Optical computation device

Abstract

This optical computation device comprises: a light source that emits laser light of one or multiple wavelengths; a first digital circuit that converts an electrical input signal into a matrix signal; one or multiple optical modulators that convert the matrix signal into an optical modulation signal; one or multiple optical computation circuits that perform an optical convolution computation on output light of the optical modulator(s), wherein the optical computation circuit(s) comprise: any one from among an optical computation circuit including a diffraction grating and a spatial optical modulator, an optical computation circuit including an MRR array, an optical computation circuit including a variable splitter and multiple delay interference systems connected to a cascade, and an optical computation circuit including a Mach-Zehnder interferometer (MZI) array; and one or multiple optical receivers that receive output light of the optical computation circuit(s) and convert the output light into an electrical signal.

Description

optical calculation device

[0001] This disclosure relates to an optical computing device.

[0002] Information processing using machine learning with deep neural networks (DNNs) is attracting attention. A DNN is a large-scale nonlinear network in which many neurons that exhibit nonlinear responses are connected by synapses, and deep learning techniques using networks in which neurons are arranged in multiple layers are widely applied. It has been reported that this technology can demonstrate excellent performance in a wide range of fields, such as image and sound recognition, robot control, and artificial data generation.

[0003] On the other hand, as the number of layers and neurons (nodes) increases, the scale of the required nonlinear matrix operations increases exponentially, leading to a significant increase in the time and power required for inference and learning. In recent years, information processing technology that utilizes optical circuits as computing elements has attracted attention as a method to solve these problems.

[0004] In particular, neural network configurations based on convolutional operations used in image recognition and other applications require the execution of a massive number of convolution matrix operations. Performing these operations optically is expected to drastically reduce processing speed and power consumption.

[0005] Trockman, A. , and Kolter, JZ, “Patches Are All You Need?”, arXiv preprint arXiv:2201.09792 (2022). Zagoruyko, S., and Komodakis, N., “Wide Residual Networks”, arXiv preprint arXiv:1605.07146 (2016). Kaiser, L., et al., “Depthwise Separable Convolutions for Neural Machine Translation”, arXiv preprint arXiv:1706.03059. (2017). Chetlur, S., et al. , “cuDNN: Efficient Primitives for Deep Learning”, arXiv preprint arXiv:1410.0759. (2014). Feldmann, J., et al., “Parallel convolutional processing using an integrated photonic tensor core”, Nature, 589(7840), 52-58. (2021). Shen, Y., et al., “Deep Learning with Coherent Nanophotonic Circuits”, Nature Photonics, 11(7), 441-446. (2017). "Illustration of the im2Col method, a convolutional processing technique," [online], July 8, 2017, [Retrieved January 22, 2025], Internet <URL: https: / / qiita.com / t-tkd3a / items / 6b17f296d61d14e12953> "Explanation of CNN convolution processing (mainly im2col) by a Python beginner (Learning Machine Learning #5)," [online], October 4, 2019, [Retrieved January 22, 2025], Internet <URL: https: / / qiita.com / MA-fn / items / 45a45a7417dfb37a5248>

[0006] However, in optical implementation, the scale of matrix operations that can be implemented is limited by manufacturing technology. Currently, it is possible to integrate approximately 1,000 to 10,000 computing elements (Mach-Zehnder interferometers or ring resonators). On the other hand, the above convolution matrix operation is generally expressed as Pconv = CinCooutK, where Cin and Cout are the number of input / output feature channels and K is the kernel size. 2 A certain number of parameters (i.e., the number of computing elements) is required. For example, if Cin = Cout = 64 and K = 4, the number of parameters (Pconv) becomes 65536, making integration difficult even with state-of-the-art optical integration processes.

[0007] Furthermore, in recent years, from the perspective of constructing general-purpose models that reduce induction bias in neural network models, research has been progressing on models that increase the number of parallel filters in the convolutional layer and the kernel size of the convolutional layer (Non-Patent Document 1 and Non-Patent Document 2), but implementation using optical elements becomes more difficult.

[0008] This disclosure addresses the above-mentioned problems and utilizes an approach that separates the convolution operation into depth and spatial directions (Non-Patent Document 3), thereby determining the required number of parameters as Psconv = Cin(K 2 The invention provides an optical computing device that can efficiently integrate optical data while reducing the overhead (+Out) to a minimum.

[0009] According to an embodiment for achieving this objective, an optical computing device is provided comprising: a light source that emits laser light of one or more wavelengths; a first digital circuit that converts an electrical input signal into a matrix signal; one or more optical modulators that convert the matrix signal into an optically modulated signal; one or more optical computing circuits that perform an optical convolution operation on the output light of the optical modulator, wherein the optical computing circuit is one of the following: an optical computing circuit including a diffraction grating and a spatial light modulator; an optical computing circuit including a microring resonator (MRR array); an optical computing circuit including a variable splitter and a plurality of cascaded delay interference systems; or an optical computing circuit including a Mach-Zehnder interferometer (MZI) array; and one or more optical receivers that receive the output light of the optical computing circuit and convert it into an electrical signal.

[0010] According to another embodiment, the first digital circuit in the optical computing device is characterized by performing a rearrangement operation on the input signal to convert it into a matrix signal.

[0011] Furthermore, according to another embodiment, the optical computing device is characterized by including a wavelength-dispersive element that is placed between the optical modulator and the optical computing circuit and linearly delays the output light of the optical modulator with respect to wavelength.

[0012] Furthermore, according to another embodiment, the optical computing device further comprises a second digital circuit that receives electrical signals from one or more optical receivers and performs subtraction processing.

[0013] Furthermore, according to another embodiment, the optical computing device is characterized by being composed of a connection between a first optical computing circuit, which is one of the optical computing circuits, and a second optical computing circuit, which is one of the optical computing circuits.

