Machine Learning Systems

Abstract

A machine learning system comprising a light weight generator configured to include a discretionary light waveform shaper that generates weight-wavelength multiplexed light by mapping weights stored within a memory in a waveform direction and in a time direction, and a light computing core configured to include an electronic circuit and an optical circuit unit, wherein the machine learning system is characterized in that the light weight generator and the light computing core are connected by one or a plurality of optical fibers, the optical circuit unit is configured to perform the product-sum computation of a value indicated by an input signal and the weight indicated by the weight-wavelength multiplexed light, and signal input / output between the electronic circuit and the optical circuit unit is carried out via an analog-digital converter and / or a digital-analog converter in a lower operating band than the signal band of the weight-wavelength multiplexed light.

Description

[Technical Field]

[0001] The present invention relates to a machine learning system that uses optical circuits to perform computations for machine learning using deep neural networks. [Background technology]

[0002] Information processing using machine learning with deep neural networks (DNNs) has been attracting attention. DNNs are large-scale nonlinear networks in which numerous neurons with nonlinear responses are connected by synapses. Deep learning techniques, particularly those using networks with multi-layered neurons, have been widely applied. This technology has been reported to demonstrate excellent performance in a wide range of fields, including image and speech recognition, robot control, and artificial data generation. However, as the number of layers and neurons (nodes) increases, the scale of the required nonlinear matrix operations increases explosively, resulting in significant challenges in terms of the time and power required for inference and learning. In recent years, information processing techniques using optical circuits as computing elements have attracted attention as a method to solve these challenges.

[0003] In general DNN-based machine learning, the Lth layer signal x (L) The output is obtained by applying the transformation of the following equation (1) to x (L+1) =f(ω (L) x (L) ) (1)

[0004] where f is a nonlinear function such as tanh or relu, and x (L) ,x (L+1) are M-dimensional and N-dimensional input / output vectors, respectively, and ω (L) is a transformation matrix with M rows and N columns called a weight matrix. In machine learning, the weight matrix ω (L) By learning and updating, the desired final output is obtained.

[0005] In a DNN based on optical computation using optical circuits, part or all of this nonlinear matrix multiplication is performed optically. A known example of a conventional optical computation device is the optical computation device described in Non-Patent Document 1. Figure 1 illustrates a conventional optical computation device. This optical computation device 10 is composed of a digital circuit 11, an optical interferometer array 12, a high-speed digital-to-analog converter (DAC) 13 for input signal x, which converts the digital signal from the digital circuit into an analog signal, a DAC 14 for weight w, a high-speed analog-to-digital converter (ADC) 15, which converts the analog output signal y from the optical interferometer array into a digital signal, an E / O converter 16, which converts the input electrical signal into an optical signal, and an O / E converter 17, which converts the output optical signal into an electrical signal. By arranging multiple optical interferometer arrays as the optical interferometer array, the intensity of the input signal x is converted according to the phase value of the interferometer. The output optical signal y can be expressed as shown in Equation (2).

number

[0006] Equation (2) is the ω of the calculation equation (1) in the DNN mentioned above. (L) x (L) Regarding nonlinear conversion, in the case of Non-Patent Document 1, the received light is converted into an electrical signal and then digitally processed, but by placing some kind of nonlinear optical element in the subsequent stage of the interference system, it is possible to completely execute equation (1) including nonlinear calculations. In equation (1), the matrix product ω (L) x (L) Since the number of operations is limited by the nonlinear operation, it is also useful to perform the nonlinear operation after the optical-to-electrical conversion of the output optical signal. In this form, learning of the weight in equation (1) is realized by learning and updating the set phase value of the interference system. The phase value of the interference system can be dynamically changed by changing the refractive index of the waveguide, for example, by the thermo-optic effect.

[0007] Because the operating bandwidth of the above-mentioned optical circuit is extremely wide (> THz), it is possible, in principle, to perform the calculation of equation (2) at a higher frequency than electronic circuits. Furthermore, by utilizing the parallelism of the space and wavelength of light, it is possible to perform calculations in parallel. Therefore, optical computing circuits can exhibit advantages such as superior power consumption and computational speed compared to existing computing elements that use electronic circuits.

[0008] In practical applications, it is desirable for the input x and output y to be input / output to a digital electronic circuit, as shown in Figure 1. This is because it is difficult to store data and perform general-purpose calculations and data manipulation, including pre- and post-processing, in the optical signal state. Therefore, as shown in Figure 1, a DAC or ADC, as well as an EO converter or OE converter, are required. In this way, in conventional optical computing technology, optical signals are generated from digital circuits via DAC or EO conversion, so the computing speed of the computing circuit is limited by the operating bandwidth of the DAC / ADC or EO / OE converter. In general, the operating bandwidth of a DAC / ADC is slower than that of an EO / OE converter, so the operating bandwidth of the DAC / ADC in particular is the main limiting factor in computing speed.

