Optical fiber-based artificial neuron unit

Abstract

An artificial neural unit for signal processing is presented. The artificial neuron unit comprises: an optical fiber-based optical processing unit having a first optical input port and a second optical input port and a first optical output port and a second optical output port, the optical processor is arranged and operable to controllably apply optical processing to the incoherent input optical signal and produce weighted first and second combined optical signals; and an electro-optical processing unit, and an electro-optical processing unit configured and operable to process the weighted first combined optical signal and the second combined optical signal and to produce a total weighted output of the artificial neuron unit by continuously performing: applying a predetermined mathematical function to the weighted first combined optical signal and the second combined optical signal, to correspond to the positive weighting and the negative weighting and generate a result signal; and applying a non-linear process to the resultant signal to transform the resultant signal into an optical output signal representative of the total weighted output of the artificial neuronal unit.

Description

Technical Field

[0001] The present disclosure belongs to the field of artificial neural networks and relates to an optical fiber-based neuron unit and a neural network using the neuron unit. Background Art

[0002] Listed below are references believed to be relevant as background to the presently disclosed subject matter: 1. R. Xu, P. Lv, F. Xu, and Y. Shi, “A survey of approaches for implementing optical neural networks,” Optics & Laser Technology, vol. 136, p. 106787, 2021, DOI: 10.1016 / j.optlastec.2020.106787.

[0003] 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.

[0004] 3. B. Shi, N. Calabretta, and R. Stabile, “Deep Neural Network Through an InPSOA-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.

[0005] 4. A. Totovic, G. Giamougiannis, A. Tsakyridis, D. Lazovsky, and N. Pleros, “Programmable photonic neural networks combining WDM with coherent linear optics,” Scientific Reports, vol. 12, p. 5605, 2022, DOI: 10.1038 / s41598-022-09370-y.

[0006] 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.

[0007] 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.

[0008] 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, p. 29080, 2016, DOI: 10.1038 / srep29080.

[0009] 8. A. N. Tait, T. Ferreira de Lima, M. A. Nahmias, H. B. Miller, H. T. Peng, B. J. Shastri, and P. R. Prucnal, ""Silicon Photonic Modulator Neuron"", Physical Review Applied, Vol. 11, p. 064043, 2019, DOI: 10.1103 / PhysRevApplied.11.064043.

[0010] 9. X. Xu, M. Tan, B. Corcoran, J. Wu, A. Boes, T. G. N. Guyen, S. T. Chu, B. E. Little, D. G. Hicks, R. Morandotti, A. Mitchell, and DJ Moss, “11 TOPS photonic convolutional accelerator for optical neural networks,” Nature, vol. 589, pp. 44–51, 2021, DOI: 10.1038 / s41586-020-03063-0.

[0011] Acknowledgment of the above references herein should not be inferred as an intention that these references are in any way relevant to the patentability of the presently disclosed subject matter. background

[0012] Photonic computing promises to enable low-power and 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. Most photonic computing solutions proposed to date rely on photonic integrated circuit (PIC) technology, silicon photonic chips (SIPH), or free-space optics [1-3] and use coherent interactions for multiply-accumulate (MAC) operations [4-6]. These technologies contain several issues, including yield and scale limitations due to large chip size, large accumulated losses of the numerous Mach-Zender interferometers (MZIs) included in most designs, the required tight phase control, and high sensitivity to local temperature or vibration.

[0013] Furthermore, in neural networks utilizing multiple cascaded MZIs, linear algebraic summation over a set of neuron inputs is achieved by coherent electric field addition using the phase of the optical carrier electric field for symbol encoding purposes. This coherent approach results in accumulation of signal errors along the cascade of MZIs and is unable to achieve sufficiently high accuracy and sufficiently low bit error rates for large-scale practical applications.

[0014] Fiber-based neural networks have been developed. These technologies offer common devices that are bulkier but based on mature technologies that have high bandwidth and low power specifications, as well as ready availability and proven reliability.

