Optical computation device

Abstract

An optical computation device (10) is provided with: an optical modulator (50) that converts an electrical input signal into a modulated optical signal; an optical computation circuit (70) that performs an optical convolution operation on light output from the optical modulator (50) and includes a plurality of delay interferometers (703-1 to 703-L) connected in cascade; and an optical receiver (90) that receives output light from the optical computation circuit (70) and converts the output light into an electrical signal.

Description

optical calculation device

[0001] The present disclosure relates to optical computing devices.

[0002] Information processing using machine learning with deep neural networks (DNNs) has been attracting attention. DNNs are large-scale nonlinear networks in which a large number of neurons with nonlinear responses are connected by synapses.

[0003] For example, DNNs are widely applied to deep learning techniques that use networks with multi-layered neurons. This technique has been reported to demonstrate excellent performance in a wide range of fields, such as image or voice recognition, robot control, and artificial data generation.

[0004] Tait, Alexander N., et al. “Neuromorphic photonic networks using silicon photonic weight banks.” Scientific reports 7.1 (2017): 7430.

[0005] For example, when a convolution operation such as that used in a neural network is optically performed, the operation performance is expected to be improved.

[0006] An exemplary objective of the present disclosure is to provide an optical computing device that can improve computing performance and / or scalability with respect to computing performance.

[0007] Therefore, an optical computing device according to one aspect of the present disclosure includes an optical modulator that converts an electrical input signal into an optical modulated signal, an optical computing circuit that performs an optical convolution operation on the output light of the optical modulator, the optical computing circuit including a plurality of delay interference systems connected in cascade, and an optical receiver that receives the output light of the optical computing circuit and converts it into an electrical signal.

[0008] 1 is a block diagram showing an exemplary configuration of an optical arithmetic device according to embodiment 1. FIG. 2 is an equivalent circuit diagram of the optical convolution circuit shown in FIG. 1. FIG. 3 is a diagram for explaining optical convolution processing according to a modified example. (a) is a diagram showing an example result of the optical convolution processing shown in FIG. 3, and (b) is a diagram showing an example result of a normal convolution processing as a reference example. FIG. 4 is a block diagram showing an example configuration of an optical arithmetic device according to embodiment 2. FIG. 5 is a block diagram showing an example configuration of an optical arithmetic device according to embodiment 3. FIG. 6 is a block diagram showing an example configuration of an optical arithmetic device according to embodiment 4. FIG. 7 is a block diagram showing an example configuration of an optical arithmetic device according to embodiment 5. FIG. 8 is a diagram showing an example filter response of an optical convolution circuit.

[0009] Hereinafter, embodiments will be described in detail with reference to the drawings. However, the accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure, and are not intended to limit the subject matter described in the claims. Furthermore, more detailed descriptions than necessary may be omitted. For example, detailed descriptions of well-known matters or redundant descriptions of substantially identical configurations may be omitted.

[0010] Furthermore, in the drawings, identical or corresponding elements are appropriately designated by the same reference numerals. The drawings are schematic, and the dimensional relationships or ratios of elements may differ from reality. The drawings may also include portions in which the dimensional relationships or ratios differ. When numerical values ​​are used in the following description, they are merely examples, and other numerical values ​​may be used in addition or instead.

[0011] <Summary> As the number of layers or neurons (or nodes) in a DNN increases, the scale of nonlinear matrix operations can explode, which can increase the time or power required for inference or learning. For this reason, optical information processing technology that uses optical circuits as computing elements is being investigated.

[0012] For example, in a neural network configuration based on convolution operations used in image recognition, etc., a huge number of convolution matrix operations are expected to be performed, and therefore, electrical convolution matrix operations can significantly increase the calculation processing time and / or power consumption.

[0013] Therefore, by optically performing the convolution matrix calculation, it is possible to expect a drastic improvement in processing speed and a reduction in power consumption. As an example of such an optical convolution calculation machine, a configuration using a microring resonator type optical filter is being considered (for example, Non-Patent Document 1).

[0014] A convolution calculator using a microring resonator optical filter can perform 2NMB convolution operations at OPS by utilizing wavelength parallelism and spatial parallelism, where N is the number of wavelength parallelism, M is the number of spatial parallelism, and B is the baud rate. Note that "OPS" is an abbreviation for "operation per second," and represents the number of operations performed per unit time, and is an index of operation performance.