[0009] Furthermore, conventional optical arithmetic devices are configured as a kind of in-memory computer in which the memory is transferred to a separate memory for the phase values ​​of an optical interference system, and calculations are performed on that interference system. [Prior art documents] [Non-patent literature]

[0010] [Non-Patent Document 1] Y. Shen et al., "Deep learning with coherent nanophotonic circuits", Nature photonics, Vol.11, P441-447, June 2017.<URL:https: / / www.nature.com / articles / nphoton.2017.93?TB_iframe=true&width=921.6&height=921.6>

Non-Patent Document 2

Non-Patent Document 3

Non-Patent Document 4

Summary of the Invention

[0011] The present disclosure is made to solve such problems of the prior art, and one of the objectives is to provide a machine learning system capable of eliminating the bottleneck of optical computing by DAC / ADC. Another objective is to provide a machine learning system configured to separately arrange an optical weight generator and an optical computing core and distribute the weight optical signal generated by the optical weight generator to the optical computing core via an optical fiber.

[0012] In order to achieve such an object, one embodiment of the present disclosure is a machine learning system having the following configuration.

[0013] The machine learning system includes an optical weight generator that generates one or more weighted wavelength-multiplexed lights, the optical weight generator including a memory that stores weights and one or more optical arbitrary waveform shapers that generate weighted wavelength-multiplexed lights by mapping the weights stored in the memory in the wavelength direction and the time direction, and an optical calculation core that includes an electronic circuit and an optical circuit unit. The optical weight generator and the optical calculation core are connected by one or more optical fibers. The optical circuit unit is configured to perform a product-sum operation between a value indicated by an input signal from the electronic circuit and a weight indicated by the weighted wavelength-multiplexed light from the optical weight generator. Signals are input and output between the electronic circuit and the optical circuit unit via an analog-to-digital converter and / or a digital-to-analog converter having an operating bandwidth lower than the signal bandwidth of the weighted wavelength-multiplexed light input to the optical circuit unit. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a diagram showing a conventional optical computing device. [Figure 2] FIG. 2 is a diagram illustrating a configuration of a machine learning system according to an embodiment of the present disclosure. [Figure 3] FIG. 3 is a diagram illustrating a first configuration example of an optical AWG included in an optical weight generator according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a diagram illustrating a second configuration example of the optical AWG included in the optical weight generator according to the embodiment of the present disclosure. [Figure 5] FIG. 5 is a diagram illustrating a third exemplary configuration of an optical AWG included in an optical weight generator according to an embodiment of the present disclosure. [Figure 6] FIG. 6 is a diagram showing a schematic configuration of an up-clocked optical arithmetic circuit according to an embodiment of the present disclosure. [Figure 7] FIG. 7 is a diagram showing a schematic configuration of a down-clocked optical arithmetic circuit according to an embodiment of the present disclosure. [Figure 8]FIG. 8 is a diagram illustrating a configuration of an optical circuit unit that implements a two-layer fully connected perceptron (MLP) according to an embodiment of the present disclosure. [Figure 9] FIG. 9 is a diagram illustrating a configuration of an optical circuit unit that implements a recurrent neural network (RNN) model according to an embodiment of the present disclosure. [Figure 10] FIG. 10 is a diagram illustrating an example of the layout configuration of a machine learning system according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0015] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. An embodiment of a machine learning system according to the present disclosure will be described with reference to Fig. 2. Fig. 2 is a diagram showing the configuration of a machine learning system 20 according to the present disclosure. In this embodiment, as shown in Fig. 2, the weight ω ijを An optical weight generator 21 that multiplexes and distributes the signal in terms of wavelength, time, and space is separated from an element 22 (hereinafter referred to as the "optical calculation core") that executes equation (1), and they are connected by one or more optical fibers 23.