[0015] The inventors have previously demonstrated optical computing units based on optical fibers, which, combined with standard equipment such as transceivers and erbium-doped fiber amplifiers, deliver the linear and nonlinear functions required by neural networks. Although the results of a single unit are affected by coherence-induced phase noise, redundancy-assisted full network simulations (ResNet-18) demonstrate performance and accuracy far superior to the prior art [7,8]. In addition, various configurations of optical neural network units are described, for example, in WO19186548 and WO21064727, assigned to the assignee of the present application. SUMMARY OF THE INVENTION

[0016] There is a need in the art for a novel method for the configuration and operation of artificial neuron units, which are the basic blocks in artificial neural networks that perform various signal processing tasks.

[0017] 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 made up of basic computational units or neurons. Artificial neurons are configured to process the input signals they are receiving and then transmit the corresponding signals to the (multiple) artificial neurons connected to them. Typically, artificial neurons are arranged in layers. Different layers can perform different kinds of transformations on their inputs and transmit corresponding output signals. Signals travel from the first layer (input layer) to the last layer (output layer), and may need to pass through different layers several times.

[0018] As mentioned above, most known photonic computing solutions rely on PIC technology, SIPH or free-space optics and use coherent interactions to perform multiply-accumulate (MAC) operations. The present disclosure provides a novel approach for photonic computing systems that utilizes hybrid fiber technology and electro-optical communication devices, characterized by negative and positive weighting schemes under incoherent data transmission conditions. The inventors have shown that this configuration enables 5-20 times acceleration while improving power efficiency by two or more orders of magnitude.

[0019] Specifically, the artificial neurons of the present disclosure are implemented as hybrid optical-electrical-optical (OEO) units, where neural network computations are performed on incoherent light propagating in an optical fiber. and output y jis optical, and the optical signal, after being appropriately weighted during propagation in the fiber-based optical portion of the neuron unit, undergoes a linear mathematical operation when interacting with the electro-optical portion of the neuron unit, which transforms the optical signal into an electronic signal, which then enters the optical output. It should be noted that operating with incoherent optical signals rather than coherent light produces less "coherent noise" associated with interference effects, and also provides more stable processing performance (compared to ambient conditions).

[0020] Therefore, according to a broad aspect of the present disclosure, there is provided an artificial neural unit for signal processing, the artificial neural unit comprising: an optical fiber-based optical processing unit having first and second optical input ports and first and second optical output ports and configured and operable to controllably apply optical processing to the incoherent input optical signal and produce weighted first and second combined optical signals; and an electro-optical processing unit configured and operable to process the weighted first and second combined optical signals and produce a total weighted output of the artificial neuron unit by successively performing the following operations: applying a predetermined mathematical function to the weighted first and second combined optical signals to correspond to positive and negative weightings and produce a result signal; and applying non-linear processing to the result signal to convert the result signal into an optical output signal representing the total weighted output of the artificial neuron unit.

[0021] In some implementations, the electro-optical processing unit includes: a linear processor having an optical input coupled to the first optical output port and the second optical output port, and configured and operable to process the weighted first combined optical signal and the second combined optical signal by applying a predetermined mathematical function to the weighted first combined optical signal and the second combined optical signal to correspond to positive weighting and negative weighting and output a resulting electronic signal; and a nonlinear processor adapted to receive an input signal indicative of the resulting electronic signal, and configured and operable to convert the input signal into an optical output signal representing a total weighted output of the artificial neuron unit.

[0022] In some embodiments, the optical fiber-based optical processing unit has the following configuration: the optical fiber-based optical processing unit includes first and second splitters at first and second optical input ports, respectively, and first and second combiners at first and second optical output ports, respectively. The first splitter is configured to divide the first optical input port into a first pair of separated first and second optical propagation paths at a predetermined ratio, and the second splitter is configured to divide the second optical input port into a second pair of separated first and second optical propagation paths at the predetermined ratio, thereby generating a first pair of optical propagation paths and a second pair of optical propagation paths. The first combiner combines the first optical propagation paths in the first pair and the 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 in the first pair and the 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. Thus, weighting is applied to the incoherent input optical signal propagated through the corresponding at least one of the first optical propagation path and the second optical propagation path of each of the first pair and the second pair, so that the first combined optical signal and the second combined optical signal are thereby the weighted first combined optical signal and the second combined optical signal, respectively.