[0014] This is a block diagram comparing the methods of a normal convolution operation and a separable convolution operation, where (a) shows a normal convolution operation and (b) shows a separable convolution operation. This is a diagram illustrating the im2col transformation in a Depthwise convolution operation. This is a block diagram illustrating an exemplary configuration of an optical computing device according to Embodiment 1 for a Depthwise convolution operation. This is a block diagram illustrating an exemplary configuration of an optical computing circuit according to this embodiment. This is a block diagram illustrating an exemplary configuration of an optical computing device according to Embodiment 2 for a Depthwise convolution operation. This is a block diagram illustrating an example of an optical computing device according to Embodiment 3 for a Depthwise convolution operation. This is a block diagram illustrating a method for performing the im2col transformation in the optical domain according to this embodiment, where (a) shows the im2col transformation in a digital circuit and (b) shows the im2col in the optical domain. This is a block diagram illustrating an exemplary configuration of an optical computing device according to Embodiment 4 for a Depthwise convolution operation. This is a block diagram showing an exemplary configuration of an optical computing device according to Embodiment 5 for Depthwise convolution operations. Figure 10(a) is a block diagram showing an exemplary configuration of an optical computing device according to Embodiment 5, and Figure 10(b) is an equivalent circuit diagram of the optical computing device illustrated in (a). This is a block diagram showing an exemplary configuration of an optical computing device according to Embodiment 1 for Pointwise convolution operations. This is a block diagram showing an exemplary configuration of an optical computing device according to Embodiment 2 for Pointwise convolution operations. This is a block diagram showing an exemplary configuration of an optical computing device integrating Depthwise and Pointwise convolution optical computing circuits according to this embodiment. This is a block diagram showing an exemplary configuration of an optical computing device integrating other Depthwise and Pointwise convolution optical computing circuits according to this embodiment.

[0015] (Separable Convolution Operation) First, the outline of the separable convolution operation will be described below. In the field of machine learning, a method called separable convolution calculation has been proposed (Non-Patent Document 2) mainly for improving the efficiency of convolution operations on terminals with a small memory capacity, such as smartphones and edge devices. In this method, the efficiency of calculation is realized by separating the convolution operation in the depth direction and the spatial direction.

[0016] Fig. 1 shows a comparison between the normal convolution operation and the separable convolution operation methods. In the normal convolution operation, as shown in Fig. 1(a), since it is necessary to independently prepare kernel filters having parameters independently in the input channel direction and the spatial direction for the number of output sheets, the number of parameters is Pconv = CinCoutK as described above. 2 It becomes.

[0017] On the other hand, in the case of the separable convolution operation shown in Fig. 1(b), first, different kernel filters are applied to each input image. This operation is called Depthwise (depthwise separable) convolution, and the number of parameters Pdw = CinK.

[0018] After that, a convolution operation with a kernel size of "1" is performed on the intermediate output of Cin dimensions obtained. Different from the Depthwise convolution, in this operation, a convolution operation is also performed in the input channel direction, but since the kernel size is 1, no spatial convolution is performed. Prepare this filter for the output dimension Cout and obtain Cout final outputs.

[0019] This operation is called Pointwise (1×1) convolution, and the required parameters are Ppw = CinCout. This operation is equivalent to taking a matrix product of Cin×Cout dimensions with respect to the input vector in the Cin direction (each matrix element corresponds to an element in the kernel).

[0020] The total number of parameters of the separable operation described above is Pscconv = Ppw + Pdw = Cin(K 2 + Cout), and compared with the normal convolution operation, the number of parameters is (Pscconv / Pconv) = (1 / Cout) + (1 / K 2It can be reduced to ( ). In particular, it has excellent advantages such as being able to reduce the number of parameters as the kernel size K and the number of parallel filters increase.

[0021] (Embodiment 1 of Optical Implementation for Depthwise Convolution Operation) Hereinafter, the Depthwise convolution operation in the method of implementing the separable convolution operation described above using optical elements will be described. In the optical computing device described below, there are a plurality of components having the same configuration for performing operations on a plurality of input signals in parallel. In some cases, one of these components will be typically described, and it should be noted that a plurality of components having the same configuration are used. In the following description, when not distinguishing individual components with the same configuration (for example, optical modulators 203a to 203m) from each other, they may be typically abbreviated as optical modulator 203.

[0022] Depthwise convolution is a convolution that applies only one filter to one input channel. Therefore, Depthwise convolution is realized by executing the convolution operation of kernel size K in parallel for each input channel in the Cin dimension.

[0023] Figure 2 shows a method of efficiently calculating the Depthwise convolution operation using the rearrangement of the input matrix and the matrix multiplication operation. Although various calculation methods have been proposed for the convolution operation, for each input matrix x ∈ R A×A a rearrangement operation called "im2col" is performed to convert it into a matrix u ∈ R K^2×T (Non-Patent Documents 7 and Non-Patent Document 8), and the converted matrix u is efficiently calculated by applying a matrix W ∈ R K^2×Cout corresponding to the kernel filter (Non-Patent Document 4). Each element ω of the matrix W is the same as the element k of the kernel filter.

[0024] Here, in the case of the Depthwise convolution operation, as described above, since each input image is operated on by one kernel filter, for each matrix u of the input signals converted by the im2col operation, W ∈ R K^2×1By performing matrix multiplication calculations using this method, Depthwise convolution operations can be executed. Therefore, as an example, Depthwise convolution operations can be realized optically using a mechanism that transforms input x into matrix u using im2col, a mechanism that optically performs matrix multiplication in parallel, and a mechanism that converts digital electronic data into optical analog information.

[0025] Figure 3 shows a first embodiment for Depthwise convolution in an optical computing device of the present disclosure. In this embodiment, the optical computing device 10 includes, exemplary, an optical modulation unit 20 including a digital circuit 50 for im2col operation, an optical computing unit for Depthwise convolution (hereinafter referred to as an optical computing unit) 30-1 including one or more optical computing circuits 301, and an optical receiving unit 40. In Figure 3, as a non-limiting example, the digital circuit 50 is shown to be mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0026] (Optical Modulation Unit) The digital circuit 50 performs the im2col operation described above and converts each input signal xi into a matrix ui∈R N×T Convert to (N=K 2 ). As a non-limiting example, the digital circuit 50 may be an FPGA or an ASIC, etc. Here, i is the input dimension number i = 1, 2, ..., Cin. In order to represent the dimension N direction in terms of each optical wavelength and the dimension T direction in terms of time, each signal matrix ui is converted into an analog signal by a Cin × N DAC 60 and introduced into optical modulators 203a to m corresponding to each wavelength.

