Fiber-based artificial neuron unit

Abstract

An artificial neuron unit for processing a signal is presented, comprising: a fiber-based optical processing unit having first and second optical input ports and first and second optical output ports, configured and operative to controllably apply optical processing to incoherent input optical signals to generate first and second weighted combined optical signals; and an electro-optical processing unit configured and operative to process the first and second weighted combined optical signals, the electro-optical processing unit configured and operative to generate a weighted sum output of the artificial neuron unit by successively performing the steps of applying a predetermined mathematical function to the first and second weighted combined optical signals corresponding to positive and negative weights to generate a resultant signal, and applying nonlinear processing to the resultant signal to convert the resultant signal into an optical output signal indicative of the weighted sum output of the artificial neuron unit.

Description

[Technical Field]

[0001] The present disclosure is in the field of artificial neural networks and relates to fiber-based neuron units and neural networks utilizing same. [Background technology]

[0002] The following references are considered relevant as background to the subject matter of this disclosure. 1. R. Xu, P. Lv, F. 2. X. Sui, Q. Wu, J. Liu, Q. Chen, and G. Gu, “A review of optical neural networks”, IEEE Access, vol. 8, pp. 70773-70783, 2020, DOI:10.1109 / ACCESS.2020.2987333. 3. B. Shi, N. Calabretta, and R. Stabile, “Deep Neural Network Through an InP SOA-Based Photonic Integrated Cross-Connect”, IEEE Journal of Selected Topics in Quantum Electronics, vol. 26, pp. 1-11, 2020, DOI: 10.1109 / JSTQE.2019.2945548. 4.A.Totovic,G.Giamougiannis,A.Tsakyridis,D.Lazovsky,and N.Pieros,「Programmable photonic neural networks combining WDM with coherent linear optics」,Scientific reports,vol.12,pp.5605,2022,DOI:10.1038 / s41598-022-09370-y. 5.H.Zhang,M.Gu,X.D.Jiang,J.Thompson,H.Cai,S.Paesani,R.Santagati,A.Laing,Y.Zhang,M.H.Yung,Y.Z.Shi,F.K.Muhammad,G.Q.Lo,X.S.Luo,B.Dong,D.L.Kwong,L.C.Kwek,and A.Q.Liu,「An optical neural chip for implementing complex-valued neural network」,Nature Communications,vol.12,pp.1-11,2021,DOI:10.1038 / s41467-020-20719-7. 6.J.Liu,Q.Wu,X.Sui,Q.Chen,G.Gu,L.Wang,and S.Li,「Research progress in optical neural networks:theory,applications and developments」,PhotoniX,vol.2,pp.1-39,2021,DOI:10.1186 / s43074-021-00026-0. 7.E.Cohen,D.Malka,A.Shemer,A.Shahmoon,Z.Zalevsky,and M.London,「Neural networks within multi-core optic fibers」,Scientific Reports,vol.6,pp.29080,2016,DOI:10.1038 / srep29080. 8. ANTait, T. Ferreira De Lima, MANahmias, HBMiller, HTPeng, BJ Shastri, and PRPrucnal, “Silicon Photonic Modulator Neuron”, Physical Review Applied, vol. 11, pp. 064043, 2019, DOI: 10.1103 / PhysRev Applied. 11.064043. 9. X. networks”, Nature, vol. 589, pp. 44-51, 2021, DOI: 10.1038 / s41586-020-03063-0.

[0003] Identification of the above documents herein should not be construed as implying that they are in any way relevant to the patentability of the subject matter of the present disclosure.

[0004] Photonic computing holds the promise of enabling low-power, high-speed solutions for real-time machine learning and artificial intelligence applications, supporting a future scalable and sustainable computing ecosystem that is expected to grow exponentially over the next decade. To date, most proposed photonic computing solutions rely on photonic integrated circuit (PIC) technology, silicon photonic chips (SIPH), or free-space optics [1-3], using coherent interactions for multiply-accumulate (MAC) operations [4-6]. These technologies face several challenges, including yield and scaling limitations due to large chip size, significant cumulative losses due to the large number of Mach-Zehnder interferometers (MZIs) included in most designs, the tight phase control required, and sensitivity to local temperature and vibration.

[0005] Additionally, in neural networks utilizing multiple cascaded MZIs, the linear algebraic sum of a set of neuron inputs is realized by coherent field addition, which utilizes the phase of the optical carrier fields for code encoding. This coherent approach suffers from signal error accumulation along the MZI cascade, preventing it from achieving high accuracy and low bit error rates sufficient for large-scale practical applications.