[0023] In some embodiments, the linear processor includes a dual balanced photodiode configured and operable to process the weighted first combined optical signal and the second combined optical signal and generate an electronic signal proportional to the difference between the weighted first combined optical signal and the second combined optical signal, thereby achieving positive weighting and negative weighting of the first combined optical signal and the second combined optical signal, respectively.

[0024] 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 electronic signal output of the linear processor. For example, as described above, the linear processor includes a double-balanced photodiode. In this case, the nonlinear processor can be implemented based on the electronic nonlinearity of the photodiode, which can be adjusted according to the desired nonlinear translation 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 any other circuit element following it (i.e., amplifier).

[0025] In some other embodiments, the non-linear processor is configured as an optically active device, and the input signal received by the non-linear processor is an optical signal corresponding to the resulting electronic signal output of the linear processor.

[0026] The artificial neural 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.

[0027] In some embodiments, the control board includes: a weighted controller configured and operable to generate a control signal to an optical fiber-based optical processing unit to apply variable optical attenuation to light propagating through the optical processing unit; a linear controller configured and operable to define coefficients of the mathematical function corresponding to a weighted sum of the weighted first combined optical signal and the second combined optical signal; and a nonlinear controller configured and operable to define the shape of the nonlinear function that transforms the weighted sum into an optical output signal representing a total weighted output of the artificial neuron unit.

[0028] According to another broad aspect of the present disclosure, there is provided an artificial neural network comprising two or more neuron layers, the two or more neuron layers being arranged such that an optical input of a subsequent layer in the two or more neuron layers is coupled to an optical output of a previous layer in the two or more neuron layers. Each of the two or more neuron layers is formed by a plurality of independently operable artificial neuron units having the above configuration. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] For a better understanding of the subject matter disclosed herein and in order to illustrate how the same may be implemented in practice, embodiments will now be described, by way of non-limiting examples only, with reference to the accompanying drawings, in which: Figure 1 schematically illustrates a functional scheme of an artificial neuron unit configured as a hybrid optical-electrical-optical processing unit according to the present disclosure; Figure 2A Schematically shows the general scheme of the artificial neuron unit of the present disclosure; Figure 2B Schematically illustrates the configuration of the artificial neuron unit of the present disclosure, and specifically illustrates the configuration of the optical processing unit based on optical fiber; Figure 2C Schematically illustrates a non-limiting embodiment of an artificial neuron unit of the present disclosure; Figure 3A Schematically illustrates the system architecture of an artificial neural network according to the present disclosure; Figure 3B A photograph of the assembled system is shown; FIG. 4A to FIG. 4C The characterization results of the timing of positive and negative outputs are shown, where Figure 4A A 20 ns pulse at 1550 nm injected into the input of a single neuron is shown; Figure 4B The same pulse is shown with a 40cm delay line (equivalent to 2ns) added to the positive optical path; and Figure 4C shows the results when a 4-level step input signal is injected into the 2 input ports: the step period is 10ns for input 1 and 40ns for input 2. The figure depicts the output when both VOAs are fully closed, fully open, and when one VOA is closed and the other open; and

[0030] Figure 4D The measured output values ​​are shown in relation to the expected values. DETAILED DESCRIPTION

[0031] refer to Figure 1 A general scheme of an exemplary artificial neuron unit (neuron) 1 is described. Such a neuron 1 can be used as an individual node of an artificial neural network 100.

[0032] The neural network 100 typically includes two or more layers that can be operated in a cascaded manner, wherein each layer includes a plurality of neurons 1 that can be operated individually. In this figure, a single layer is shown. As shown in the figure, the neuron 1 is configured and operable as follows: the neuron 1 includes two or more inputs (three such inputs are shown in the figure) for receiving corresponding input signals , for example, the input signal comes from a neuron in the previous layer (not shown). The input signal received by neuron 1 is composed of synapses S (with weights ) are linearly weighted and then summed to produce a linear neuron signal, which is nonlinearly transformed (via a nonlinear transfer (activation) function) to generate a single neuron output (y j ). The weight applied by the synapse S can have positive or negative values, where a positive weight activates the neuron and a negative weight inhibits the neuron. Depending on the neural model used, the nonlinear function applied to the linear neuron signal can include a logistic function (sigmoid), a rectified linear unit (ReLU), an inverse square root linear unit (ISRU), etc.