[0027] In the optical modulator 203, N wavelength signals are extracted from the N' wavelength comb ray generated from the light source 201 using, for example, a wavelength filter, and the signal matrix ui for each is converted into an optical signal. The total number of wavelengths N' generated from the light source 201 is greater than N (N' = pN, p ≥ 1), and it is possible to use N wavelength signals, which are a part of this total. In addition, as a non-limiting example, the optical modulator 203 can 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 phase.

[0028] Here, the light source 201 may, in example, be a frequency comb light source. In this case, it is possible to generate input light with multiple wavelengths (e.g., N wavelengths) all at once, which has the excellent feature of enabling simplification and integration of the device.

[0029] The output ray from the light source 201 can be split into multiple rays according to the number of input signals xi. This splitting may be performed spatially or using optical fibers or the like. The split rays are input to wavelength demultiplexers 205a to m (WDM-DEMUX) located on the output side of the light source 201, and can be further split for different wavelengths. Subsequently, the rays split for each wavelength are input to a plurality of optical modulators 203 located on the output side of the wavelength demultiplexer 205, and can be modulated by a matrix ui introduced into each optical modulator 203.

[0030] Each optical signal modulated by the matrix ui introduced into each optical modulator 203 is combined by a wavelength multiplexer (WDM-MUX) 207 located on the output side of the optical modulator 203. The optical signals can then be input to the optical computing unit 30-1 via an optical fiber array 209 located between the optical modulation unit 20 and the optical computing unit 30-1.

[0031] Figure 3 shows a configuration in which the optical modulation unit 20 and the optical computing unit 30-1 are modularized and connected by fiber, but both units may be mounted on a single substrate. In the above modularization, multiple optical signals are propagated together using a wavelength multiplexer (WDM-MUX) 207 to reduce the number of optical fibers, but each optical signal output from multiple optical modulators 203 may be introduced into each optical matrix computer 303 located in the optical computing unit 30-1.

[0032] (Optical Computing Unit) The optical computing unit 30-1 in this embodiment comprises one or more optical computing circuits 301a to m having the same configuration depending on the number of input signals. Each optical computing circuit 301 also includes, as an example, a wavelength demultiplexer (WDM-DEMUX) 305, one or more wavelength division multiplexing (WDM) type optical matrix computers 303, and a wavelength multiplexer (WDM-MUX) 307.

[0033] Each optical signal ui from the optical modulation unit is input to the wavelength demultiplexer 305 and wavelength-separated. Each wavelength-separated optical signal ui is used to perform matrix calculations in parallel using one or more WDM type optical matrix computers 303 (e.g., variable optical attenuators (VOAs)) located in the optical computing unit 30-1.

[0034] Each wavelength-separated optical signal is subjected to optical attenuation (or amplification or phase modulation) corresponding to the weight of each wavelength by a variable optical attenuator (VOA), then wavelength-combined again by a wavelength multiplexer (WDM-MUX) 307, and output to an optical receiver 401 (e.g., a photodetector (PD)).

[0035] Hereinafter, with reference to Figure 4, several embodiments for configuring the above-described optical computing circuit will be explained.

[0036] Figure 4(a) shows the configuration of the optical processing circuit and the propagation state of the optical signal in this embodiment. As shown in Figure 4(a), the WDM optical signal ui of the i-th optical path is input to the i-th optical input port in the optical fiber array 209. This WDM optical signal ui is emitted into space through the first lens 308a (for example, a collimating lens, etc.) and wavelength-delimited in the x-direction by the diffraction grating 305-1.

[0037] The demultiplexed optical signal is focused onto a spatial light modulator (SLM) 303-1 at different positions depending on the optical path i and wavelength λj via a second lens 309a (e.g., a collimating lens). On the SLM 303-1, different spatial light can be modulated by displaying independent patterns in each region. The optical signal modulated by the SLM 303-1 is then wavelength-combined again by a diffraction grating 305-1 and coupled to an output port corresponding to the optical path i.

[0038] The coupling efficiency in this case can be arbitrarily set according to the spatial modulation pattern on the SLM. This corresponds to providing an arbitrary attenuation state ωij depending on each wavelength λj and optical path i. This configuration has the excellent function of realizing a large number of optical circuits for large-scale matrix operations in a simple configuration, because it uses a single diffraction grating for wavelength multiplexing and demultiplexing and a single SLM for the variable optical attenuator (VOA).

[0039] SLMs are broadly classified into phase-modulated types (e.g., Liquid crystal on silicon, LCOS) and intensity-modulated types (e.g., Digital mirror device, DMD), but either type of device can be used in this configuration. In this case, the method of controlling ωij differs. In the case of a spatial light modulator capable of multi-level intensity modulation, ωij can be generated by simply displaying pixels corresponding to the desired attenuation intensity ωij. If multi-level modulation is not possible and only binarized values ​​can be set in the SLM, it is possible to implement multi-level ωij by controlling the ratio of coupled pixel values ​​(corresponding to 1) and attenuated pixel values ​​(corresponding to 0) for the area irradiated by each light beam.

[0040] In the case of a phase-modulated SLM, it becomes necessary to convert the given phase pattern information into intensity modulation. For example, by displaying a periodic phase pattern (e.g., sine wave, sawtooth wave, etc.) on the SLM and controlling the period for each region, the diffraction angle from the SLM changes. This makes it possible to control the coupling efficiency to the output fiber, and the control of ωij becomes possible through the phase pattern display. This embodiment is not limited to the above and can be applied to various SLM configurations and control methods.

[0041] Alternatively, as shown in Figures 4(b, c), an optical fiber 209 may be used in the front end. Figure 4b shows the optical arrangement in the x-z plane, and Figure 4c shows the optical arrangement in the y-z plane. Similar to the front end described above, the WDM optical signal ui output from the optical fiber 209 is emitted into space through the first lens 308a (for example, a collimating lens), and wavelength-deselected in the x direction by the diffraction grating 305-1.