[0006] Fiber-based neural networks are being developed that, although large in volume, are based on mature technologies with high bandwidth and low power specifications, providing off-the-shelf, readily available devices with proven reliability.

[0007] We have previously demonstrated an in-fiber-based optical computing unit that, when combined with standard devices such as transceivers and erbium-doped fiber amplifiers, realizes both the linear and nonlinear functions required for neural networks. While single-unit results suffer from coherence-induced phase noise, redundancy-assisted full-network emulation (ResNet-18) has demonstrated significantly better performance and accuracy than existing techniques [7, 8]. Various configurations of optical neural network units are also described, for example, in WO19186548 and WO21064727, both of which are assigned to the assignee of the present application. Summary of the Invention

[0008] There is a need in the art for novel approaches to the construction and operation of artificial neuron units, the building blocks of artificial neural networks that perform a variety of signal processing tasks.

[0009] In general, an artificial neural network (ANN) is a computational model inspired by the way biological nervous systems, such as the brain, process information. It consists of a large number of highly interconnected systems consisting of basic computational units or neurons. The artificial neurons are configured to process the input signals they receive and send corresponding signals to the artificial neurons connected to them. The artificial neurons are usually arranged in layers. Different layers can perform different types of transformations on their inputs and send corresponding output signals. A signal travels from the first (input) layer to the last (output) layer, possibly after passing through different layers several times.

[0010] As mentioned above, most known photonic computing solutions rely on PIC technology, SIPH, or free-space optics, using coherent interactions for multiply-accumulate (MAC) operations. This disclosure provides a novel approach for photonic computing systems that utilizes hybrid fiber technology and electro-optical communication devices, featuring negative and positive weighting schemes under incoherent data transmission conditions. We demonstrate that such a configuration can achieve speed improvements of 5-20 times while improving power efficiency by more than two orders of magnitude.

[0011] Specifically, the disclosed artificial neurons are realized as hybrid optical-electrical-optical (OEO) units, where the computation of the neural network is performed using incoherent light propagating through a fiber. Each neuron unit receives an input {x i} and output y j is light, and the optical signal, after being appropriately weighted during propagation in the fiber-based optical part of the neuron unit, undergoes linear mathematical operations while interacting with the electro-optical part of the neuron unit, which converts the optical signal into an electrical signal and then into an optical output. Note that working with incoherent optical signals rather than coherent light results in less "coherence noise" associated with interference effects and also more stable processing performance (over environmental conditions).

[0012] That is, according to one broad aspect of the present disclosure, there is provided an artificial neuron unit for processing a signal, the artificial neuron unit comprising: a fiber-based optical processing unit having first and second optical input ports and first and second optical output ports, the fiber-based optical processing unit configured and operable to controllably apply optical processing to the incoherent input optical signal to generate weighted first and second combined optical signals; an electro-optical processing unit configured and operable to process the weighted first and second combined optical signals, the electro-optical processing unit configured and operable to generate a weighted sum output of the artificial neuron unit by successively performing the steps of: applying a predetermined mathematical function to the weighted first and second combined optical signals corresponding to positive and negative weightings to generate a resultant signal; and applying non-linear processing to the resultant signal to convert it into an optical output signal indicative of the weighted sum output of the artificial neuron unit.

[0013] In some embodiments, the electro-optical processing unit has optical inputs coupled to the first and second optical output ports and includes a linear processor configured and operable to process the weighted first and second combined optical signals by applying a predetermined mathematical function to them, corresponding to positive and negative weightings, and output a resulting electrical signal; and a non-linear processor configured and operable to receive an input signal indicative of the resulting electrical signal, and to convert the input signal into an optical output signal indicative of the weighted sum output of the artificial neuron unit.

[0014] In some embodiments, the fiber-based optical processing unit has the following configuration: first and second splitters provided at first and second optical input ports, respectively, and first and second combiners provided at first and second optical output ports, respectively. The first splitter is configured to split the first optical input port into a first pair of distinct first and second optical propagation paths at a predetermined ratio, and the second splitter is configured to split the second optical input port into a second pair of distinct first and second optical propagation paths at a predetermined ratio, thereby generating the first and second pairs of optical propagation paths. The first combiner combines the first optical propagation paths of the first and second pair at the first optical output port to generate a first combined optical signal, and the second combiner combines the second optical propagation paths of the combined first and second pair at the second optical output port to generate a second combined optical signal. At least one of the first and second optical propagation paths of each of the first and second pairs is configured to apply a variable optical attenuation (VOA) to light propagating therethrough, whereby a weighting is applied to the incoherent input optical signal propagating through at least one of the first and second optical propagation paths of each of the first and second pairs, resulting in the first combined optical signal and the second combined optical signal being a weighted first and second combined optical signal, respectively.