[0016] The optical weight generator 21 generates the weight ω ij and a memory 24 for storing and holding the weight signal ω ijThe optical weight generator 21 also includes one or more optical arbitrary waveform generators 25 (hereinafter referred to as "optical AWGs") that generate a weighted wavelength-division multiplexed (WDM) optical signal by mapping the weights in the wavelength λ direction and the time t direction. The optical weight generator 21 also includes a memory 24, a processor for controlling the optical AWG 25, and an input / output interface. The optical weight generator distributes the weights as a wavelength-division multiplexed optical signal via an optical fiber 23. As will be described later, the optical computation core 22 may perform multiple matrix operations depending on the network topology. Therefore, it is desirable that the optical weight generator 21 has the function of distributing multiple weighted wavelength-division multiplexed optical signals in parallel. This can also be said to have the function of distributing a spatially multiplexed (SDM) wavelength-division multiplexed (WDM) optical signal. In this embodiment, the modulation speed of the weighted multiplexed optical signal is desirably faster than the sampling rate of the DAC / ADC of the optical computation core, which will be described later. The ratio of the modulation speed (i.e., operating speed) of the optical AWG to the sampling rate (i.e., operating speed) of the DAC / ADC at this time is hereinafter referred to as the oversampling ratio (OS ratio).

[0017] The optical calculation core 22 constituting the machine learning system of this embodiment is composed of, for example, a PC 26 including a memory and a processor for storing input and output data, a DAC 27 for converting digital input data into an analog signal, an ADC 28 for converting output data from the optical circuit unit into a digital signal, and an optical circuit unit 30 that performs photoelectric calculations using the analog RF signal from the DAC. The optical circuit unit 30 and PC 26 are connected via an FPGA or DAC / ADC board 29 including the DAC 27 and DAC 28 and a communication interface 31 configured, for example, by PCIE or Ethernet (registered trademark). The optical circuit unit 30 is configured to perform a product-sum calculation between a value indicated by an input signal input from an electronic circuit such as PC 26 via the DAC 27 and a weight indicated by weighted wavelength-multiplexed light from an optical weight generator.

[0018] In the optical computation core 22 of this embodiment, the DAC 27 needs to parallelize M vector dimensions of the input signal, and the ADC 28 needs to parallelize K vector dimensions of the output signal, but the sampling rate (hereinafter also referred to as the "operating bandwidth") of each can be slower than the weight signal of the optical AWG described above. As will be described later, this low-speed input signal has its operating bandwidth upconverted to the optical signal bandwidth by the optical circuit unit of this embodiment, and calculation based on equation (1) is performed. The final output of the calculation is then downconverted again to a signal of a low-speed bandwidth. This provides an excellent function, such as enabling calculation in the optical signal bandwidth without being limited by the operating bandwidth of the DAC or ADC.

[0019] (Light Weight Generator) Next, a specific example of the configuration of the optical AWG 25 included in the optical weight generator that constitutes the machine learning system of this embodiment will be described with reference to FIGS.

[0020] A first example of the optical AWG configuration is a system using an SDM-WDM optical modulator. The SDM-WDM optical modulator 35 is realized by, for example, parallelizing Mach-Zehnder modulators for spatial channels × wavelength channels. The example shown in Figure 3 is composed of optical modulators 361, 362, ... 36 that modulate the phase of an optical signal according to the input weight. M and a wavelength MUX 37 that multiplexes the output light from each optical modulator. This example shows a configuration in which optical modulators are parallelized for the number of wavelength channels, but multiple configurations like those in Figure 3 can be used to parallelize for the number of spatial channels. This first configuration example has the excellent feature of ease of implementation, as it can be realized using commercially available WDM modulators, etc. However, it has disadvantages in that the signal bandwidth of the optical multiplexed signal depends on the operating bandwidth of the optical modulator, making it difficult to sufficiently increase the OS ratio, and an optical modulator array for the number of spatial channels x wavelength channels is still required.

[0021] Next, a second configuration example of the optical AWG will be described with reference to Fig. 4. As the second configuration example of the optical AWG, there is a method of introducing an optical pulse δ(t) into a waveguide-type optical filter. The second configuration example 40 of the optical AWG shown in Fig. 4 includes a wavelength DEMUX 42 that wavelength-divides pulse light input from a pulse light source 41, and waveguide-type optical filters 431, 432, ... 43 to which the wavelength-divided wavelength light is input. M and a wavelength MUX 44 that multiplexes the output light from each waveguide-type optical filter. Fig. 4 shows an example of a configuration in which optical filters are arranged in parallel for the number of wavelength channels. For example, a waveguide-type optical filter can generate any designed impulse response by configuring it like optical filter 43 shown in the bubble in Fig. 4. Optical filter 43 is made up of optical waveguides 471, 472, 473, ... 47 that are branched by optical coupler 46 and coupled by optical coupler 49. i The optical waveguides 471, 472, 473, . . . 47 i Each of the optical waveguides includes a delay line in which the delay difference increases by θ. i Each optical waveguide may be provided with a variable attenuator as needed. i With this configuration, the optical pulse δ(t) undergoes the following transformation:

number

[0022] In equation (3), θ represents the delay time difference of the delay interference system, and ω ’ i represents the weight component by the i-th delay line. δ(t) is repeated for a period of N ’Assuming that a signal has intensity 1 only at θ, the configuration shown in Figure 4 can generate an arbitrary weighted signal over time depending on the weight setting of the optical filter. Furthermore, by parallelizing the configuration shown in Figure 4 as many times as the number of spatial channels, weights for the spatial channel × wavelength channel can be multiplexed and generated. This method has the advantage of not using an optical modulator and can be configured using a passive waveguide. Furthermore, although the OS ratio depends on θ, θ can be set arbitrarily by adjusting the delay length of the waveguide, which allows for arbitrary OS ratio setting. On the other hand, this method has the disadvantage of requiring an optical filter array equal to the number of spatial channels × wavelength channels. For simplicity, the optical filter shown in Figure 4 is an example using a finite impulse response (FIR) filter. However, as long as the impulse response can be designed, the optical circuit configuration is not limited to this example. Another method for obtaining an arbitrary impulse response is to use an infinite impulse response (IIR) filter. For example, a configuration such as that described in Non-Patent Document 2 can be used. Furthermore, as another configuration using an FIR filter, a lattice type filter may be used. One example of such a configuration is a configuration in which an optical interference system is formed in multiple stages, as shown in Non-Patent Document 3. Such a configuration is also included in the second configuration example.

[0023] Furthermore, a third example of the configuration of the optical AWG will be described with reference to Fig. 5. The third example of the configuration of the optical AWG is a method of introducing an optical pulse δ(t) into a spatial optical system type optical filter. The calculation principle is the same as in the second example of the configuration, but this method makes it possible to perform the calculation of equation (3) for the required spatial ch × wavelength ch in a single device by using a spatial optical system. For example, this can be realized by a configuration in which an input optical pulse signal is introduced into a spatial optical system as shown in Fig. 5. Fig. 5 shows the third example of the configuration 50. The third example of the configuration 50 includes a pulse light source 51, an input pulse signal from the pulse light source, and a weight ω ijThe spatial optical filter 52 includes an input optical fiber 521, a collimating lens 522, a diffraction grating 523, a Fourier lens 524, a spatial light modulator (SLM) 525, a Fourier lens 526, a diffraction grating 527, a collimating lens 528, and an output optical fiber 529. The pulsed light from the input optical fiber 521 passes through the collimating lens 523 and is split by the diffraction grating 523 to form different beam spots on the SLM 525 for each spatial channel and wavelength channel. The SLM 525 can provide different impulse responses for each channel by setting a phase pattern for each channel. The transmitted (or reflected) light from the SLM 525 is combined again by the diffraction grating 527, passes through the collimating lens 528, and is output from the output optical fiber 529. This configuration example has the excellent advantage of enabling the calculation of the above-mentioned equation (3) to be performed collectively in a single device. Using a reflective element (e.g., reflective liquid crystal on silicon, LCOS, or digital mirror device, DMD) in the SLM allows the optical system to be folded, so the input optical fibers 521 and 529, collimating lenses 522 and 528, diffraction gratings 523 and 527, and Fourier lenses 524 and 526 can each be configured using the same components. This contributes to achieving excellent functionality, such as reducing the number of parts and the housing size of the spatial optical filter 52. For simplicity's sake, the spatial optical filter 52 in FIG. 5 illustrates a configuration in which only wavelength channels are manipulated collectively. However, as described in Non-Patent Document 4, for example, spatial multiplexing can be achieved by spatially separating multiple input pulse beams in the y-axis direction, and then dispersing each input pulse beam separated in the y-axis direction in the x-axis direction. This also allows spatial channels to be integrated into a single device.

[0024] (Optical computing core) Next, with reference to Figures 6 and 7, the configuration of the optical computation core constituting the machine learning system of this embodiment will be described. For simplicity, the following description will be given assuming M = K, but the scope of the present invention is not limited to this condition. First, the main components of the optical computation core will be described. To optically perform a calculation such as that shown in Equation (1), it is necessary to be able to perform the matrix multiplication operation and nonlinear operation shown in Equation (2) above. The calculation operation shown in Equation (2) above can be decomposed into a vector dot product and its sum. In this embodiment, the optical circuit unit is configured by combining optical computation circuits that perform vector dot product and sum operations using two different methods, making it possible to upconvert low-speed signals from the DAC to the signal bandwidth of the optical signal and perform the calculation. Hereinafter, the two different types of optical computation circuits will be referred to as the up-clocked type and the down-clocked type, respectively.