[0042] The decomposed optical signal is focused onto a spatial light modulator (SLM) 303-1 at different positions depending on the optical path i and wavelength λj via a second lens 309a (e.g., a collimating lens). On the SLM 303-1, different spatial light can be modulated by displaying independent patterns in each region.

[0043] To suppress light diffusion, it is desirable to arrange each of the above components at a predetermined distance. For example, if the focal length between the optical fiber and the first lens 308a is f1, and the focal length between the second lens 309a and the SLM 303-1 is f2, it is desirable that the focal length between the first lens 308a and the second lens 309a be f1 + f2.

[0044] Figure 4(d) shows another embodiment, illustrating the configuration of an optical computing circuit using an optical waveguide instead of an optical fiber. While the above exemplifies an optical system at the front end using an optical fiber system, a waveguide-type front end 211, for example, as described in Non-Patent Literature 6, may also be used. In this case, excellent advantages such as the ability to increase the number of achievable parallelisms M compared to the fiber system are realized.

[0045] Furthermore, the ability to incorporate the same function as a lens within the waveguide without using the first lens 308a is advantageous from an implementation standpoint. In addition, the focal length f between each component can be made smaller than the aforementioned focal length f2, which is also advantageous for miniaturization.

[0046] (Optical receiving unit) Returning to Figure 3, the optical receiving unit 40 comprises at least one or more PDs 401a to m. The optical receiving unit 40 may also optionally or additionally comprise one or more ADCs 403 and digital circuits 405.

[0047] The optical signal output from the optical processing unit 30-1 can be input to PD401. Since PD401 generally has low wavelength selectivity (dependence), its output signal can be considered as the sum of all wavelengths. Therefore, the output yi from each PD can be considered as the result of a Depthwise convolution operation on each input xi.

[0048] Optionally, the output yi from each PD can be converted into a digital signal by the ADC 403 and subtracted by the digital circuit 405 (e.g., FPGA or ASIC).

[0049] The configuration of this disclosure performs calculations in intensity space, and therefore, simple configurations cannot perform subtraction; some additional processing is required when performing subtraction. For example, similar to the technology described in Non-Patent Document 5, subtraction can be achieved by separating it into positive and negative operations.

[0050] For example, this can be done by separating the positive and negative components of ωij into ω+ and ω-, obtaining the calculation results for each independently, and then performing a subtraction operation on a digital circuit. Alternatively, by separating the positive and negative calculations for each optical path and receiving them with a differential receiver, it is also possible to optically perform the calculation in equation (4) described later.

[0051] However, the ADC 403 and digital circuit 405 are optional, and when the optical computing unit 30-2 that performs Pointwise convolution operations (described later) is connected downstream of the optical computing unit 30-1 that performs Depthwise convolution operations, the ADC 403 and digital circuit 405 may be placed downstream of the optical computing unit 30-2 that performs Pointwise convolution operations.

[0052] (Embodiment 2 of Optical Implementation for Depthwise Convolution Operations) Figure 5 shows a second embodiment for Depthwise convolution operations in an optical computing device of the present disclosure. In this embodiment, the optical computing device 10A includes, exemplary, an optical modulation unit 20 including a digital circuit 50 for im2col operation, an optical computing unit for Depthwise convolution operations (hereinafter referred to as the optical computing unit) 30-1, and an optical receiving unit 40. In Figure 5, as a non-limiting example, the digital circuit 50 is shown to be implemented inside the optical modulation unit 20, but the digital circuit 50 may be implemented outside the optical modulation unit 20.

[0053] However, in the following description, the optical modulation unit 20 and optical receiving unit 40, which have the same configuration as the optical computing device 10 of the first embodiment described above, will not be explained, and only the optical computing unit 30-1, which has a different configuration, will be described.

[0054] As shown in Figure 5, the optical computing unit 30-1 includes one or more optical computing circuits composed of wavelength filters using microring resonator (MRR) arrays 70a to m.

[0055] The optical signal ui output from the optical modulation unit 20 is input to the input port of each microring resonator (MRR). Subsequently, the transmittance of the optical signal ui to the output port (corresponding to the weighting coefficient ω of W) is controlled by the MRR array 70, which is arranged for each wavelength λj, and the output light, which is the result of the calculation, is output to the PD401. Because this configuration is realized with a waveguide structure, it has the excellent advantage of being able to reduce the device size compared to the spatial optical type method described above.

[0056] (Embodiment 3 of optical implementation for Depthwise convolution operation) In the above embodiment, the im2col conversion was performed in the digital domain, so K was generated for each input channel. 2 This presents challenges such as the need for multiple DACs and optical modulators, which limits scalability. In this embodiment, we disclose a configuration that reduces the number of optical modulators and DACs required for each input channel to one by performing im2col in the optical domain.

[0057] Figure 6 shows the configuration of this embodiment. In this embodiment, the optical computing device 10B includes, as an example, an optical modulation unit 20 including a digital circuit 50, an optical computing unit 30-1, and an optical receiving unit 40. In Figure 6, as a non-limiting example, the configuration in which the digital circuit 50 is mounted inside the optical modulation unit 20 is shown, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0058] However, in the following description, the optical computing unit 30-1 and the optical receiving unit 40, which have the same configuration as the optical computing device 10 of the first embodiment described above, will not be explained, and the optical modulation unit 20 and the optical fiber array 209-1, which have different configurations and are arranged between the optical modulation unit 20 and the optical computing unit 30-1, will be described.

[0059] The optical modulation unit 20 in this embodiment includes a digital circuit 50, a DAC 60 for converting an input signal to an analog signal, a light source 201, and one or more optical modulators 203a to m. The optical modulation unit 20 can also be connected to the optical computing unit 30-1 by a linear wavelength dispersion element 209-1.

[0060] In the optical modulator 203, N wavelength signals are extracted from the N' wavelength comb ray generated from the light source 201, for example using a wavelength filter, and the signal matrix ui for each is converted into an optical signal. The total number of wavelengths N' generated from the light source 201 is greater than N (N' = pN, p ≥ 1), and it is possible to use N wavelength signals, which are a portion of this total.