[0015] In some embodiments, the linear processor includes a dual balanced photodiode configured and operable to process the weighted first and second combined optical signals and generate an electrical signal proportional to a difference between the weighted first and second combined optical signals, thereby achieving positive and negative weighting by the first and second combined optical signals, respectively.

[0016] In some embodiments, the nonlinear processor is configured as an electro-optical device, and the input signal received by the nonlinear processor is the resulting electrical signal output of the linear processor. For example, as described above, the linear processor includes a dual-balanced photodiode. In this case, the nonlinear processor can be realized based on the electronic nonlinearity of the photodiode, which can be adjusted depending on the desired nonlinear transformation to be performed. The nonlinear response of the circuit can be modulated / adjusted by changing the photodiode, changing the operating point of the photodiode, or subsequently changing other circuit elements (i.e., amplifiers).

[0017] In some other embodiments, the nonlinear processor is configured as an optically active device, and the input signal received by the nonlinear processor is an optical signal that corresponds to the resulting electrical signal output of the linear processor.

[0018] The artificial neuron unit may further include a control board configured and operable to control the operation of the fiber-based optical processing unit and the electro-optical processing unit.

[0019] In some embodiments, the control board includes a weighting controller configured and operative to generate control signals to the fiber-based optical processing unit to apply variable optical attenuation to light propagating through the optical processing unit; a linear controller configured and operative to prescribe coefficients of a mathematical function corresponding to a weighted sum of the weighted first and second combined optical signals; and a nonlinear controller configured and operative to prescribe the shape of a nonlinear function that converts the weighted sum into an optical output signal indicative of the weighted sum output of the artificial neuron unit.

[0020] According to another broad aspect of the present disclosure, there is provided an artificial neural network including two or more neuron layers arranged such that an optical input of a successive one of the two or more neuron layers is coupled to an optical output of a preceding one of the two or more neuron layers, each of the two or more neuron layers being formed by a number of independently operable artificial neuron units having the configuration described above. [Brief explanation of the drawings]

[0021] In order to better understand the subject matter disclosed herein and to illustrate how it may be carried out in practice, embodiments will now be described, by way of non-limiting example only, with reference to the accompanying drawings in which: [Figure 1] FIG. 1 shows a schematic functional scheme of an artificial neuron unit constructed in accordance with the present disclosure as a hybrid optical-electrical-optical processing unit. [Figure 2] Figure 2A shows a schematic diagram of the overall scheme of the disclosed artificial neuron unit, Figure 2B shows a schematic diagram of the configuration of the disclosed artificial neuron unit, specifically showing the configuration of a fiber-based optical processing unit, and Figure 2C shows a schematic diagram of a non-limiting embodiment of the disclosed artificial neuron unit. [Figure 3] Figure 3A shows a schematic diagram of the system architecture of an artificial neural network according to the present disclosure, and Figure 3B shows a photograph of the assembled system. [Figure 4]Figures 4A-4C show the results of timing characterization of the positive and negative outputs. Figure 4A shows a 20 ns pulse at 1550 nm injected into a single neuron input. Figure 4B shows the same pulse with a 40 cm delay line (equivalent to 2 ns) added to the positive path. Figure 4C shows the results when a four-level step input signal is injected into two input ports (input 1 with a step period of 10 ns and input 2 with a step period of 40 ns). The plots show the output when both VOAs are fully closed, fully open, and when one VOA is closed and the other is open. Figure 4D shows the measured output values ​​and the expected value. DETAILED DESCRIPTION OF THE INVENTION

[0022] 1, there is shown a schematic scheme of an exemplary artificial neuron unit (neuron) 1. Such a neuron 1 can be used as an individual node of an artificial neural network 100.

[0023] Neural network 100 typically includes two or more layers operable in a cascaded fashion, with each layer including a number of neurons 1 operable independently. A single layer is shown in the figure. As shown, neurons 1 are configured and operable as follows: they receive corresponding input signals (x1, x2, ... x1) from, for example, neurons in a previous layer (not shown). i ) by synapse S. 1j ,w 2j ,...w ij ) and then summed to produce a linear neuron signal, which is then transformed nonlinearly (by a nonlinear transfer (activation) function) to produce a single neuron output (y j) is generated. The weights added by the synapses S can be positive or negative, with positive weights activating the neuron and negative weights inhibiting it. Nonlinear functions applied to linear neuronal signals include logistic functions (sigmoids), rectified linear units (ReLUs), and inverse square root linear units (ISRUs), depending on the neural model used.