[0025] FIG. 6 is a diagram for explaining the outline of the configuration of an arithmetic element that constitutes an up-clocked optical arithmetic circuit. FIG. 6(a) shows an example of an arithmetic element 60 that performs a vector dot product and that constitutes an up-clocked optical arithmetic circuit. The arithmetic element that performs a vector dot product and that constitutes this up-clocked optical arithmetic circuit is called a Type I vector dot product arithmetic element. The Type I vector dot product arithmetic element 60 comprises a wavelength DEMUX 61 and optical modulators 621, 622, ... 62. M and a wavelength MUX 63. The Type I vector dot product calculation element 60 converts the weighted wavelength-multiplexed light from the optical weight generator into wavelengths λ1, λ2, . . . λ at the wavelength DEMUX 61. M and optical modulators 621, 622, . . . 62 M The RF signals X1, X2, X are introduced from the DAC. M is introduced to the optical modulator as an RF signal, where the signal bandwidth of the RF signal is slower than the signal bandwidth of the weighted wavelength multiplexed light according to the OS ratio.

[0026] In each optical modulator, as shown by the signal waveform in Fig. 6(b), the weighted optical signal from the optical weight generator is i RF signal every X iAs a result, the input value indicated by the RF signal and the weighted optical signal are multiplied by a vector dot product, and the resulting optical signal is arranged in time and wavelength space and output.

[0027] This output can be expressed as the matrix shown in FIG. 6(c). As shown in FIG. 6(c), the column direction of this matrix is ​​the time axis direction, and the row direction is the wavelength axis direction. As described above, the Type I vector dot product calculation element 60 of FIG. 6(a) can perform MN calculations for each clock of the low-speed RF signal using M modulators, so the signal bandwidth of the output light output as a result of this calculation is faster by the OS ratio than the signal bandwidth of the original RF signal. This output signal is multiplexed by the wavelength MUX 63 and output as wavelength-multiplexed light. This Type I vector dot product calculation element 60 can speed up the signal bandwidth of the output light by the OS ratio than the signal bandwidth of the original RF signal.

[0028] FIG. 6(d) is a diagram showing the outline of the configuration of the summing operation element 64 that constitutes the up-clocked optical arithmetic circuit. FIG. 6(d) shows an example of the arithmetic element 64 that performs summation and that constitutes the up-clocked optical arithmetic circuit. This arithmetic element 64 that performs summation is called a Type I summing operation element. The Type I summing operation element 64 shown in FIG. 6(d) can be configured with a light-receiving element, for example, a photodetector (PD) 65. Generally, PDs do not have a large wavelength dependency in the band where WDM communication is used (for example, C band). Therefore, the reception by the PD 65 of the wavelength-multiplexed optical signal output from the vector dot product operation circuit can be regarded as a summing operation in the wavelength direction. Specifically, as shown in the area enclosed by the line in the matrix of FIG. 6(e), addition is performed in the row direction, which is the wavelength axis direction, to generate the RF signal y i will be output.

[0029] In this way, the matrix operation of Equation (2) can be performed by an up-clocked optical arithmetic circuit that combines the arithmetic elements shown in Figure 6(a) and (d). Because the signal bandwidth of the RF signal output from this up-clocked optical arithmetic circuit is increased in accordance with the OS ratio, an ADC with a low operating bandwidth cannot resolve the output signal.

[0030] Therefore, in the optical circuit unit of this embodiment, this up-clocked optical arithmetic circuit and the down-clocked optical arithmetic circuit shown in FIG. 7 are used in combination.

[0031] Next, the down-clocked optical arithmetic circuit will be described with reference to FIG. 7. FIG. 7(a) is a diagram showing an arithmetic element 70 that constitutes the down-clocked optical arithmetic circuit and that calculates a vector dot product. This arithmetic element 70 that calculates a vector dot product is called a Type II vector dot product arithmetic element. The Type II vector dot product arithmetic element 70 shown in FIG. 7(a) is configured to include an optical modulator 71 to which weighted wavelength-multiplexed light delivered from an optical weight generator and a high-speed RF signal output from the up-clocked optical arithmetic circuit are introduced as an RF input signal. The RF signal output from the up-clocked optical arithmetic circuit is speeded up according to the OS ratio, and a signal y according to the dimension N of the vector is generated. N are multiplexed in the time series direction. This high-speed RF signal becomes the RF input signal to the optical modulator 71. It is desirable that the weights introduced into the optical modulator 71 here be different from the weights introduced in the up-clocked optical arithmetic circuit. The vector dot product calculation by this optical modulator 71 can also be described by the matrix of FIG. 7(b). As shown in the matrix of FIG. 7(b), the relationship between wavelength and time is reversed compared to the matrix of FIG. 6(b). In this way, the Type II vector dot processor 70 can use the optical modulator 71 to calculate the vector dot product of the input value indicated by the RF signal and the weight indicated by the weighted optical signal.