[0061] Here, the light source 201 may, in example, be a frequency comb light source. In this case, it is possible to generate input light with multiple wavelengths (e.g., N wavelengths) all at once, which has the excellent feature of enabling simplification and integration of the device.

[0062] The output ray from the light source 201 can be split into multiple rays according to the number of input signals xi. This splitting may be performed spatially or using optical fibers or the like. The split rays are input to multiple optical modulators 203a to m located on the output side of the light source 201, and can be modulated by a matrix ui introduced into each optical modulator 203.

[0063] In the configuration of this embodiment, the im2col conversion using the digital circuit 50 performs a conversion on only a part of the input matrix x, and the signal x'∈R 1×T This signal x' is converted to [a specific format]. This signal x' is introduced to the optical modulator 203 via the DAC 60, which generates the same optical signal x' for all N wavelengths obtained from the comb light source 201.

[0064] The optical signal is dispersed using a linear wavelength-dispersive element 209-1 (for example, an element approximated by second-order wavelength dispersion such as an optical fiber, a periodic grating, or a photonic crystal) to obtain an optical signal ui corresponding to im2col.

[0065] Figure 7 compares the details of the conversion mechanism. The im2col conversion in the digital domain described in Embodiments 1 and 2 corresponds to individually modulating each lane of the generated signal u to each wavelength channel (see Figure 7(a)). Here, the signal between each lane corresponds to a signal delayed by p time steps from the adjacent lane. The time step p is a value related to the sliding window of the convolution filter, and typically p is 1.

[0066] The digital circuit of this disclosure executes im2col on only one lane to generate x'. x' is modulated collectively into all wavelength signals and converted into an optical signal (see Figure 7(b)). In a wavelength-dispersive element, the signal is delayed linearly with respect to the wavelength according to the dispersion value D, so by appropriately setting D, it is possible to generate a signal delayed by a p time step.

[0067] This makes it optically possible to convert the input signal x' to the input matrix u, and enables Depthwise convolution using only a single optical modulator 203 and DAC 60 for each channel. Here, it is desirable that the dispersion D satisfies the relationship in equation (1) such that it has a delay of p time step with respect to adjacent wavelengths.

[0068]

[0069] Here, "baudrate" is the modulation rate of the DAC, and Δλ is the difference from adjacent wavelengths (corresponding to the optical frequency comb spacing).

[0070] (Fourth Embodiment of Optical Implementation for Depthwise Convolution Operations) Figure 8 shows a fourth embodiment for Depthwise convolution operations in an optical computing device of the present disclosure. The optical computing device 10C in this embodiment exemplifies a configuration in which an optical modulation unit 20 including a digital circuit 50, an optical computing unit 30-1, and an optical receiving unit 40. In Figure 8, as a non-limiting example, the digital circuit 50 is shown to be mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20. The optical modulation unit 20 may also be connected to the optical computing unit 30-1 by a linear wavelength dispersion element 209-1.

[0071] However, in the following description, the optical modulation unit 20 and optical receiving unit 40, which have the same configuration as the optical computing device 10B of the third embodiment described above, will not be explained, and only the optical computing unit 30-1, which has a different configuration, will be described.

[0072] As shown in Figure 8, the optical computing unit 30-1 includes one or more optical computing circuits 70a to m, each composed of a wavelength filter using a microring resonator (MRR) array.

[0073] The optical signal ui output from the optical modulation unit 20 is input to the input port of the microring resonator (MRR). Subsequently, the transmittance of the optical signal ui to the output port (corresponding to the weighting coefficient ω of W) is controlled by the MRR array 70, which is arranged for each wavelength λj, and the output light, which is the result of the calculation, is output to the PD401. Since this configuration is realized with a waveguide structure, it has the excellent advantage of being able to reduce the device size compared to the spatial optical type method described above.

[0074] (Embodiment 5 of optical implementation for Depthwise convolution operation) Wavelength-dispersive elements and comb light sources having dispersion values ​​D lead to larger size and higher cost of the optical computing device. In this embodiment, a fifth implementation form that performs Depthwise convolution without using the above-mentioned wavelength-dispersive elements and comb light sources will be described with reference to Figure 9.

[0075] In this embodiment, the optical computing device 10D includes, as an example, an optical modulation unit 20 including a digital circuit 50, an optical computing unit 30-1, and an optical receiving unit 40. In Figure 9, as a non-limiting example, the digital circuit 50 is shown mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0076] However, in the following description, the optical receiving unit 40, which has the same configuration as the optical computing device 10C of the fourth embodiment described above, will not be described, and the optical modulation unit 20 and optical computing unit 30-1, which have different configurations, will be described.

[0077] The optical modulation unit 20 in this embodiment comprises a digital circuit 50, a DAC 60, a laser light source 201 that oscillates at a single wavelength (e.g., λ1), and one or more optical modulators 203a to m.

[0078] The optical modulation unit 20 performs only the conversion from the input signal x to the optical signal x' described above, and outputs the optical signal to the optical processing unit 30-1. The optical processing unit 30-1 includes delay interference optical circuits 80a to m, which are composed of one or more delay interference systems.

[0079] Within the delay interference optical circuit 80, which is composed of a multi-stage (for example, L-stage) delay interference system, a convolution operation in the time direction is performed based on the branching ratio set for each delay line arm and the weighting corresponding to the phase modulation amount, leading to the output.

[0080] Referring to Figure 10, the principle of convolution operation within the wire interference optical circuit 80 will be explained. The delay interference optical circuit 80 comprises a variable splitter 811 and L delay interference systems 833-1 to 833-L (where L is an integer of 2 or more) optically cascaded to the variable splitter 811. Each of the delay interference systems 833-1 to 833-L may be referred to as a "unit cell".

[0081] Through calculations performed by the wire interference optical circuit 80, the variable splitter 811 splits the output light from the optical modulator 203 at a variable splitting ratio, and inputs each of the split beams to the first-stage delay interference system 833-1. For example, the variable splitter 811 has two input ports and two output ports, and the output light from the optical modulator 203 is input to one of the two input ports (the upper input port in the example in Figure 10).