[0024] It should be noted that interactions between incoherent optical signals input to neurons do not affect their phase and therefore do not by themselves enable negative weighting. Therefore, while it is difficult to implement negative weighting using neurons operating with incoherent signals, the use of incoherent signals is important and suitable for various applications. In particular, as mentioned above, operating with incoherent optical signals reduces "coherence noise" and provides more stable processing performance.

[0025] Referring to FIG. 2A, there is shown a schematic diagram of the structure and operation of an artificial neuron unit 10 according to the present disclosure, configured to operate with incoherent input signals and capable of providing appropriate positive and negative weights.

[0026] The inventors were inspired by the push-pull mechanism, which describes the interplay between excitation and inhibition during neural processing throughout the central nervous system. Briefly, when inhibition is coupled with excitation in a push-pull manner, where inhibition decreases as excitation increases, neuronal excitability can be increased. The inventors used the principles of such a push-pull mechanism to achieve positive and negative weighting.

[0027] The artificial neuron unit 10 of the present disclosure is configured as a hybrid system of a fiber-based optical processing unit 12 and an electro-optical processing unit 16. The fiber-based optical processing unit 12 (fiber configuration) has first and second optical input ports 14A, 14B for receiving first and second incoherent input signals Lin1, Lin2, respectively, and has first and second optical output ports 15A, 15B. The fiber-based optical processing unit 12 controllably performs optical processing on the incoherent input optical signals to generate first and second weighted combined optical signals (L (com) 1) w and (L (com) 2) w Each of these weighted combined optical signals is formed by combining weighted portions of the first and second incoherent input signals Lin1, Lin2, respectively.

[0028] The electro-optical processing unit 16 includes a linear processor 16A and a non-linear processor 16B. The linear processor 16A has an optical input OI coupled to the first and second optical output ports 15A, 15B of the optical unit 12, and generates first and second weighted combined optical signals (L,L) by applying predetermined mathematical functions thereto corresponding to positive and negative weightings. (com) 1) w , (L (com) 2) w to generate an electrical signal ES that is output via an electrical output port EO of the linear processor 16A.

[0029] The nonlinear processor 16B can be configured as an electro-optical device, for example, the nonlinearity can be realized based on the electronic nonlinearity of the photodiode of a linear processor (i.e., the photodiode response has a linear range and then saturates when the electrical capacitor is full). Adjusting the nonlinearity of the photodiode can be realized by changing the type of photodiode or by changing one or more of the circuit voltages to change the operating point.

[0030] Alternatively, the nonlinear processor 16B can be configured as an optically active device, such as an erbium-doped fiber amplifier (EDFA) or semiconductor optical amplifier (SOA), in which one or more input signals compete for gain resources (cross-gain modulation) or compete via a nonlinear process such as the Kerr effect (cross-phase modulation). Such optical nonlinear processors / operators are known and are described, for example, in WO2021064727 and US2022327372 (assigned to the assignee of the present application), which are incorporated herein by reference.

[0031] Thus, in general, the nonlinear processor 16B is configured to receive an input signal indicative of the electrical signal output by the linear processor 16A, and convert this input signal into an optical signal L indicative of the weighted sum output of the artificial neuron unit 10. out This optical signal L is configured and operable to be converted into out is allowed to propagate in the output optical fiber 18, which is for example the input fiber of the neuronal unit of the next layer.

[0032] Considering the electro-optical implementation of nonlinear processor 16B, the nonlinear processor is directly coupled to the electrical output EO of linear processor 16A and converts the electrical signal ES output from processor 16A into an optical output signal L propagating in output optical fiber 18. outAlthough not specifically shown in the figure, in the case of an optical implementation of the nonlinear processor 16B, it should be noted that the neuron unit 10 also includes an electro-optical converter of any known suitable configuration, such as a laser diode with a low coherence length, arranged upstream of the nonlinear processor 16B.

[0033] The neuron unit 10 is associated with (e.g., includes) a control board 20. The control board 20 includes a weighting controller 20A and two power controllers 20B, 20C. The weighting controller 20A is configured and operable to apply a control signal CS1 to the optical processing unit 12, for example, to cause variable optical attenuation of light propagating through the optical processing unit 12, as described below. The power controllers 20B, 20C are configured and operable to control the linear processor 16A and the nonlinear processor 16B, respectively. More specifically, the controller 20B is configured and operable to define coefficients of the weighted sum and generate corresponding operating data / signal CS2 for the linear processor 16A, and the controller 20C is configured and operable to apply the weighted sum of the inputs embedded in the electrical signal ES output from the linear processor 16B to the optical output signal L. out and generates corresponding control data / signal CS3 to operate the non-linear processor 16B.