[0015] However, in a convolution computing device using a microring resonator optical filter, there is room for improvement in computing performance and / or scalability related to the computing performance. In some embodiments described below, optical computing devices that can improve computing performance (e.g., OPS) and / or scalability in a configuration different from that using a microring resonator optical filter are exemplified.

[0016] An example of the configuration of an optical computing device that can improve the OPS is a configuration in which multiple delay interference systems are connected in multiple stages (e.g., cascade). With such a configuration, for example, when the number of stages of delay interference systems is L, the OPS can perform up to 4LMNB operations. Therefore, it is possible to achieve better computing performance and / or scalability in terms of computing performance than a configuration using a microring resonator optical filter.

[0017] 1 is a block diagram showing an exemplary configuration of an optical arithmetic device 10 according to embodiment 1. As shown in Fig. 1, the optical arithmetic device 10 illustratively includes a light source 30, an optical modulator 50, an optical convolution circuit 70, and an optical receiver 90.

[0018] The light source 30 outputs, for example, a continuous wave (CW) laser beam. The optical modulator 50 modulates, for example, the laser beam input from the light source 30 with an electrical input signal x(t). For example, the optical modulator 50 may be an optical intensity modulator that converts input information into optical intensity, or an IQ modulator that converts input information into a complex electric field value including a phase.

[0019] The optical convolution circuit 70 performs optical convolution on, for example, the optical modulated signal input from the optical modulator 50. In the following description, the optical convolution circuit 70 may be abbreviated as the "operation circuit 70."

[0020] The optical receiver 90 receives the output light of the optical convolution circuit 70 and converts it into an electrical signal. For example, the optical receiver 90 may be a photodetector (PD) that outputs a voltage signal corresponding to the intensity of the received light as light intensity information, or a coherent receiver that measures a complex electric field value including phase information.

[0021] When an IQ modulator is used as the optical modulator 50 and a coherent receiver is used as the optical receiver 90, the excellent effect of being able to perform a convolution operation in complex space can be obtained.

[0022] The optical convolution circuit 70 is a non-limiting example of an optical circuit that optically implements a finite impulse response (FIR) filter (e.g., a lattice filter). For example, when an input signal x(t) such as image data is input to the optical convolution circuit 70, characteristic portions (in other words, feature quantities) of the image data can be detected or extracted by filtering.

[0023] As an example of this optical circuit, the optical convolution circuit 70 includes a variable splitter 701 and L (L is an integer of 2 or more) delay interference systems 703-1 to 703-L optically cascaded to the variable splitter 701.

[0024] Each of the delay interference systems 703-1 to 703-L may be referred to as a “unit cell.” In the following description, when the individual delay interference systems 703-1 to 703-L are not to be distinguished from one another, they may be abbreviated as delay interference systems 703.

[0025] The variable splitter 701 splits the output light from the optical modulator 50 at a variable splitting ratio, and inputs each of the split lights to the first-stage delay interference system 703-1. Illustratively, the variable splitter 701 has two input ports and two output ports, and the output light from the optical modulator 50 is input to one of the two input ports (the upper input port in the example of FIG. 1 ).

[0026] The variable splitter 701 can be configured using a Mach-Zehnder interferometer (MZI), and the branching ratio of the light output from the two output ports is varied by the phase value of a phase shifter 711 arranged in at least one of the two arm waveguides that make up the MZI.

[0027] Each of the delay interferometers 703 illustratively has two input ports and two output ports, which are optically coupled to the two output ports of the preceding delay interferometer 703, respectively, and which are optically coupled to the input ports of the succeeding delay interferometer 703, respectively.

[0028] The optical receiver 90 is optically coupled to one of the two output ports (for example, the upper output port in FIG. 1) of the delay interference system 703-L in the final stage (Lth stage) and receives the output light from that output port.

[0029] Each of the delay interference systems 703 illustratively includes a first MZI having two asymmetric arm waveguide lengths and a second MZI having two symmetric arm waveguide lengths. The first MZI delays one of two branched lights input from the variable splitter 701 or the preceding delay interference system 703 by a delay difference θ and causes it to interfere with the other. The second MZI branches the output light of the first MZI at a variable branching ratio and outputs it to the next delay interference system.