[0082] The variable splitter 811 may be configured using a Mach-Zehnder interferometer (MZI), and the splitting ratio of the light output from the two output ports is varied by the phase value of a phase shifter 801 located on at least one of the two arm waveguides forming the MZI.

[0083] Each of the delay interference systems 833 has, exemplarily, two input ports and two output ports. The two input ports are optically coupled to the two output ports of the preceding delay interference system, and the two output ports are optically coupled to the input ports of the subsequent delay interference system, respectively.

[0084] The optical receiver 401 is optically coupled to one of the two output ports of the final stage (L stage) delay interference system 833-L (for example, the upper output port in Figure 10(a)) and receives the output light from that output port.

[0085] Furthermore, each of the delay interference systems 833 includes, as an example, a first MZI with asymmetrical two-arm waveguide lengths and a second MZI with symmetrical two-arm waveguide lengths. The first MZI, for example, delays one of two branched beams input from a variable splitter 811 or a preceding delay interference system by a delay difference θ and interferes with the other beam. The second MZI, for example, branches the output beam of the first MZI at a variable branching ratio and outputs it to the next stage delay interference system.

[0086] 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 803 or 805. At least one of the arm waveguides of the first MZI includes, for example, an optical delay line for providing a delay difference θ between the arm waveguides.

[0087] With this configuration, in the wire-extending interference optical circuit 80, the optical branching of the branching ratio according to the phase value of the phase shifter 803 and / or 805, and the application of a delay difference θ to the branched light are repeated a number of times corresponding to the number of stages in the delay interference system 833.

[0088] Here, the transfer function in the z-transform for the wire interference optical circuit 80 can be expressed, for example, by the equivalent circuit (lattice filter) in Figure 10(b) and the following equation (2). In equation (2), j = 1, 2, ..., L, and the coefficient kj representing the branching ratio of each stage is given by equation (3). In Figure 10(b), "E0" represents the electric field strength signal corresponding to the input signal x.

[0089]

[0090] When j=1, A0(z) and B0(z) are determined in the preceding variable splitter 811 by the branching ratio k0 corresponding to the phase value of the phase shifter 801, and are given, for example, by the following equations (4) and (5).

[0091]

[0092] Here, in equation (2), φ j and ψ j These represent the phase values ​​of the phase shifters 803 and 805 of each stage, respectively. -1 This represents a time shift, and its unit delay time is equal to the delay difference θ in each delay interference system 833.

[0093] In Figure 10(b), in the j-th delayed interference system 833-j, the signal input from the lower output port of the previous stage (stage j-1) to the lower input port is delayed by an amount equivalent to the delay difference θ and multiplied by a coefficient kj. The signal multiplied by the coefficient kj 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 (stage j+1).

[0094] Furthermore, the signal input from the upper output port of the previous stage to the upper input port is multiplied by a coefficient kj, then added to the signal input from the lower output port of the previous stage (stage j-1) to the lower input port, which has undergone a delay equivalent to the delay difference θ, and then input from the lower output port to the lower input port of the next stage (stage j+1).

[0095] When the delay difference θ is matched to the sampling rate of the input signal x(t), in the equivalent circuit illustrated in Figure 10(b), the z-transform X(z) of the input signal x(t) can be transformed by Y(z) = AL(z)X(z) and Y(z) = BL(z)X(z) for the two output ports of the final stage (L stage).

[0096] The former, Y(z) = AL(z)X(z), represents the transformation with respect to the upper output port in Figure 10(b), for example, while the latter, Y(z) = BL(z)X(z), represents the transformation with respect to the lower output port in Figure 10(b). Here, if the kernel in the convolutional filter is c = [c0, c1, c2, ..., cL], its z-transform is expressed by the following equation (6).

[0097]

[0098] The parameter φ of each step is such that AL(z) = C(z) j and ψ j By controlling this, any convolution operation can be performed using this optical circuit. By arranging these in parallel as shown in Figure 9, depthwise convolution calculations can be performed.

[0099] Next, the parameter φ for obtaining the desired kernel filter characteristic C(z) j and ψ j The method for determining this will be explained. To simplify the explanation, the phase value of the phase shifter 803 at each stage will be φ j Let's consider the case where we set it to =0 and obtain a real kernel filter C(z). In this case, equation (2) above can be transformed into equation (7) below, based on equations (8) and (9) below.

[0100]

[0101] Equation (7) can be separated into elements P of a real product expressed by the following equation (10) and elements of a matrix-vector product expressed by the following equation (11).

[0102]

[0103] The matrix-vector product term represented by equation (11) determines the characteristics of the kernel filter, and the real product P represented by equation (10) determines the gain or attenuation of the filter. Therefore, by transforming equation (6) into the form of equation (12) below, the parameters of the desired kernel filter can be converted to the parameters of a grid filter.

[0104]

[0105] For parameter transformation, for example, a method using simultaneous equations can be applied. For instance, by expanding equation (12) and comparing it with equation (6), a system of simultaneous equations can be obtained, and the parameters can then be determined based on this system of equations.

[0106] As a non-restrictive example of parameter determination methods, we will describe the case where the kernel c = [c0, c1, c2] is set in a lattice filter with L = 2 stages. In this case, the target filter response is c(z) = c0 - c1z -1 -c2z -2 This is equivalent to being.

[0107] From equation (12), the response of the first stage is A1(z) = A0 + q1B0z -1 , B1(z)=q1A0−B0z -1 Therefore, the second stage response is A2(z) = A1(z) + q2B1(z)z -1 =A0+(q1B0+q2q1A0)z -1 +q2B0z -2 This is the result. Comparing C(z) and A2(z), we obtain the following relationships (13) to (15).

[0108]

[0109] By solving equations (3) to (5) and equations (13) to (15) simultaneously, the phase parameter ψj of each filter stage can be determined. In this example, the case where the number of stages is L = 2 was explained, but it can be extended to any dimension (number of stages).

[0110] While solving simultaneous equations generally tends to be computationally intensive, imposing constraints on the transfer function, for example, allows for the use of a successive method, thereby reducing the computational cost of parameter determination. For instance, by imposing the constraint A0 = B0 = 1 / √2 (ψ0 = 45°), filter parameters can be efficiently determined using a successive method.