[0034] 2B shows a schematic diagram of the neuron unit 10 of the present disclosure and details the operation of the fiber-based optical processing unit 12. The same reference numerals are used to identify functionally common components in all examples.

[0035] The neuron unit 10 of FIG. 2B is configured generally similarly to that of FIG. 2A, i.e., as a hybrid system of a fiber-based optical processing unit 12 and an electro-optical processing unit 16, and is associated with a control board 20. The fiber-based optical processing unit 12 has first and second optical input ports 14A, 14B for receiving an incoherent input signal, and first and second optical output ports 15A, 15B optically coupled to optical inputs of a linear processor 16A, which outputs first and second weighted combined optical signals (L (com) 1) w , (L (com) 2) w and applies to it a preset mathematical function corresponding to positive and negative weighting. The electrical signal ES thus generated is converted into an optical output signal L representing the weighted sum output of the artificial neuron unit 10. out is received and processed by a non-linear processor 16B which converts it into

[0036] As shown in FIG. 2B, the optical input signals at the optical input ports 14A, 14B of the optical processing unit 12 are x (1) and x (2) The fiber-based optical processing unit 12 includes splitters 24A and 24B at the input ports 14A and 14B, respectively. The splitter 24A splits the input optical field Lin1 into two respective fibers F1 A , F2 A Two light parts Lin1 propagating along p , Lin1 n Similarly, splitter 24A is configured to split the input optical field Lin2 into two respective fibers F1 and F2 in a preset ratio. B , F2 B Two light parts Lin2 propagating along P , Lin2 nThus, each of the first and second optical input signals entering processor 12 via input ports 14A, 14B is split in a predetermined ratio onto a pair of separate first and second optical propagation paths.

[0037] Thus, the two arms of the fiber-based processor associated with the two inputs 14A, 14B receive the respective optical signals x p (1) and x n (1) The light portion Lin1 of the input light Lin1 having p , Lin1 n Light propagation path F1 A , F2 A and the first pair of optical signals x p (2) and x n (2) The light portion Lin2 of the input light Lin2 having P , Lin2 n Light propagation path F1 B , F2 B and a second pair of.

[0038] First and second pair F1 A -F2 A and F1 B -F2 B At least one of the first and second light propagation paths of each of the optical attenuators (VOA) 22 is configured to apply variable optical attenuation (VOA) to the light propagating therethrough, thereby providing a weighting (e.g., w p (1) , w n (1) , w p (2) , w n (2) ) are the first fiber F1 A , F1 B and / or the second fiber F2 A , F2 B The incoherent input light portion Lin1 propagates through p and Lin2 Pand / or input optical part Lin1 n and / or Lin2 n applies to.

[0039] Furthermore, the fiber-based optical processor 12 is provided with optical combiners 26A and 26B, and the combiner 26A is connected to the propagation path F1. A , F1 B , so the light part Lin1 p , Lin2 p is coupled to output port 15A, and combiner 26B couples propagation path F2 A , F2 B , so the light part Lin1 n , Lin2 n As a result, the output optical signal propagating through output ports 15A and 15B is a weighted combined signal (L (com) 1) p , (L (com) 2) n More specifically, the first light propagation path F1 of the first and second pair A , F1 B is connected to the first optical output port 15A and combined thereto to generate a first combined optical signal (L (com) 1) p (e.g., w p (1) x p (1) +w p (2) x p (2) ) is generated, while the first and second pair of propagation paths F2 A , F2 B is connected to the second optical output port 15B and combined thereto to produce a second combined optical signal (L (com) 2) n (e.g., w n (1) x n (1) +w n (2) x n (2) ) is generated.

[0040] The combined optical signal (L (com) 1) p , (L (com) 2) n is fed to a linear processor 16A to produce a combined optical signal (L (com) 1) p , (L (com) 2) n In this non-limiting example, a linear processor may generate an output electrical signal ES (e.g., w ) that is proportional to the difference between the weighted first and second combined optical signals. p (1) x p (1) +w p (2) x p (2) -(w n (1) x n (1) +w n (2) x n (2) )) and thus the first and second combined optical signals (L (com) 1) p , (L (com) 2) n Positive and negative weighting is realized by the following.

[0041] Note that the subscripts "p" and "n" used herein refer to positive and negative weights, respectively. In this non-limiting example, positive and negative weights are assigned to the first propagation path F1A-F1B and the second propagation path F2A-F2B of a pair, respectively. However, it should be understood that this can be defined the other way around. It should also be understood that the negative and positive weights are actually realized through the interaction of the combined optical signal with the linear processor 16A, and the magnitude of the weights is determined by the controlled operation of one or more VOAs 22. Therefore, the combined optical signal output via output ports 15A and 15B is referred to herein as a "weighted" combined optical signal.