[0030] In each of the first MZI and the second MZI, at least one of the two arm waveguides is provided with, for example, a phase shifter 711 or 731. At least one arm waveguide of the first MZI includes, for example, an optical delay line for applying a delay difference θ between the arm waveguides.

[0031] With this configuration, in the optical convolution circuit 70, the optical branching with a branching ratio according to the phase value of the phase shifter 731 and / or 733 and the application of a delay difference θ to the branched light are repeated a number of times according to the number of stages of the delay interference system 703.

[0032] Here, the transfer function in the z-transform for the optical convolution circuit 70 can be expressed, for example, by the equivalent circuit (lattice filter) in FIG. 2 and the following equation (1): In equation (1), j=1, 2, ..., L, and the coefficient k j is given by equation (2): In Fig. 2, "E0" represents the field strength signal corresponding to the input signal x(t).

[0033]

[0034] A(z) and B(z) when j=1 are determined by the branching ratio k in the upstream variable splitter 701, which corresponds to the phase value of the phase shifter 711, and are given by, for example, the following equations (3) and (4).

[0035]

[0036] Here, in equation (1), φ j and ψ jrepresent the phase values ​​of the phase shifters 731 and 733 of each stage, respectively. -1 represents a time shift, the unit delay time of which is equal to the delay difference θ in each delay interferometer 703 .

[0037] In FIG. 2, in the j-th delay interference system 703-j, the signal input from the lower output port of the previous stage (j-1 stage) to the lower input port is delayed by an amount corresponding to the delay difference θ, and the coefficient k j is multiplied by the coefficient k j The signal multiplied by is added to the signal input from the upper output port of the previous stage to the upper input port, and the result is input from the upper output port to the upper input port of the next stage (j+1 stage).

[0038] In addition, the signal input from the upper output port of the previous stage to the upper input port is j After being multiplied by , the signal is added to the signal that is input from the lower output port of the previous stage (j-1 stage) to the lower input port and is delayed by a delay equivalent to the delay difference θ, and is input from the lower output port to the lower input port of the next stage (j+1 stage).

[0039] When the delay difference θ is made to match the sampling rate of the input signal x(t), in the equivalent circuit illustrated in FIG. 2, for X(z), which is the z-transform of the input signal x(t), the following is true for the two output ports of the final stage (Lth stage): Y(z) = A L (z) X(z) and Y(z) = B L (z)X(z) can be transformed.

[0040] The former Y(z) = A L (z) X(z) represents the transformation for the upper output port of FIG. 2, and the latter Y(z)=B L (z)X(z) represents the transformation for the lower output port in FIG.

[0041] Here, the kernel in the convolution filter is c = [c0, c1, c2, ..., c L ], the z-transform is expressed by the following equation (5).

[0042]

[0043] AL The parameter φ of each stage is set so that the condition (z) = C(z) is satisfied. j and ψ j By setting or controlling the number of parallel wavelengths N and M, it is possible to perform any convolution operation in the optical convolution operation circuit 70. In the configuration of the first embodiment, since the number of parallel wavelengths N is 1 and the number of parallel spatial connections M is 1, the OPS can perform 2BL convolution operations for the baud rate B of the input signal x(t).

[0044] In the existing configuration using a microring resonator filter, the cases of M=1 and N=1 are not anticipated in principle. However, in the first embodiment, the configuration illustrated in FIG. 1 can also accommodate the case where the number of parallel wavelengths N=1 or the number of parallel spatial channels M=1.

[0045] Furthermore, unlike conventional configurations, the convolution operation in the optical convolution operation circuit 70 is performed in the time domain, leaving room for utilizing the degree of freedom in the space domain and / or wavelength domain. Therefore, it is possible to further improve the operation speed by parallelizing the space and / or wavelength (in other words, by setting the values ​​of N and / or M to 2 or more). Examples of parallelization will be described later in embodiments 3 to 5.

[0046] Next, the parameter φ for obtaining the desired kernel filter characteristic C(z) is j and ψ j For the sake of simplicity, the phase value of the phase shifter 731 at each stage is determined as φ j = 0 to obtain a real kernel filter C(z). In this case, the above-mentioned formula (1) can be converted to the following formula (6) based on the following formulas (7) and (8):

[0047]

[0048] Equation (6) can be separated into an element P of a real product expressed by the following equation (9) and an element P of a matrix vector product expressed by the following equation (10).