[0032] Next, referring to FIG. 7(c), the operation element 72 that performs addition and constitutes the down-clocked optical operation circuit will be described. This operation element 72 is called a Type II addition operation element. The Type II addition operation element 72 demultiplexes the wavelength multiplexed light output from the vector dot product operation circuit using a wavelength DEMUX 73, and outputs each wavelength light to different PDs 741, 742, ..., 744 for each wavelength channel. M The operating bandwidth of these PDs is set lower than the signal bandwidth of the optical signal output from the vector dot product processor. For example, the relationship between the optical signal bandwidth and the PD operating bandwidth can be set as shown in Figure 7(e). This effectively averages the signals for each wavelength channel in the time domain. This averaging operation can be effectively interpreted as a summation operation in matrix calculations performed along the time domain. Specifically, in the matrix in Figure 7(e), as shown by the box, the values ​​of each row are added in the row direction corresponding to the time axis. Note that instead of setting the operating bandwidth of the PD lower, an RF filter or RF amplifier with a bandwidth lower than the signal bandwidth of the optical signal can be inserted after the PD. This configuration makes it possible to downconvert the high-speed wavelength-multiplexed light in the optical signal bandwidth output from the vector dot product processor to a low signal bandwidth that can be resolved by a low-speed ADC and output it as an RF signal.

[0033] Therefore, by combining the up-clocked optical arithmetic circuit shown in Figure 6 and the down-clocked optical arithmetic circuit shown in Figure 7, an optical circuit unit can be realized that has a low signal bandwidth for the input and output RF signals and can perform vector dot product operations at a speed equivalent to the OS ratio.

[0034] Furthermore, by using an optical unit circuit that combines an up-clocked optical arithmetic circuit and a down-clocked optical arithmetic circuit as the optical unit circuit 30 of the optical arithmetic core of the machine learning system shown in Figure 2, it is possible to provide a machine learning system that can eliminate the bottleneck of optical arithmetic caused by DAC / ADC.

[0035] FIG. 8(a) illustrates the configuration of an optical circuit unit 80 that implements a two-layer fully connected perceptron (MLP) by combining the up-clocked optical arithmetic circuit shown in FIG. 6 and the down-clocked optical arithmetic circuit shown in FIG. 7. FIG. 8(b) shows an equivalent network model of the MLP realized by this optical circuit unit 80. As shown in FIG. 8(a), a two-layer MLP can be realized by combining a Type I vector dot-product element 81, a Type I addition element 82, a Type II vector dot-product element 83, and a Type II addition element 84. The RF signal input to the Type I vector dot-product element 81 and the RF signal output from the Type II addition element 84 are low-speed RF signals that are input to and output from the electronic circuit via a low-speed DAC 85 or ADC 86. High-speed RF signals are input from the Type I addition element 82 to the Type II vector dot-product element 83. A nonlinear activation function is required in the hidden layer. For example, as shown by f in Figure 8(a), by introducing a Mach-Zehnder modulator in place of the Type II vector dot product calculation element 83, it is possible to apply a sinusoidal nonlinear function. Here, the case of a two-layer MLP is shown, but by extending the same operation, it can be expanded to calculations with any number of layers. In addition, by imposing constraints on the weights delivered by the optical weight signal input by the optical weight generator, calculations such as convolution calculations can also be performed with the same configuration.

[0036] The optical circuit unit 80 in Figure 8(a) described above executes feedforward neural network calculations without recursive connections, but by configuring it as shown in Figure 9(a), it is possible to implement an optical circuit unit that calculates the recursive neural network (RNN) model described by equations (4) and (5).

number

number

[0037] Here, Ω, M, and ω represent the connections of the input layer, intermediate layer (RNN layer), and output layer, respectively.