[0111] In this case, due to the relationship A0 = c0, the settable value of c0 is fixed to 1 / √2, but the kernel to be set can be, for example, c' = [1 / √2, c1 / c0, c2 / c0, ..., c L By changing it as shown in [ / c0], it is possible to configure a kernel filter with the same characteristics as when no constraints are applied.

[0112] The output intensity becomes 1 / √2 * c0 times the original intensity, but this can be compensated for in subsequent processing. Under these constraints, the output response B L (z) is A L Since the relationship is a transpose-conjugate relationship with respect to (z), it is possible to sequentially determine the coefficients based on this relationship.

[0113] The following describes a non-restrictive example of the calculation procedure. From equation (12), the L-stage filter response A L z in (z) -L The coefficient multiplied by is q L Therefore, by comparing equation (4) and equation (12), the parameter of the final stage (L) of the lattice filter is q L = c L It can be converted like this.

[0114] Next, q L-1 To make a similar decision regarding A, L-1 Let's consider the response of (z). From equation (10), the (j-1)th filter response A j-1 (z) is the j-th filter response A j (z), Bj When expressed using (z), it can be represented as shown in equation (16) below.

[0115]

[0116] A' L (z)=C(z), B' L (z) = C † Substituting (z) into equation (16), we get A' L-1 (z) is operable. Note that † represents the conjugate transpose of the response, and in the case of real numbers, the transpose matrix C † (z) = c L +c L-1 z -1 +c L-2 z -2 +...+c0z -L That is the case.

[0117] The result obtained z -1 The coefficient of the term is the coefficient q of the lattice filter. L-1 This corresponds to the same calculation for L-2 and beyond, where the parameter q j It is possible to determine these sequentially. This method can be performed by solving equation (16) L times, thus reducing the amount of computation compared to the simultaneous equation solution method described above. Although the above example shows calculations using real numbers, parameter determination using complex numbers can also be performed using the same procedure.

[0118] (Embodiment 1 of Optical Implementation for Pointwise Convolution Operation) Below, the Pointwise convolution operation in the method of implementing the above-described separable convolution operation using optical elements will be explained. As mentioned above, the Pointwise convolution operation is equivalent to a Cin × Cout dimension matrix operation. Therefore, the configuration of the optical implementation for the Depthwise convolution operation described above can be used as is.

[0119] Figure 11 shows a first embodiment for Pointwise convolution in an optical computing device of the present disclosure. In this embodiment, the optical computing device 90 includes, exemplary, an optical modulation unit 20 including a digital circuit 50, an optical computing unit for Pointwise convolution (hereinafter referred to as the optical computing unit) 30-2, and an optical receiving unit 40. In Figure 11, as a non-limiting example, the digital circuit 50 is shown to be mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0120] (Optical Modulation Unit) The optical modulation unit 20 in this embodiment includes a digital circuit 50, a DAC 60 that converts an input signal to an analog signal, one or more light sources 201-1 to n, and one or more optical modulators 203a to n.

[0121] The optical modulation unit 20 may have a WDM-MUX (e.g., AWG) to send the optical signals ui output from one or more optical modulators 203a to n to the optical computing unit 30-2. The optical modulation unit 20 may also be connected to the optical computing unit 30-2 via an optical fiber array or the like.

[0122] The digital circuit 50 processes the input signal x and outputs it to the optical modulator 203 via the DAC 60. In the Pointwise convolution operation of this disclosure, the digital circuit 50 does not perform operations like the Depthwise convolution operation described above, so operations such as im2col are unnecessary. The digital circuit 50 may be, for example, an FPGA or an ASIC.

[0123] The light source 201 emits, for example, a predetermined wavelength (λ1 to λ1). N A continuous wave (CW) laser beam having the following characteristics is output to each optical modulator 203. In Figure 11, a light source array composed of multiple light sources is shown as a non-limiting example, but it is also possible to output multiple light rays with different wavelengths using a comb light source or the like.

[0124] Each optical modulator 203 modulates, for example, the laser light input from the light source 201 with the input signal x from the digital circuit 50 and the DAC 60. Specifically, signal modulation is performed using the optical modulator 203 for different wavelengths, such that the i-th input signal xi is encoded into each wavelength channel λi. As non-limiting examples, the optical modulator 203 can 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 phase.

[0125] (Optical Computing Unit) The optical computing unit 30-2 comprises one or more optical computing circuits 301a to m having the same configuration depending on the number of input signals. Each optical computing circuit 301 also includes, as an example, a wavelength demultiplexer (WDM-DEMUX) 305, one or more wavelength division multiplexing (WDM) type optical matrix computers 303, and a wavelength multiplexer (WDM-MUX) 307.

[0126] The optical computing unit 30-2 includes an optical element 311 (for example, a beam splitter) for arbitrarily branching the optical signal ui, which is sent in bulk from the optical modulation unit 20 using WDM-MUX, to multiple WDM-type matrix computing units.

[0127] Each optical signal ui from the optical modulation unit is input to the wavelength demultiplexer 305 and wavelength-separated. Each wavelength-separated optical signal ui is used to perform matrix calculations in parallel using one or more WDM type optical matrix computers 303 (e.g., variable optical attenuators (VOAs)) located in the optical computing unit 30-2.

[0128] Each wavelength-separated optical signal is subjected to optical attenuation (or amplification or phase modulation) corresponding to the weight of each wavelength by a variable optical attenuator (VOA), then wavelength-combined again by a wavelength multiplexer (WDM-MUX) 307, and output to an optical receiver 401 (e.g., a photodetector (PD)).

[0129] The WDM-type optical matrix computer 303 in this embodiment has the same configuration as the first and third embodiments for Depthwise convolution operations described above. Furthermore, the WDM-type optical matrix computer 303 can be realized with the components described with reference to Figure 4.

[0130] (Optical receiving unit) The optical receiving unit 40 comprises at least one or more PDs 401a to m. The optical receiving unit 40 may also optionally or additionally comprise one or more ADCs 403 and digital circuits 405.