[0049]

[0050] The matrix-vector product terms expressed by equation (10) determine the characteristics of the kernel filter, and the real product P expressed by equation (9) determines the gain or attenuation of the filter. Therefore, by converting equation (5) into the form of equation (11) below, the parameters of the desired kernel filter can be converted into the parameters of a lattice filter.

[0051]

[0052] For example, a determination method using simultaneous equations can be applied to the parameter conversion. For example, simultaneous equations can be obtained by expanding equation (11) and comparing it with equation (5), and parameters can be determined based on these simultaneous equations.

[0053] As a non-limiting example of a method for determining parameters, a case will be described in which a kernel c=[c0, c1, c2] is set in a lattice filter with L=2 stages. This means that the target filter response is c(z)=c0-c1z. -1 -c2z -2 This is equivalent to:

[0054] From equation (11), the response of the first stage is A1(z) = A0 + q1B0z -1 , B1(z)=q1A0−B0z -1 Therefore, the response of the second stage is A2(z) = A1(z) + q2B1(z)z -1 =A0+(q1B0+q2q1A0)z -1 +q2B0z -2 Comparing C(z) and A2(z), the following relations (12) to (14) are obtained.

[0055]

[0056] By simultaneously solving the equations (2) to (4) and the equations (12) to (14), the phase parameter ψ of the filter at each stage is calculated. j In this example, the case where the number of stages is L=2 has been described, but the present invention can be extended to any dimension (number of stages).

[0057] Although solutions using simultaneous equations generally tend to require a large amount of calculation, for example, by imposing constraints on the transfer function, it becomes possible to use an iterative method, thereby reducing the amount of calculation required to determine the parameters. For example, by imposing the constraint A = B = 1 / √2 (ψ = 45°), it is possible to efficiently determine the filter parameters using the iterative method.

[0058] In this case, the settable value of c is fixed to 1 / √2 due to the relationship A = c. However, the kernel to be set can be, for example, c' = [1 / √2, c / c, c / c, ..., c L / c0], it is possible to configure a kernel filter with the same characteristics as when no constraint is imposed.

[0059] The output intensity is 1 / √2·c0 times the original intensity, but this can be compensated for in the post-processing. Under these constraints, the output response B L (z) is A L Since the coefficients are in a transposed conjugate relationship with respect to (z), it is possible to determine the coefficients sequentially based on this relationship.

[0060] A non-limiting example of the calculation procedure will be described below. From equation (11), the Lth stage filter response A L z in (z) -L The coefficient multiplied by is q L Therefore, by comparing equation (3) with equation (11), the parameters of the final stage (L) of the lattice filter are q L = c L can be converted as follows.

[0061] Next, q L-1 Similarly, to determine L-1 Consider the response of (z). From equation (9), the j-1th filter response A j-1 (z) is the jth filter response A j (z), B j When expressed using (z), it is expressed as the following equation (15).

[0062]

[0063] A' L(z)=C(z), B' L (z)=C † Substituting (z) into equation (15), A' L-1 (z) can be calculated. Note that † means the conjugate transposed response, and in the case of real numbers, the transposed matrix C † (z) = c L +c L-1 z -1 +c L-2 z -2 +...+c0z -L is.

[0064] The resulting z -1 The coefficients of the term are the coefficients q of the lattice filter. L-1 By carrying out the same calculation for L-2 and onwards, the parameter q j This method can be performed by repeatedly solving equation (15) L times, so the amount of calculation can be reduced compared to the above-mentioned solution method using simultaneous equations.

[0065] Although the above example is an example of calculation using real numbers, it is also possible to determine parameters using calculation using complex numbers using a similar procedure. Furthermore, although the above-mentioned optical convolution calculation circuit 70 is a one-dimensional circuit, it is also possible to perform multidimensional optical convolution calculations by, for example, performing preprocessing on the input side.

[0066] For example, as shown in FIG. 3, by applying a time series data transformation such as the im2col function to the input side of the optical convolution circuit 70, two-dimensional image data can be converted into a one-dimensional input information sequence.