[0033] It should be noted that the interaction between incoherent optical signals input to the neuron does not allow negative weighting by itself, since such interaction does not affect the phase. Therefore, it is challenging to achieve negative weighting by neurons operating with incoherent signals, while the use of incoherent signals is important and preferred for various applications. In particular, as described above, operation with incoherent optical signals produces less "coherent noise" and provides more stable processing performance.

[0034] refer to Figure 2A, schematically illustrates the configuration and operation of an artificial neuron unit 10 according to the present disclosure, which is configured to operate on an incoherent input signal and is capable of correctly providing positive and negative weights.

[0035] The inventors were inspired by the push-pull mechanism, which describes the interaction between excitation and inhibition during neural processing throughout the central nervous system. In short, when inhibition is coupled with excitation in a push-pull manner, neuronal excitability can increase while inhibition decreases as excitation increases. The inventors used the principle of this push-pull mechanism to implement positive and negative weighting.

[0036] The artificial neuron unit 10 of the present disclosure is configured as a hybrid system having an optical processing unit 12 based on an optical fiber and an electro-optical processing unit 16. The optical processing unit 12 based on an optical fiber (optical fiber device) has a first optical input port 14A and a second optical input port 14B for inputting a corresponding first incoherent input signal Lin 1 and the second incoherent input signal Lin 2 , and having a first optical output port 15A and a second optical output port 15B. The fiber-based optical processing unit 12 is configured and operable to controllably apply optical processing to the incoherent input optical signal and generate a first weighted combined optical signal (L (com) 1 ) w and the second weighted combined optical signal (L (com) 2 ) w Each of these weighted combined optical signals is composed of a first incoherent input signal Lin 1 and the second incoherent input signal Lin 2 The combination of the corresponding weighted parts forms the

[0037] The electro-optical processing unit 16 comprises a linear processor 16A and a nonlinear processor 16B. The linear processor 16A has an optical input terminal OI coupled to the first optical output port 15A and the second optical output port 15B of the optical unit 12, and is configured and operable to perform a weighted combination of the first optical signal (L (com) 1 ) w and the second weighted combined optical signal (L (com) 2 ) w To process the first weighted combined optical signal (L (com) 1 ) w and the second weighted combined optical signal (L (com) 2 )w , to correspond to the positive weighting and the negative weighting, to produce a resulting electronic signal ES that is output via the electrical output port EO of the linear processor 16A.

[0038] The nonlinear processor 16B may be configured as an electro-optical device, for example, the nonlinearity may be implemented based on the electronic nonlinearity of the photodiode of the linear processor (i.e., the response of the photodiode has a linear range and then reaches saturation when the capacitor is filled). Adjustment of the photodiode nonlinearity may be achieved by changing the type of photodiode or by changing the operating point via changing one or more circuit voltages.

[0039] Alternatively, the nonlinear processor 16B may be configured as an optical active device, such as an erbium doped fiber amplifier (EDFA) or a semiconductor optical amplifier (SOA), where (multiple) input signals compete for gain resources (cross-gain modulation), or through nonlinear processes such as the Kerr effect (cross-phase modulation), etc. Such optical nonlinear processors / operators are known and described, for example, in WO2021064727 and US2022327372 (assigned to the assignee of the present application), which are incorporated herein by reference.

[0040] Thus, in general, the non-linear processor 16B is adapted to receive an input signal indicative of an electronic signal output by the linear processor 16A, and is configured and operable to convert the input signal into an optical signal L out , the optical signal L out represents the total weighted output of the artificial neuron unit 10. This optical signal L out are not allowed to propagate in the output optical fiber 18, which is, for example, the input fiber of a neuronal unit of a subsequent layer.

[0041] Considering the electro-optical configuration of the non-linear processor 16B, it can be directly coupled to the electrical output EO of the linear processor 16A and operate to convert the electronic signal ES output from the processor 16A into an optical output signal L out , to propagate in the output optical fiber 18. It should be noted that, although not specifically shown in the drawings, in the case of an optical configuration of the nonlinear processor 16B, the neuron unit 10 also includes an electro-optical converter of any known suitable configuration arranged upstream of the nonlinear processor 16B, for example, a laser diode with a low coherence length.