[0042] It is known in the art that linearity alone is not enough in neural networks, and that nonlinear activation functions are necessary, similar to the function of synapses in the brain nervous system. Nonlinear functions are necessary to speed up the convergence speed of the network and improve recognition accuracy, and are an essential part of neural networks. Nonlinearities vary depending on the neural model used, from simple sigmoids to complex dynamical systems.

[0043] As mentioned above, the nonlinear function can be realized electronically or optically by the nonlinear calculator unit 16B. For example, the electrical signal ES from the dual balanced photodiode 16A can be optically pumped (from the controller 20C of the control board 20) to drive the electro-optic modulator 16B to generate a nonlinearly converted optical signal. In another non-limiting example, optical nonlinearity can be implemented using an optically active device such as an EDFA, in which case a laser diode (electron-to-photon converter) is placed upstream of the nonlinear module with respect to the normal direction of signal propagation through the neuron unit.

[0044] Therefore, the resulting analog optical signal L out denotes the weighted sum optical output that propagates down the output fiber 18 of the artificial neuron unit 10 and inputs, for example, to the next layer of the network.

[0045] As mentioned above, and as shown schematically in the drawings, the artificial neuron unit 10 is associated with a control board 20 configured and operable to control the electro-optical components of the system, such as one or more VOAs 22, dual balanced photodiode 16A, and nonlinear operator unit 16B.

[0046] Below we provide more concrete examples of the system architecture and neuron performance, and estimate the performance of the complete neural network based on the measurements.

[0047] 2C, the operation of an exemplary single neuron 10 is shown schematically. The neuron unit 10 includes an optical processing unit 12 having two inputs 14A, 14B for receiving incoherent input optical signals Lin1, Lin2. As mentioned above, the neuron 10 employs a push-pull mechanism to achieve positive and negative weighting. Each of the first and second input optical signals Lin1, Lin2 is split by a respective fiber-coupled splitter 24A, 24B in a predetermined ratio (e.g., a 30 / 70 ratio) to produce an optical portion Lin1. p -Lin1 n , Lin2 P -Lin2 n For each pair of light propagation paths F1A-F1B and F2A-F2B, a pair of light propagation paths F1A-F1B and F2A-F2B is provided. In this non-limiting example, the light portion (Lin1) passing through 70% of each light propagation path (F1A or F1B) acts as a positive weight. p or Lin2 p ) passes through variable optical attenuators (VOAs) 22, which are individually controlled by the control board 20, and a light portion (Lin1) passing through 30% of each light propagation path (F2A or F2B) is n or Lin2 n ) acts as a negative weight and is not attenuated. The positive weight sections F1A, F1B are coupled to the positive weight output port 15A by the fiber coupling combiner 26A to produce a positive weighted combined optical signal (L (com) 1) P and the negative weighted sections F2A, F2B are coupled to the negative weighted output port 15B by the fiber coupled combiner 26B to generate a negative weighted combined optical signal (L (com) 2) n These positive and negative weighted combined optical signals (L (com) 1) P , (L (com) 1) n is guided to the optical input of the dual balanced photodiode 16A (linear processor) to generate an electrical signal ES corresponding to the differential output voltage, which is then converted by the nonlinear calculator unit 16B into the sum output L of the neuron 10.OUT is converted into an analog optical signal that indicates

[0048] In addition to providing negative weights, as described above, the artificial neuron unit 10 is based on a push-pull mechanism of control. If we consider the first input 14A as "excitatory" and the second input 14B as "inhibitory," it is easy to show that by increasing inhibition, we can achieve an increase in sum output, i.e., an increase in the excitability of the neuron. This illustrates the paradox at the heart of the push-pull configuration: by increasing background inhibition and increasing disinhibition, we can achieve an increase in force output. Such a configuration may be advantageous in designing more robust multilayer artificial neural networks.

[0049] FIG. 3A shows a schematic diagram of a complete photonic / computing system architecture 100 configured as a multi-layer neural network, with each layer formed by an array of independently operable neuron units 10 configured according to the present disclosure as described above. Analog electrical signals are generated by a control board 20 and converted to analog optical signals using a modulator. These analog optical signals are appropriately weighted, biased, and injected into the first neural layer. Each neuron unit 10 is configured and operable as described above to perform an analog nonlinear function on the weighted sum of its inputs and send the resulting amplitude to the next layer. The outputs of the output layer neurons are read by the control board 20 via photodiodes.