[0067] In this case, the output obtained by the optical convolution circuit 70 can be converted back into two-dimensional image data by the col2im function, which is the inverse operation of the im2col function on the input side. As a non-limiting example, Fig. 4(a) shows an example of a lattice filter output when using a kernel (filter) expressed by the following equation (16). Fig. 4(b) shows an example of the output of a normal convolution operation for reference.

[0068]

[0069] The filter parameters were determined by the aforementioned sequential method. As can be seen from Figures 3, 4(a), and 4(b), two-dimensional convolution filter operations can also be reproduced by the optical convolution operation circuit 70.

[0070] The modifications illustrated in FIGS. 3, 4(a) and 4(b) are also applicable to the following second to fifth embodiments.

[0071] 5 is a block diagram showing an example of the configuration of an optical arithmetic device 10A according to embodiment 2. The configuration shown in Fig. 5 is a configuration in which an optical convolution circuit 70 is capable of performing both forward propagation convolution and backward propagation convolution.

[0072] Therefore, optical computing device 10A differs from the configuration illustrated in FIG. 1 in that, for example, the input light to optical convolution circuit 70 can be modulated by input signal x(t) or error signal e(t), and optical receivers 90A and 90B are provided.

[0073] The optical receiver 90A and the optical receiver 90B are optically coupled to two output ports of the delay interference system 703-L in the final stage (Lth stage) of the optical convolution circuit 70, respectively, and receive the output light from each of the two output ports.

[0074] As in the first embodiment, each of the optical receivers 90A and 90B may be, for example, a PD that outputs a voltage signal corresponding to the received light intensity as light intensity information, or a coherent receiver that measures a complex electric field value including phase information. When an IQ modulator and a coherent receiver are used, the excellent effect of being able to perform a convolution operation in complex space can be obtained.

[0075] In machine learning, in addition to the forward propagation of the convolution operation, a backward propagation operation is also performed for parameter learning. In the optical processing device 10A having the configuration illustrated in Fig. 5, for example, the light reception result by the optical receiver 90A coupled to the upper output port indicates the forward propagation convolution operation result. In contrast, the light reception result by the optical receiver 90B coupled to the lower output port indicates the backward propagation convolution operation result.

[0076] The forward propagation method is the same as in the first embodiment. The back propagation of the convolution operation corresponds to the convolution operation in the conjugate transposed kernel filter. For example, in the above equation (11), when A = B (ψ = 45°), the output response B L (z) is A L It is the transposed conjugate of (z).

[0077] Therefore, by inputting the error signal e(t), which is a backpropagation signal, from the previous layer as the input signal and measuring the output of the lower port with the optical receiver 90B, it becomes possible to perform a conjugate transpose convolution operation equivalent to backpropagation.

[0078] Therefore, for example, both the forward propagation convolution operation used in the inference phase of machine learning and the back propagation convolution operation used in the learning phase of machine learning can be performed by the same convolution operation circuit 70.

[0079] Third Embodiment FIG. 6 is a block diagram showing an example configuration of an optical arithmetic device 10B according to a third embodiment. The configuration shown in FIG. 6 corresponds to a configuration in which optical convolution is spatially parallelized. As shown in FIG. 6, the optical arithmetic device 10B differs from the configuration of the second embodiment shown in FIG. 5 in that it includes a 1:M splitter 60 to which output light from an optical modulator 50 is input, and M parallel optical convolution circuits 70-1 to 70-M, where M is an integer greater than or equal to 2. Furthermore, optical receivers 90A and 90B are optically coupled to the two output ports of each of the optical convolution circuits 70-1 to 70-M.

[0080] The 1:M splitter 60, for example, branches the output light of the optical modulator 50 into M segments, and inputs the M-sequence branched light to M parallel optical convolution circuits 70-1 to 70-M, respectively. The configuration of each of the optical convolution circuits 70-1 to 70-M may be the same as that of embodiment 2. Each of the M optical receivers 90A is, for example, for forward propagation, and each of the M optical receivers 90B is, for example, for reverse propagation.

[0081] According to the optical arithmetic device 10B of the third embodiment, the optical convolution operation (forward propagation, or both forward propagation and backward propagation) described in the first or second embodiment can be executed for the number of ports that are spatially parallelized to M.