[0042] The neuron unit 10 is associated with (eg, includes) a control board 20. The control board 20 includes a weighting controller 20A and two power controllers 20B and 20C. The weighting controller 20A is configured and operable to apply a control signal CS to the optical processing unit.1 , to cause variable optical attenuation of light propagating through the optical processing unit 12, for example, as will be further described below. The power controllers 20B and 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 the coefficients of the weighted sum and to generate corresponding operating data / signals CS to the linear processor 16A 2 and the controller 20C is configured and operable to define the shape of a nonlinear function that transforms the weighted sum of the inputs embedded in the electronic signal ES (which is the output of the linear processor 16B) into an optical output signal L out , and generate corresponding control data / signal CS 3 To operate the non-linear processor 16B.

[0043] refer to Figure 2B , schematically illustrates a neuron unit 10 of the present disclosure and describes in detail the operation of a fiber-based optical processing unit 12. The same reference numerals are used to identify functionally common components of all examples.

[0044] therefore, Figure 2B The neuron unit 10 is configured roughly similar to Figure 2A The neuron unit is configured as a hybrid system having an optical processing unit 12 based on an optical fiber and an electro-optical processing unit 16, and can be associated with a control board 20. The optical processing unit 12 based on an optical fiber has a first optical input port 14A and a second optical input port 14B for receiving an incoherent input signal, and has a first optical output port 15A and a second optical output port 15B optically coupled to an optical input end of a linear processor 16A, and the linear processor 16A receives the first weighted combined optical signal (L (com) 1 ) w and the second weighted combined optical signal (L (com) 2 ) w , and applying a predetermined mathematical function corresponding to positive and negative weightings to it. The electronic signal ES thus generated is received and processed by a nonlinear processor 16B, which converts the electronic signal into an optical output signal L representing the total weighted output of the artificial neuron unit 10. out .

[0045] like Figure 2B As shown, the optical input signal at the optical input ports 14A and 14B of the optical processing unit 12 is 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 is configured to split the input light field Lin1 into two optical fibers F1 at a predetermined ratio. A and F2 A The two light parts Lin1 p and Lin1 n Similarly, the splitter 24A is configured to divide the input optical field Lin2 into two optical fibers F1 and F2 at a predetermined ratio. B and F2 B The two light parts Lin2 propagated p and Lin2 n Therefore, each of the first optical input signal and the second optical input signal entering the processor 12 via the input ports 14A and 14B is divided into a pair of separated first and second optical propagation paths at a predetermined ratio.

[0046] Thus, the two arms of the fiber-based processor associated with the two inputs 14A and 14B provide a first pair of light propagation paths F1 A and F2 A And the second pair of light propagation paths F1 B and F2 B , the first pair of light propagation paths F1 A and F2 A Light portion Lin1 of the transmitted input light Lin1 p and Lin1 n , with a corresponding optical signal and , the second pair of light propagation paths F1 B and F2 B Light portion Lin2 of the transmitted input light Lin2 p and Lin2 n , with a corresponding optical signal and .

[0047] The first pair of F1 A -F2 A And the second pair of F1 B -F2 B At least one of the first light propagation path and the second light propagation path of each pair carries a controlled optical attenuator (VOA) 22 configured to apply a variable optical attenuation (VOA) to light propagating therethrough. Thus, the weighting (e.g., ) are applied to the first optical fiber F1 respectively A and F1 B and / or second optical fiber F2 A and F2B The propagated incoherent input light portion Lin1 p and Lin2 p and / or input optical part Lin1 n and / or Lin2 n .

[0048] Optical combiners 26A and 26B are also provided in the fiber-based optical processor 12. The optical combiners 26A and 26B are configured such that the combiner 26A combines the propagation path F1 A and F1 B And therefore the light part Lin1 p and Lin2 p Combined into output port 15A, and combiner 26B transmits propagation path F2 A and F2 B And therefore the light part Lin1 n and 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 and (L (com) 2 ) n More specifically, the first light propagation path F1 in the first pair and the second pair A and F1 B The optical signal L is coupled to the first optical output port 15A and combined at the first optical output port 15A to generate a first combined optical signal (L (com) 1 ) p (For example, ); and the second light propagation path F2 in the first pair and the second pair A and F2 B The optical signal L is coupled to the second optical output port 15B and combined at the second optical output port 15B to generate a second combined optical signal (L (com) 2 ) n (For example, ).