[0038] FIG. 9(a) illustrates the configuration of an optical circuit unit 90 that implements an RNN by combining the up-clocked optical arithmetic circuit shown in FIG. 6 and the down-clocked optical arithmetic circuit shown in FIG. 7. FIG. 9(b) shows an equivalent network model of the RNN realized by this configuration. As shown in FIG. 9(a), RNN operations are possible by combining a Type I vector dot-product arithmetic element 91, a Type I addition arithmetic element 92, a Type II vector dot-product arithmetic element 93, and a Type II addition arithmetic element 94. The RF signal input to the Type I vector dot-product arithmetic element 91 and the RF signal output from the Type II addition arithmetic element 94 are low-speed RF signals that are input to and output from the electronic circuit via a low-speed DAC 95 or ADC 96. High-speed RF signals are input from the Type I addition arithmetic element 92 to the Type II vector dot-product arithmetic element 93. In this case, a nonlinear activation function is still required in the hidden layer. However, for example, by introducing a Mach-Zehnder modulator at the location of the Type II vector-dot product optical processor 93 indicated by f in Figure 9(a), it is possible to apply a sinusoidal nonlinear function.

[0039] In the configuration shown in FIG. 9(a), the input to the optical circuit unit is also an RF signal input from a low-speed DAC, and the output signal is an RF signal that can be resolved by a low-speed ADC.

[0040] Next, let us consider the operation speed of the configurations shown in Figures 8 and 9. The optical circuit unit 80 that realizes the MLP shown in Figure 8 is considered to be equivalent to decomposing Equation (2) into two layers per unit time. The operation speed (operations per sec [OPS]) is expressed by the following Equation (6): OPS = (2MN+(M-1)+(N-1)) f da / ad (6)

[0041] Here, M and N indicate the number of dimensions of the input / output layer and the number of dimensions of the hidden layer, respectively. However, here, it is assumed that the number of dimensions of the input / output layer is the same. M corresponds to the number of parallel DAC / ADCs, and N corresponds to the OS ratio. For example, if the operating band of the optical AWG of the optical weight generator and the sampling rate of the DAC / ADC are respectively f awg =50 GSa / s, f da / ad = 100 MSa / s. In this case, the OS ratio in equation (2) is 500. In this case, assuming that the number of parallel DACs M = 100, it is estimated that a calculation speed of approximately 10 TOPS can be achieved.

[0042] Similarly, let us consider the calculation speed of the configuration that realizes the RNN in Figure 9. In the configuration that realizes the RNN in Figure 9, it is considered to be equivalent to solving equations (4) and (5) per unit time. The calculation speed is expressed by the following equation (7). OPS = (2MN+N 2 +(M-1)+(N-1)) f da / ad (7)

[0043] Assuming the same sampling rate and parallelism as in the case of Fig. 6(a), it is estimated that a computation speed of approximately 250 TOPS can be achieved. These computation speed values ​​are performance indices that exceed the theoretical computation performance of a typical CPU (approximately 500 G OPS).

[0044] (Form of use) Conventional configurations (e.g., Figure 1) are a type of in-memory computer, in which the memory is transferred to a separate memory for the phase values ​​of an optical interferometer and calculations are performed on that interferometer. In contrast, the machine learning system disclosed herein separates the optical weight generator, which has the memory function for storing weights and the distribution function, from the optical calculation core that performs the calculations, resulting in a configuration similar to that of general-purpose computing. Furthermore, in conventional general-purpose computers, weight distribution is performed via electronic cables, so it is desirable to place the memory and calculation circuit (e.g., CPU) as close as possible from the perspective of power and wiring. In contrast, the machine learning system disclosed herein distributes weights from the optical weight generator to the optical calculation core via optical fiber, enabling weight information to be transferred to the optical calculation core with almost no power consumption (no transmission loss). Furthermore, by utilizing the wavelength, time, and spatial multiplexing of light, weight information can be transferred using several optical fibers (or a single multi-core fiber). Therefore, unlike conventional computing units, the optical weight generator 101 and the optical computation cores 102, 102 can be arranged at a distance from each other, as shown in FIG. 10, and the optical weight generator 101 and the optical computation core 102 can be connected by an optical fiber 105. Note that FIG. 7 shows the optical circuit unit 103 and the digital circuit 104 included in the optical computation core 102. In this way, in the machine learning system of this embodiment, the optical weight generator and the optical computation core can be arranged at a distance from each other, so that the optical weight generator and the optical computation core can be arranged in different racks or different buildings. For example, it is possible to use the system in a manner that separates the client side where the optical computation core is arranged from the server side that distributes the weights. In this case, the client side only needs to consume memory for input and output [O(N)], and the weight information [O(N 2 )] can be stored and distributed to the server side. Also, the client side can perform calculations on the server side, up-converted to a high-speed optical signal band via a low-speed DAC / ADC. In addition, by distributing the signal from the optical weight generator to multiple clients, distributed learning is also possible.