[0050] FIG. 3B shows a photograph of the assembled exemplary system.

[0051] We first characterized the timing of the positive and negative weights. To that end, we inserted a 20-ns square pulse at 1500 nm into the neuron's single input. The output was measured with an oscilloscope (Keysight MXR604A), as shown in Figure 4A. Next, to display the negative and positive weights separately, we added a 40-cm delay line, equivalent to a 2-ns delay, to the positive weight's optical path, as shown in Figure 4B. This graph shows that the negative weight precedes the positive weight, and the added spike is indeed 2 ns long, as expected.

[0052] Next, we inserted a four-level analog step function into the neuron's two inputs (input 1 with a step period of 10 ns and input 2 with a step period of 40 ns). We recorded the output for various VOA states, as shown in Figure 4C. When both VOAs are closed, the positive weight is equal to zero, resulting in a negative downward step. When both VOAs are fully open, the intensity of the combined positive input is much higher than the intensity of the combined negative input, resulting in an upward step. The intermediate state is when one VOA is open and the other is closed. Because the VOA for the short-period input is open, that input has a positive weight, thus indicating a rising output. Because the VOA for the long-period input is closed, it has a negative weight, thus indicating a downward step.

[0053] To evaluate the accuracy of the MAC calculation shown in Figure 4D, we compared the output values ​​measured by the dual balanced photodiode with the expected output values. For an ideal MAC calculator, we expect the plot to be linear. The plot in Figure 4D shows the R ∧ 2 = 0.9995, showing a linear approximation, suggesting that the MAC accuracy of the neuron is excellent.

[0054] The inventors completed a prototype system with 16 input channels and a four-layer classifier and tested the performance of this system. They compared the performance of the prototype system with the industry standard Nvidia DGX A100 and another photonic accelerator, the LightMatter Envise server. The results showed up to 20 times faster speeds than competing systems and two orders of magnitude better power efficiency.

[0055] Thus, the present disclosure provides a neuron unit based on hybrid fiber technology and electro-optical communication devices, featuring positive and negative weighting schemes under incoherent data transmission conditions, and a photonic computing system utilizing such a neuron unit. The inventors have shown that such a design can achieve 5-20 times faster speeds while improving power efficiency by over 100 times.

Claims

1. 1. An artificial neuron unit for processing a signal, comprising: a fiber-based optical processing unit having first and second optical input ports and first and second optical output ports, the fiber-based optical processing unit configured and operable to controllably apply optical processing to the incoherent input optical signal to generate first and second weighted combined optical signals; an electro-optical processing unit configured and operable to process the weighted first and second combined optical signals, the electro-optical processing unit configured and operable to generate a weighted sum output of the artificial neuron unit by successively performing the steps of: applying a predetermined mathematical function to the weighted first and second combined optical signals corresponding to positive and negative weightings to generate a resultant signal; and applying non-linear processing to the resultant signal to convert it into an optical output signal indicative of a weighted sum output of the artificial neuron unit. An artificial neuron unit comprising:

2. 2. The artificial neuron unit according to claim 1, the electro-optical processing unit having optical inputs coupled to the first and second optical output ports, the electro-optical processing unit comprising: a linear processor configured and operative to process the weighted first and second combined optical signals corresponding to positive and negative weightings by applying the predetermined mathematical function to them and to output a resultant electrical signal; and a non-linear processor configured and operative to receive an input signal indicative of the resultant electrical signal, and to convert the input signal into an optical output signal indicative of a weighted sum output of the artificial neuron unit.

3. 2. The artificial neuron unit according to claim 1, the fiber-based optical processing unit comprises first and second splitters provided at the first and second optical input ports and first and second combiners provided at the first and second optical output ports, the first splitter configured to split the first optical input port into a first pair of distinct first and second optical propagation paths in a preset ratio, the second splitter configured to split the second optical input port into a second pair of distinct first and second optical propagation paths in the preset ratio, the first combiner combining the first and second pair of first optical propagation paths at the first optical output port to generate a first combined optical signal, and the second combiner combining the first and second pair of second optical propagation paths combined at the second optical output port to generate a second combined optical signal; at least one of the first and second optical propagation paths of each of the first and second pairs is configured to apply a variable optical attenuation (VOA) to light propagating therethrough, thereby applying a weighting to the incoherent input optical signals propagating through at least one of the first and second optical propagation paths of each of the first and second pairs, such that the first combined optical signal and the second combined optical signal are the weighted first and second combined optical signals, respectively.

4. The artificial neuron unit according to claim 2, An artificial neuron unit, wherein the nonlinear processor is configured as an electro-optical device.