[0049] The combined optical signal (L) from the first optical output port 15A and the second optical output port 15B of the optical processing unit 12 (com) 1 ) p and (L (com) 2 ) n is fed to the linear processor 16A, generating an indication combined optical signal (L (com) 1 )p and (L (com) 2 ) n In this non-limiting example, the linear processor includes a dual-balanced photodiode configured and operable to provide an output electronic signal ES that is proportional to the difference between the weighted first combined optical signal and the second combined optical signal, for example, Therefore, the first combined optical signal (L (com) 1 ) p and the second combined optical signal (L (com) 2 ) n Positive and negative weighting are implemented.

[0050] It should be noted that the indices "p" and "n" used here are associated with 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 in these pairs, respectively. However, it should be understood that this can be defined in reverse. It should also be understood that negative and positive weights are actually implemented in the interaction of the combined optical signals with the linear processor 16A, and the size of the weights is defined by the controlled operation of the VOA 22. Therefore, these combined optical signals output via the output ports 15A and 15B are referred to herein as "weighted" combined optical signals.

[0051] It is known in the art that having linearity in neural networks is not enough and that one needs to deal with nonlinear activation functions, similar to the function of synapses in the brain's nervous system. Nonlinear functions are needed to speed up the convergence of the network and improve recognition accuracy, which is an indispensable part of neural networks. Nonlinearity can vary from a simple sigmoid to a complex dynamic system, depending on the neural model used.

[0052] As described above, the nonlinear function can be implemented electronically or optically by the nonlinear operator unit 16B. For example, the electrical signal ES from the double-balanced photodiode 16A can drive the electro-optic modulator 16B with an optical pump (controller 20C from the control board 20) to generate a nonlinearly transformed optical signal. In another non-limiting example, optical nonlinearity implemented by an optical active device (such as an EDFA) can be used, in which case a laser diode (converter from electrons back to photons) is placed upstream of the nonlinear module relative to the general direction of signal propagation through the neuronal unit.

[0053] Therefore, the resulting simulated optical signal Lout represents the total weighted light output that propagates to and through the output optical fiber 18 of the artificial neural unit 10, for example, to be input to subsequent layers of the network.

[0054] As described above and schematically shown in the figures, 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 the VOA 22, the dual-balanced photodiode 16A, and the nonlinear operator unit 16B.

[0055] In the following, the system architecture and neuron performance are illustrated in more detail, and the performance of the complete neural network is estimated based on measurement results.

[0056] Figure 2C The operation of an exemplary single neuron 10 is schematically illustrated. The neuron unit 10 includes an optical processing unit 12 having two inputs 14A and 14B for receiving incoherent input optical signals Lin1 and Lin2. As described above, the neuron 10 uses a push-pull mechanism to implement positive weighting and negative weighting. Each of the first input optical signal Lin1 and the second input optical signal Lin2 is divided by a corresponding fiber-coupled splitter 24A, 24B at a predetermined ratio (e.g., a 30 / 70 ratio) to provide a pair of optical propagation paths F1A-F1B, F2A-F2B for a corresponding pair of optical portions Lin1 p -Lin1 n 、Lin2 p -Lin2 n In this non-limiting example, the portion of light (Lin1) passing through 70% of the legs (F1A or F1B) of each light propagation path p or Lin2 p ) is used as a positive weight, which passes through the variable optical attenuator (VOA) 22 which is individually controlled by the control board 20; and the optical portion (Lin1) of the 30% branch (F2A or F2B) of each optical propagation path n or Lin2 n ) acts as a negative weight and is not attenuated. The positive weighted branches F1A and F1B are combined by the fiber coupled combiner 26A into the positive weighted output port 15A, generating a positive weighted combined optical signal (L (com) 1 ) p ; and the negative weighted branches F2A and F2B are combined by the fiber coupled combiner 26B to the negative weighted output port 15B, generating a negative weighted combined optical signal (L (com) 2 ) n These positively weighted combined optical signals (L (com) 1 )p and negatively weighted combined optical signal (L (com) 1 ) n is directed to the optical input of a double-balanced photodiode 16A (linear processor), generating an electronic signal ES corresponding to a differential output voltage, which is then converted by a nonlinear operator unit 16B into a signal representing the total output L of the neuron 10. OUT The simulated optical signal.