[0082] Therefore, for example, by independently setting or controlling the parameters of each of the optical convolution circuits 70-1 to 70-M, it is possible to perform convolution operations of different kernels in parallel on the same input (e.g., input signal x(t) or error signal e(t)).

[0083] Therefore, according to the optical calculation device 10B of the third embodiment, when the spatial parallelism number is M, the number of stages of the delay interference systems 703 cascaded in each of the optical convolution calculation circuits 70-1 to 70-M is L, and the baud rate is B, the calculation speed is improved to 2 MLB in OPS.

[0084] In machine learning, multiple kernel filters may be used in one convolutional neural network (CNN) layer. The parallel configuration exemplified in the third embodiment is useful for parallel calculation of such multiple filters.

[0085] In addition, in the configuration illustrated in FIG. 6, either the optical receiver 90A for forward propagation or the optical receiver 90B for backward error propagation may be omitted from some or all of the M parallel optical convolution circuits 70-1 to 70-M.

[0086] 7 is a block diagram showing an example of the configuration of an optical arithmetic device 10C according to embodiment 4. The configuration shown in Fig. 7 corresponds to a configuration in which optical convolution operations are spatially parallelized by inputting light from both the forward and reverse directions using the reciprocity of optical elements.

[0087] Embodiment 4 may be understood as a modified example of Embodiment 2. In Embodiment 4, the "forward direction" refers to the direction from left to right in Fig. 7, and the "reverse direction" refers to the direction from right to left in Fig. 7.

[0088] For example, as shown in FIG. 7, an optical calculation device 10C includes an optical convolution calculation circuit 70, light sources 30A and 30B, optical modulators 50A and 50B, and 2×2 couplers 62A and 62B corresponding to the forward and reverse directions, respectively.

[0089] The configuration of the light sources 30A and 30B may be the same as the configuration of the light source 30 illustrated in Fig. 1, and the configuration of the optical modulators 50A and 50B may be the same as the configuration of the optical modulator 50 illustrated in Fig. 1. Furthermore, the configuration of each of the optical receivers 90A to 90D may be the same as the configuration of the optical receiver 90 illustrated in Fig. 1.

[0090] When the input light is propagated in the forward direction in the optical convolution circuit 70, an optical receiver 90A for forward propagation and an optical receiver 90B for backward error propagation are optically coupled to the two output ports of the optical convolution circuit 70, respectively.

[0091] On the other hand, when the input light in the optical convolution circuit 70 is propagated in the reverse direction, an optical receiver 90C for forward propagation and an optical receiver 90D for backward error propagation are optically coupled to the two input ports of the optical convolution circuit 70, respectively.

[0092] The first 2×2 coupler 62A is optically coupled to the output of the optical modulator 50A, one of the two input ports of the optical convolution circuit 70, and the input of the optical receiver 90C. The 2×2 coupler 62A introduces the output light of the optical modulator 50A in the forward direction into one of the two input ports of the optical convolution circuit 70, while outputting the light output in the reverse direction from that input port to the optical receiver 90C.

[0093] The second 2×2 coupler 62B is optically coupled to the output of the optical modulator 50B, one of the two output ports of the optical convolution circuit 70, and the input of the optical receiver 90A. The 2×2 coupler 62B introduces the output light of the optical modulator 50B in the reverse direction into one of the two output ports of the optical convolution circuit 70, while outputting the light output in the forward direction from that output port to the optical receiver 90A.

[0094] In such a configuration, the signal introduced in the forward direction from the input side of the arithmetic circuit 70 by the light source 30A and the optical modulator 50A and the signal introduced in the reverse direction from the output side of the arithmetic circuit 70 by the light source 30B and the optical modulator 50B can be different signals.

[0095] The optical convolution circuit 70 executes the same convolution operation as in embodiment 1 or embodiment 2. Here, due to the reciprocity of the optical convolution circuit 70, which is an optical circuit, the light propagating through the operation circuit 70 undergoes the same transformation on the input side and the output side.

[0096] Therefore, by introducing different optical signals in the forward and reverse directions into the optical convolution circuit 70, it is possible to perform convolution operations using a common kernel in parallel on different input optical signals. This type of configuration is useful, for example, for mini-batch processing in machine learning.

[0097] The above-described fourth embodiment can also be combined with the spatial parallelization configuration exemplified in the third embodiment. In a configuration in which the fourth embodiment is combined with the third embodiment, parallel operations with a parallel number M can be performed in both the forward and backward directions. Therefore, the operation speed can be improved to 4 MLB in OPS.

