for metro link setup. . . . . . . . . . . . .Transmitter parameters for high-speed short-reach link setup. . . .808286
List of AcronymsAADCANNAnalog-to-Digital ConverterArtificial Neural NetworkBBPBERBPDCBTOBackpropagationBit error rateBackpropagation-DecorrelationBarium TitanateCCMOSCDCWComplementary metal oxide semiconductorChromatic DispersionContinuous waveDDSPDBPDFEDFBDigital Signal ProcessorDigital BackpropagationDecision Feedback EqualizerDistributed Feedback
xviEEDFAEAMErbium Doped Fiber AmplifierElectro-absorptiion ModulatorFFDTDFFEFFNNFIRFinite-Difference Time-DomainFeed Forward EqualizerFeed Forward Neural NetworkFinite Impulse ResponseHHD-FECHard Decision Forward Error CorrectionIIBMIM-DDIoTInternational Business Machines CorporationIntensity Modulation - Direct-DetectionInternet of ThingsMMLMITMachine LearningMetal-Insulator TransitionNNRZNEPNon-return-to-zeroNoise Equivalent Power
xviiOOOKOBPFOn-Off keyingOptical Bandpass Filter sPPhRCPICPMDPhotonic Reservoir ComputingPhotonic Integrated CircuitPolarization Mode DispersionRRCRNNReservoir ComputingRecurrent Neural NetworkSSOASOISNNSNRSiSD-FECSPMSSMFSemiconductor Optical Amplifiersilicon on insulatorSpiking Neural NetworkSignal-to-noise ratioSiliconSoft Decision Forward Error CorrectionSelf Phase ModulationStandard single-mode fiberTTETMVTransverse ElectricTransverse Magnetic
xviiiVO2Vanadium Oxide
Nederlandse samenvatting–Summary in Dutch–Telecommunicatie en computers zijn de twee meest toonaangevende voorbeeldenvan technologische vooruitgang in onze moderne maatschappij. Hun impact opons leven is zo wijdverbreid dat het ondenkbaar zou zijn dat er een dag voorbij zougaan zonder dat we gebruik maken van een slimme telefoon of een ander gelijkaardig toestel. In belangrijke mate steunen alle facetten van onze moderne maatschappij (gezondheidszorg, financies, zakenleven, onderwijs, .) op vooruitgangin deze domeinen.Maar hoewel onze maatschappij deze technologieën omarmd heeft, heeft deonderliggende infrastructuur last om gelijke tred te houden met de als maar groeiende behoefte aan meer informatie, meer gegevens, meer rekenkracht. Er wordtveel werk verricht om iedere laatste gram capaciteit uit de onderliggende digitaleelektronische infrastructuur te persen, en men probeert nieuwe manieren uit omdeze systemen te bouwen, of om belangrijke flessenhalssubsystemen te vervangen.Als inspiratie voor oplossingen zijn sommige mensen op zoek naar antwoorden in de manier waarop ons menselijk brein werkt. Ons brein heeft vaardigheden in communicatie en informatieverwerking die voor sommige toepassingenefficiënties halen ver boven die van de meest geavanceerde supercomputers. Eenmogelijke aanpak die succesvol gebleke
Examencommissie: Prof. Dr. Ir. Filip de Turck (voorzitter) Universiteit Gent, INTEC Prof. Dr. Ir. Peter Bienstman (promotor) Universiteit Gent, INTEC Prof. Dr. Ir. Joni Dambre (promotor) Universiteit Gent, ELIS Dr. Ir. Thomas Van Vaerenbergh Hewlett Packard Enterprise, USA Prof. Dr. Ir. Guy Van der Sande Vrije Universiteit Brussel
Introduction to Neuromorphic Computing Insights and Challenges What is a neuromorphic computer? Why is neuromorphic computing confusing? What about building a brain? Todd’s Top 10 list of challenges Alternative ways of thinking about a building a brain Closing thoughts Outline . What is a Neuromorphic Computer? A neuromorphic computer is a machine comprising many .
electronic and in the photonic domain focusing on the role and the applications of memristors. We discuss the need and the possibility to conceive a photonic memristor and we offer a positive outlook on the challenges and opportunities for the ambitious goal of realising the next generation of full-optical neuromorphic hardware.
neuromorphic vision datasets [ , ] have been released in recent years, which facilitate the neuromorphic vision application forobject detectionand tracking. Recent years also witness the various applications for detection and tracking tasks with neuromorphic vision sensor such as featuretracking[ , ],linetracking[],andmicroparticle tracking[ ].
Neuromorphic computing spans a broad range of scientific disciplines from materials science to devices, to computer science, to neuroscience, all of which are required to solve the neuromorphic computing grand challenge. In our workshop we focus on the computer science aspects, specifically from a neuromorphic device through an application. .
terial in the silicon-on-insulator (SOI) fabrication process. 2 SILICON PHOTONIC LINKS A photonic link is by fi equipped with a transmit unit (Tx) and a receive unit (Rx). The link architecture based on an inte-grated photonic platform such as Silicon Photonics can be envis-aged as the ones presented in Fig. 1 [7, 10]. The Tx requires in-
The degree of a polynomial is the largest power of xwith a non-zero coe cient, i.e. deg Xd i 0 a ix i! d if a d6 0 : If f(x) Pd i 0 a ixiof degree d, we say that fis monic if a d 1. The leading coe cient of f(x) is the coe cient of xd for d deg(f) and the constant coe cient is the coe cient of x0. For example, take R R.
imaging focus. The new report, Neuromorphic Sensing & Computing delivers an in-depth understanding of the neuromorphic landscape with key technology trends, competitive landscape, market dynamics and segmentation per application. It presents key technical insights and analysis regarding future technology trends and challenges. This
to computer science, to neuroscience, all of which are required to solve the neuromorphic computing grand challenge. In our workshop we focus on the computer science aspects, specifically from a neuromorphic device through an application.