[0057] As already mentioned above, in addition to providing negative weights, the artificial neuron unit 10 is based on a push-pull control mechanism. If the first input 14A is considered "excitatory" and the second input 14B is considered "inhibitory", it can be easily shown that by increasing inhibition, an increased total output can be achieved, i.e. the neuron excitability increases. This illustrates the core paradox of the push-pull organization: increased force output can be achieved by increasing the background inhibition to provide greater disinhibition. Such an organization may be advantageous in designing more robust multi-layer artificial neural networks.

[0058] Figure 3A The complete photonic / computing system architecture 100 is schematically shown, which is configured as a multi-layer neural network, where each layer is formed by an array of independently operable neuronal units 10, which are configured as described above in accordance with the present disclosure. Analog electronic signals are generated by the control board 20 and converted into analog optical signals using a modulator. These analog optical signals are appropriately weighted and injected into the first neural layer with the addition of a bias. Each neuronal unit 10 is configured and operable as described above to perform an analog nonlinear function on the sum of the weighted inputs and send the resulting amplitude to the next layer. The output of the output layer neurons is read by the control board 20 via a photodiode.

[0059] Photos of exemplary assembled systems are at Figure 3B Shown in.

[0060] The inventors first characterized the timing of positive and negative weights. To do this, a 20ns square pulse at 1500nm was inserted into a single input of the neuron. The output was measured by an oscilloscope (Keysight MXR604A), as shown in Figure 2. Figure 4A A delay line of 40 cm in length (equivalent to a 2 ns delay) is then added to the optical path of the positive weight so that the negative and positive weights can be shown separately, as shown in Figure 4B The graph shows that the negative weights come before the positive weights and that the added spike is indeed 2ns long, as expected.

[0061] Then, the inventors inserted a simulated step function with 4 levels into the two inputs of the neuron, input 1 with a 10ns step period and input 2 with a 40ns step period. The outputs were recorded at different states of the VOA, such as Figure 4C As shown in Figure 2. When both VOAs are closed, the positive weights are equal to zero, resulting in a negative, falling step. When both VOAs are fully open, the strength of the combined positive inputs is much higher than the strength of the combined negative inputs, resulting in a rising step. The state where one VOA is open and the other is closed is an intermediate state. The VOA with the shorter period input is open, so the input has a positive weight, and therefore shows a rising output. The VOA with the longer period input is closed, so it has a negative weight and shows a falling step.

[0062] The inventors compared the output value measured by the dual-balanced photodiode with the expected output value as an evaluation of the accuracy of the MAC operation, such as Figure 4D The inventors expect that for an ideal MAC calculator, this graph is linear. Figure 4D The plot in shows a linear fit with R^2=0.9995, which means the neuron has excellent MAC accuracy.

[0063] The inventors completed a prototype system with 16 input channels and a 4-layer classifier and tested the performance of the system. The inventors compared the performance of the prototype system with the industry standard Nvidia DGX A100 and another photonic accelerator (LightMatter Envise server). The results showed an acceleration of up to 20 times that of the competing systems and a 2-order-of-magnitude improvement in power efficiency.

[0064] Therefore, the present disclosure provides a neuron unit based on hybrid fiber technology and electro-optical communication equipment and a photonic computing system using such a neuron unit, characterized by positive weighting and negative weighting schemes under incoherent data transmission conditions. The inventors show that such a design can achieve 5 to 20 times acceleration while improving power efficiency by more than 100 times.

Claims

1. An artificial neural unit for signal processing, the artificial neural unit include: a fiber-based optical processing unit having first and second optical input ports and first and second optical output ports and configured and operable to controllably apply optical processing to the incoherent input optical signal and produce 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 and produce a total weighted output of the artificial neuron unit by successively performing the following operations: applying a predetermined mathematical function to the weighted first and second combined optical signals to correspond to positive and negative weightings and produce a result signal; and applying non-linear processing to the result signal to convert the result signal into an optical output signal representing the total weighted output of the artificial neuron unit.