[0098] 8 is a block diagram showing an example of the configuration of an optical arithmetic device 10D according to embodiment 5. The configuration shown in Fig. 8 corresponds to a configuration capable of collectively processing input signals of different wavelengths in parallel by wavelength division multiplexing (WDM).

[0099] For example, the optical arithmetic device 10D includes, in the upstream (or input side) of the optical convolution circuit 70, light sources 30-1 to 30-N and optical modulators 50-1 to 50-N corresponding to a plurality of wavelengths λ1 to λN (N is an integer of 2 or greater), and a wavelength multiplexing unit (WDM-MUX) 64.

[0100] The optical operation device 10D also includes a wavelength demultiplexer 80 and optical receivers 90-1 to 90-N corresponding to wavelengths λ1 to λN at the downstream (or output side) of the optical convolution operation circuit 70. The configuration of the optical convolution operation circuit 70 may be the same as that of the first or second embodiment, for example.

[0101] The light source 30-i (i = 1, 2, ..., N) outputs laser light (e.g., CW light) of wavelength λi. The optical modulator 50-i, for example, modulates the laser light of wavelength λi input from the light source 30-i with an electrical input signal x i (t). The input signal x i (t) can be different for each wavelength λi. The optical modulator 50-i can be, for example, an optical intensity modulator that converts input information into optical intensity, or an IQ modulator that converts input information into a complex electric field value including a phase.

[0102] The wavelength multiplexing unit 64 wavelength-division multiplexes the optically modulated signals input from the optical modulators 50-i, and inputs the resulting WDM light to the optical convolution circuit 70. The optical convolution circuit 70 performs the same optical convolution operation as in the first or second embodiment on the WDM light input from the wavelength multiplexing unit 64.

[0103] The wavelength demultiplexer 80 demultiplexes the output light of the optical convolution circuit 70 into wavelengths λi and outputs the separated light of wavelength λi to the corresponding optical receiver 90-i. The optical receiver 90-i receives the light of wavelength λi input from the wavelength demultiplexer 80 and converts it into an electrical signal.

[0104] For example, each of the optical receivers 90-i may be a PD that outputs a voltage signal corresponding to the intensity of received light as optical intensity information, or a coherent receiver that measures a complex electric field value including phase information. When an IQ modulator and a coherent receiver are used, the excellent effect of being able to perform convolution operations in complex space is obtained.

[0105] Here, the filter response (for example, transmitted field intensity versus frequency) of the optical convolution circuit 70 has a periodicity of 1 / θ (θ: delay difference in the delay interferometer 703) in frequency space, as illustrated in FIG.

[0106] Therefore, by setting the interval between each wavelength λi to 1 / θ in frequency space, it is possible to obtain a common kernel filter response for input light of different wavelengths λi. Therefore, by modulating the light of each wavelength λi with a different input signal x i (t), it is possible to perform convolution operations on the different input signals x i (t) in parallel.

[0107] The configuration illustrated in FIG. 8 can also implement the backpropagation calculation illustrated in the second embodiment by receiving and measuring the light output from the two output ports of the optical convolution circuit 70 using optical receivers.

[0108] In addition, the configuration illustrated in FIG. 8 can be combined with the spatial parallel configuration illustrated in embodiment 3 and / or embodiment 4, and when the number of wavelength multiplexing is N, the calculation speed is improved to 4LMNB in ​​OPS.

[0109] <Summary> As described above, according to the optical computing device disclosed herein, when the number of stages of the cascaded delay interference system 703 is L, for example, a maximum of 4LMNB calculations can be performed in the OPS, making it possible to achieve scalability in calculation speed superior to existing technologies.

[0110] <Terminology, etc.> The terms "connect" and "couple" used in this disclosure may be interpreted interchangeably. "Connect" or "couple" may be understood to mean any direct or indirect "connection" or "coupling" between two or more elements. For example, the term may be understood to include an indirect "connection" or "coupling" in which one or more intermediate elements are interposed between two elements that are "connected" or "coupled" to each other.