5. The artificial neuron unit according to claim 4, the linear processor includes a dual balanced photodiode configured and operable to process the weighted first and second combined optical signals and generate an electrical signal proportional to a difference between the weighted first and second combined optical signals, thereby achieving positive and negative weighting by the first and second combined optical signals, respectively.

6. The artificial neuron unit according to claim 5, The artificial neuron unit is characterized in that the nonlinear processor is configured as an electro-optical device based on the electronic nonlinearity of the dual balanced photodiode.

7. The artificial neuron unit according to claim 2, 1. An artificial neuron unit, wherein the nonlinear processor is configured as an optically active device, and the input signal received by the nonlinear processor is an optical signal corresponding to the resulting electrical signal output of the linear processor.

8. 2. The artificial neuron unit according to claim 1, 10. An artificial neuron unit, further comprising: a control board configured and operable to control the operation of the fiber-based optical processing unit and the electro-optical processing unit.

9. The artificial neuron unit according to claim 8, the control board includes: a weighting controller configured and operative to generate control signals to the fiber-based optical processing unit to apply variable optical attenuation to light propagating through the optical processing unit; a linear controller configured and operative to define coefficients of the mathematical function corresponding to a weighted sum of the weighted first and second combined optical signals; and a nonlinear controller configured and operative to define the shape of a nonlinear function that converts the weighted sum into an optical output signal indicative of a weighted sum output of the artificial neuron unit.

10. The artificial neuron unit according to claim 2, the fiber-based optical processing unit comprises first and second splitters provided at the first and second optical input ports and first and second combiners provided at the first and second optical output ports, the first splitter configured to split the first optical input port into a first pair of distinct first and second optical propagation paths in a preset ratio, the second splitter configured to split the second optical input port into a second pair of distinct first and second optical propagation paths in the preset ratio, the first combiner combining the first and second pair of first optical propagation paths at the first optical output port to generate a first combined optical signal, and the second combiner combining the first and second pair of second optical propagation paths combined at the second optical output port to generate a second combined optical signal; at least one of the first and second optical propagation paths of each of the first and second pairs is configured to apply a variable optical attenuation (VOA) to light propagating therethrough, thereby applying a weighting to the incoherent input optical signals propagating through at least one of the first and second optical propagation paths of each of the first and second pairs, such that the first combined optical signal and the second combined optical signal are the weighted first and second combined optical signals, respectively.

11. The artificial neuron unit according to claim 2, An artificial neuron unit, wherein the nonlinear processor is configured as an electro-optical device.

12. The artificial neuron unit according to claim 11, the linear processor includes a dual balanced photodiode configured and operable to process the weighted first and second combined optical signals and generate an electrical signal proportional to a difference between the weighted first and second combined optical signals, thereby achieving positive and negative weighting by the first and second combined optical signals, respectively.

13. The artificial neuron unit according to claim 12, The artificial neuron unit is characterized in that the nonlinear processor is configured as an electro-optical device based on the electronic nonlinearity of the dual balanced photodiode.

14. The artificial neuron unit according to claim 2, 1. An artificial neuron unit, wherein the nonlinear processor is configured as an optically active device, and the input signal received by the nonlinear processor is an optical signal corresponding to the resulting electrical signal output of the linear processor.

15. The artificial neuron unit according to claim 2, 10. An artificial neuron unit, further comprising: a control board configured and operable to control the operation of the fiber-based optical processing unit and the electro-optical processing unit.

16. The artificial neuron unit according to claim 15, the control board includes: a weighting controller configured and operative to generate control signals to the fiber-based optical processing unit to apply variable optical attenuation to light propagating through the optical processing unit; a linear controller configured and operative to define coefficients of the mathematical function corresponding to a weighted sum of the weighted first and second combined optical signals; and a nonlinear controller configured and operative to define the shape of a nonlinear function that converts the weighted sum into an optical output signal indicative of a weighted sum output of the artificial neuron unit.

17. 10. An artificial neural network comprising two or more neuron layers arranged such that an optical input of a successive one of the two or more neuron layers is coupled to an optical output of a preceding one of the two or more neuron layers, each of the two or more neuron layers being formed by a plurality of independently operable artificial neuron units, each artificial neuron unit being constructed in accordance with claim 1.

18. 1. An artificial neural network comprising two or more neuron layers arranged such that an optical input of a successive one of the two or more neuron layers is coupled to an optical output of a preceding one of the two or more neuron layers, each of the two or more neuron layers being formed by a plurality of independently operable artificial neuron units, each artificial neuron unit being constructed in accordance with claim 2.

Images