2. The artificial neural unit according to claim 1, in, The electro-optical processing unit comprises: a linear processor having an optical input coupled to the first optical output port and the second optical output port, and configured and operable to process the weighted first and second combined optical signals by applying the predetermined mathematical function to the weighted first and second combined optical signals to correspond to positive and negative weightings and output a resulting electronic signal; and a non-linear processor adapted to receive an input signal indicative of the resulting electronic signal, and configured and operable to convert the input signal into an optical output signal representing the total weighted output of the artificial neuron unit.

3. The artificial neural unit according to claim 1, in: The fiber-based optical processing unit includes a first splitter and a second splitter at the first optical input port and the second optical input port, and a first combiner and a second combiner at the first optical output port and the second optical output port, wherein the first splitter is configured to split the first optical input port into a first pair of separated first and second optical propagation paths at a predetermined ratio, the second splitter is configured to split the second optical input port into a second pair of separated first and second optical propagation paths at the predetermined ratio, the first combiner combines the first and second optical propagation paths in the first pair and the 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 in the first and second pair at the second optical output port to generate a second combined optical signal, and At least one of the first light propagation path and the second light propagation path of each of the first pair and the second pair is configured to apply variable optical attenuation (VOA) to light propagating through the path, thereby applying weighting to the incoherent input optical signal propagated through at least one of the first light propagation path and the second light propagation path of each of the first pair and the second pair, so that the first combined optical signal and the second combined optical signal are thereby the weighted first combined optical signal and the second combined optical signal, respectively.

4. The artificial neural unit according to claim 2, in, The non-linear processor is configured as an electro-optical device.

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

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

7. The artificial neural unit according to claim 2, in, The non-linear processor is configured as an optically active device, and the input signal received by the non-linear processor is an optical signal corresponding to a resulting electronic signal output of the linear processor.

8. The artificial neural unit according to claim 1, further comprising a control board configured and operable to control operations of the fiber-based optical processing unit and the electro-optical processing unit.

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

10. The artificial neural unit according to claim 2, in: The fiber-based optical processing unit includes a first splitter and a second splitter at the first optical input port and the second optical input port, and a first combiner and a second combiner at the first optical output port and the second optical output port, wherein the first splitter is configured to split the first optical input port into a first pair of separated first and second optical propagation paths at a predetermined ratio, the second splitter is configured to split the second optical input port into a second pair of separated first and second optical propagation paths at the predetermined ratio, the first combiner combines the first and second optical propagation paths in the first pair and the 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 in the first and second pair at the second optical output port to generate a second combined optical signal, and At least one of the first light propagation path and the second light propagation path of each of the first pair and the second pair is configured to apply variable optical attenuation (VOA) to light propagating through the path, thereby applying weighting to the incoherent input optical signal propagated through at least one of the first light propagation path and the second light propagation path of each of the first pair and the second pair, so that the first combined optical signal and the second combined optical signal are thereby the weighted first combined optical signal and the second combined optical signal, respectively.

11. The artificial neural unit according to claim 2, in, The non-linear processor is configured as an electro-optical device.

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

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

14. The artificial neural unit according to claim 2, in, The non-linear processor is configured as an optically active device, and the input signal received by the non-linear processor is an optical signal corresponding to a resulting electronic signal output of the linear processor.

15. The artificial neural unit of claim 2, further comprising a control board configured and operable to control operations of the fiber-based optical processing unit and the electro-optical processing unit.

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

17. An artificial neural network comprising two or more neuron layers, wherein the two or more neuron layers are arranged so that an optical input end of a subsequent layer in the two or more neuron layers is coupled to an optical output end of a previous layer in the two or more neuron layers, wherein each of the two or more neuron layers is formed by a plurality of independently operable artificial neuron units, each artificial neuron unit being configured according to claim 1.

18. An artificial neural network comprising two or more neuron layers, wherein the two or more neuron layers are arranged so that an optical input end of a subsequent layer in the two or more neuron layers is coupled to an optical output end of a previous layer in the two or more neuron layers, wherein each of the two or more neuron layers is formed by a plurality of independently operable artificial neuron units, each artificial neuron unit being configured according to claim 2.

Images