[0111] Any reference to an element followed by a designation such as "first...," "second...," etc. does not limit the quantity or order of those elements. These designations are merely used as a convenient way to distinguish between two or more elements. For example, a reference to a first and a second element does not imply that only two elements may be employed, nor does it imply that the first element must precede the second element in any physical quantity.

[0112] Although the present disclosure has been described in detail above, it is clear to those skilled in the art that the spirit and scope of the present disclosure are not limited to the contents described throughout the present disclosure. The present disclosure can be implemented in modified and altered forms without departing from the spirit and scope of the present disclosure as defined by the claims. Therefore, the description of the present disclosure is intended for illustrative purposes only and does not have any limiting meaning on the spirit and scope of the present disclosure.

[0113] The present disclosure is useful, for example, for neural networks based on convolution operations.

[0114] 10, 10A, 10B, 10C, 10D Optical arithmetic unit 30, 30A, 30B, 30-1 to 30-N Light source 50, 50A, 50B, 50-1 to 50-N Optical modulator 60 1:M splitter 62A, 62B 2×2 coupler 64 Wavelength multiplexing unit (WDM-MUX) 80 Wavelength demultiplexing unit (WDM-DEMUX) 70 Optical convolution arithmetic circuit 701 Variable splitter 703-1 to 703-L Delay interference system 711, 731, 733 Phase shifter 90, 90A, 90B, 90C, 90D, 90-1 to 90-N Optical receiver

Claims

1. An optical arithmetic device comprising: an optical modulator that converts an electrical input signal into an optical modulated signal; an optical arithmetic circuit that performs optical convolution on the output light of said optical modulator, said optical arithmetic circuit including a plurality of cascaded delay interference systems; and an optical receiver that receives the output light of said optical arithmetic circuit and converts it into an electrical signal.

2. The optical arithmetic device according to claim 1, wherein the optical arithmetic circuit includes a variable splitter that splits the output light of the optical modulator at a variable splitting ratio, and each of the delay interference systems includes: a first Mach-Zehnder interferometer that delays one of the two split lights split by the variable splitter and input from a delay interference system of a preceding stage, causing it to interfere with the other, and a second Mach-Zehnder interferometer that splits the output light of the first Mach-Zehnder interferometer at a variable splitting ratio and outputs it to the delay interference system of the next stage.

3. The optical arithmetic device of claim 1, wherein the input signal is a first signal corresponding to a forward propagation signal of a neural network or a second signal corresponding to an error backpropagation signal of the neural network, the optical receiver is a first optical receiver optically coupled to a first of two output ports of the optical arithmetic circuit, and the optical arithmetic device further comprises a second optical receiver optically coupled to a second of two output ports of the optical arithmetic circuit.

4. The optical arithmetic device according to claim 1, further comprising a 1:M splitter that splits the output light of the optical modulator at 1:M (M is an integer of 2 or more), and M optical arithmetic circuits are arranged in parallel corresponding to each of the M-sequence split lights split by the 1:M splitter.

5. The optical arithmetic device comprises, in addition to the optical modulator as a first optical modulator that converts the input signal, which is an electrical first input signal, into an optical modulated signal, a second optical modulator that converts a second electrical input signal into an optical modulated signal, and comprises first to fourth optical receivers as the optical receivers provided corresponding to the first input port, second input port, first output port and second output port of the optical arithmetic circuit, respectively, wherein the first optical modulator and the first optical receiver are optically coupled to the first input port of the optical arithmetic circuit via a first coupler, the second optical modulator and the third optical receiver are optically coupled to the first output port of the optical arithmetic circuit via a second coupler, the second optical receiver is optically coupled to the second input port of the optical arithmetic circuit, and the fourth optical receiver is optically coupled to the second output port of the optical arithmetic circuit, 2. The optical arithmetic device according to claim 1, wherein the first input signal and the second input signal are signals corresponding to forward propagation signals of a neural network or signals corresponding to back propagation signals of an error network.

6. The optical arithmetic device according to claim 1, comprising: optical modulators provided for a plurality of wavelengths corresponding to a plurality of said input signals; a wavelength multiplexing section that wavelength-division multiplexes the output light of each of said optical modulators and inputs it to said optical arithmetic circuit; a wavelength separation section that separates the output light of said optical arithmetic circuit according to said wavelength; and a plurality of said optical receivers provided corresponding to each of the light separated according to said wavelength.

Images