[0045] As described above, the machine learning system disclosed herein enables the separation of the computational load between the client side and the server side as described above, enabling low-load, memory-saving calculations on the client side. [Industrial Applicability]

[0046] It is possible to provide a machine learning system that can eliminate the bottleneck of optical calculations caused by DAC / ADC.

[0047] It is also possible to provide a machine learning system in which the optical weight generator and the optical calculation core are arranged separately, and the weight optical signal generated by the optical weight generator is distributed by the optical calculation core via an optical fiber.

Claims

1. an optical weight generator for generating one or more weighted wavelength multiplexed lights, a memory in which weights are stored; one or more optical arbitrary waveform shapers that generate weighted wavelength-multiplexed light by mapping the weights stored in the memory in the wavelength direction and the time direction; an optical weight generator comprising: an optical calculation core including an electronic circuit and an optical circuit unit; A machine learning system comprising: the optical weight generator and the optical calculation core are connected by one or more optical fibers; the optical circuit unit is configured to perform a product-sum operation on a value indicated by an input signal from the electronic circuit and a weight indicated by the weighted wavelength-multiplexed light from the optical weight generator, The input and output of signals between the electronic circuit and the optical circuit unit is performed via an analog-to-digital converter and / or a digital-to-analog converter having an operating band lower than the signal band of the weighted wavelength multiplexed light input to the optical circuit unit. Machine learning systems.

2. The optical circuit unit of the optical calculation core comprises: an up-clocked optical arithmetic circuit that performs a product-sum operation on an input value indicated by an RF signal in the operating band of the analog-to-digital converter input from the electronic circuit and a weight indicated by the weighted wavelength multiplexed light from the optical weight generator, and outputs the result as an RF signal in the signal band of the weighted wavelength multiplexed light; a down-clocked optical arithmetic circuit that outputs a result of a product-sum operation between a value indicated by the RF signal in the signal band of the weighted wavelength multiplexed light and a weight indicated by the weighted wavelength multiplexed light from the optical weight generator as an RF signal in the operating band of the digital-to-analog converter; The machine learning system according to claim 1, characterized in that it is configured by combining the above.

3. The up-clocked optical arithmetic circuit comprises: a plurality of optical modulators; a wavelength demultiplexer that demultiplexes the weighted wavelength multiplexed light and inputs the demultiplexed light to each of the optical modulators; a wavelength MUX that multiplexes the output light from each of the optical modulators and outputs it as wavelength multiplexed output light; a vector dot product calculation element including: a summing operation element including a light receiving element that converts the wavelength multiplexed output light into an RF signal, Each of the optical modulators is configured to modulate the input light of each wavelength using an RF signal input via the analog-to-digital converter. The machine learning system according to claim 2 .

4. The down-clocked optical arithmetic circuit comprises: a vector dot product calculation element including one optical modulator that modulates and outputs the weighted wavelength multiplexed light using the RF signal in the signal band of the weighted wavelength multiplexed light; A plurality of light receiving elements; A wavelength demultiplexer that demultiplexes the input wavelength multiplexed light into individual wavelengths and outputs them to the respective light receiving elements; A summation operation element including: The machine learning system according to claim 2, characterized in that it is configured to include:

5. The optical arbitrary waveform shaper of the optical weight generator includes: a plurality of optical modulators that modulate light of different wavelengths according to input weights; a wavelength MUX for multiplexing the wavelength lights output from the respective optical modulators; The machine learning system of claim 1 , comprising:

6. The optical arbitrary waveform shaper of the optical weight generator includes: a pulsed light source; a wavelength demultiplexer that demultiplexes the impulse light from the pulse light source to generate light of a plurality of wavelengths; a plurality of waveguide-type optical filters that generate impulse responses according to the weights of the wavelength lights output from the wavelength demultiplexer; a wavelength MUX for multiplexing the wavelength lights output from the optical filters; The machine learning system of claim 1 , comprising:

7. The optical arbitrary waveform shaper of the optical weight generator includes: a pulsed light source; a spatial filter that receives impulse light from the pulse light source and generates one or more weighted wavelength multiplexed lights; The machine learning system of claim 1 , comprising:

8. 8. The machine learning system according to claim 1, wherein the optical weight generator and the optical calculation core are arranged in separate racks or separate buildings.

9. The machine learning system according to any one of claims 1 to 7, characterized in that it comprises a plurality of optical calculation cores, and the weight wavelength multiplexed light from the optical weight generator is distributed to each optical calculation core.

Images