[0131] The optical signal output from the optical computing unit 30-2 can be input to PD 401. Optionally, the output yi from each PD can be converted into a digital signal by ADC 403 and subtracted by a digital circuit 405 (e.g., FPGA or ASIC).

[0132] (Embodiment 2 of Optical Implementation for Pointwise Convolution Operation) Figure 12 shows a second embodiment for Pointwise convolution operation in an optical computing device of the present disclosure. In this embodiment, the optical computing device 90A includes, exemplary, an optical modulation unit 20 including a digital circuit 50, an optical computing unit 30-2, and an optical receiving unit 40. In Figure 12, as a non-limiting example, the digital circuit 50 is shown to be mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0133] However, in the following description, the optical modulation unit 20 and optical receiving unit 40, which have the same configuration as the optical computing device 90 of the first embodiment described above, will not be explained, and only the optical computing unit 30-2, which has a different configuration, will be described.

[0134] As shown in Figure 12, the optical computing unit 30-2 includes one or more optical computing circuits composed of wavelength filters using microring resonator (MRR) arrays 70a to m.

[0135] The WDM optical signal ui output from the optical modulation unit 20 is input to the input port of the microring resonator (MRR). Subsequently, the transmittance of the WDM optical signal ui to the output port (corresponding to the weighting coefficient ω of W) is controlled by the MRR array 70, which is arranged for each wavelength λj, and the resulting optical output is output to the PD401. This configuration has the excellent advantage of being realized with a waveguide structure, which allows for a smaller device size compared to the spatial optical method described above.

[0136] (Embodiment 3 of Optical Implementation for Pointwise Convolution Operation) Figure 13 shows a third embodiment for Pointwise convolution operation in an optical computing device of the present disclosure. In this embodiment, the optical computing device 90B exemplary comprises an optical modulation unit 20 including a digital circuit 50, an optical computing unit 30-2, and an optical receiving unit 40. In Figure 12, as a non-limiting example, the digital circuit 50 is shown to be mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0137] However, in the following description, the optical modulation unit 20 and optical receiving unit 40, which have the same configuration as the optical computing units 90 and 90A of the first and second embodiments described above, will not be described, and only the optical computing unit 30-2, which has a different configuration, will be described.

[0138] The optical computing unit 30-2 in this embodiment may be configured using a Mach-Zehnder interferometer (MZI) array 80-1 as shown in Reference 6. The MZI array 80-1 consists of a plurality of programmable MZIs, each MZI having, as a non-limiting example, a thermo-optic phase shifter (θ) positioned between two 50% evanescent directional couplers, followed by another phase shifter (φ). The branching ratio of the MZI is controlled by an internal phase shifter, and the differential output phase is controlled by an external phase shifter (see Reference 6).

[0139] (Embodiments for Integration) It is also possible to integrate the above-described Depthwise and Pointwise convolution calculations into a single device, and specific embodiments of these are shown in Figures 14 and 15. Note that the module configurations shown in Figures 14 and 15 are non-limiting examples, and it should be understood that any combination of the above-described optical computing circuits can be implemented.

[0140] Figure 14 shows an example of an integrated optical computing module 100A, which performs a wavelength-direction depthwise convolution operation using a comb light source and an MRR array, followed by a spatial-direction pointwise convolution operation using an MZI array.

[0141] The integrated optical computing module 100A includes, as an example, an optical modulation unit 20 including a digital circuit 50, an optical computing unit 30-1 for Depthwise convolution, an optical computing unit 30-2 for Pointwise convolution, and an optical receiving unit 40. In Figure 14, as a non-limiting example, the digital circuit 50 is shown mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0142] In one or more optical modulators 203, N wavelength signals are extracted from the N' wavelength comb light rays generated from the comb light source 201, for example, using a wavelength filter, and the signal matrix ui of each is converted into an optical signal.

[0143] The optical signal is dispersed using a linear wavelength-dispersive element 209-1 (for example, an element approximated by second-order wavelength dispersion such as an optical fiber, a periodic grating, or a photonic crystal) to obtain a WDM optical signal ui corresponding to im2col.

[0144] The WDM optical signal ui is input to the input port of the MRR array 70. Subsequently, the WDM optical signal ui is processed by the MRR array 70, which is arranged for each wavelength λj, controlling the transmittance to the output port (corresponding to the weighting coefficient ω of W), and the resulting output light is output to the Pointwise convolution optical calculation unit 30-2.

[0145] The Pointwise convolution optical computing unit 30-2 can be configured, as a non-limiting example, using the Mach-Zehnder interferometry (MZI) array 80-1 described in Reference 6. The Pointwise convolution optical computing unit 30-2 uses the MZI array 80-1 to perform a convolution operation with a kernel size of "1" on the obtained intermediate output. Unlike Depthwise convolution, this operation also performs a convolution operation in the input channel direction, but since the kernel size is 1, no convolution is performed in the spatial direction.

[0146] The optical receiving unit 40 includes at least one or more PD401a to m. Optical signals output from the optical processing unit 30-2 can be input to the PD401a to m.

[0147] Furthermore, the optical receiving unit 40 may include one or more ADCs 403 and digital circuits 405. The output from each PD is converted into a digital signal by the ADC 403 and can be subtracted by the digital circuit 405 (e.g., FPGA or ASIC).

[0148] Figure 15 shows an optical computing module 100B integrated with a different configuration from the optical computing module 100A described above, and illustrates an example in which an optical computing unit 30-1 for depthwise convolution operations using a delayed interference optical circuit and an optical computing unit 30-2 for spatial pointwise convolution operations using an MZI array are combined.

[0149] The integrated optical computing module 100B includes, as an example, an optical modulation unit 20 including a digital circuit 50, an optical computing unit 30-1 for Depthwise convolution, an optical computing unit 30-2 for Pointwise convolution, and an optical receiving unit 40. In Figure 14, as a non-limiting example, the digital circuit 50 is shown mounted inside the optical modulation unit 20, but the digital circuit 50 may be mounted outside the optical modulation unit 20.

[0150] The optical modulation unit 20 in this embodiment includes a digital circuit 50, a DAC 60 that converts an input signal to an analog signal, a laser light source 201 that oscillates a single wavelength (e.g., λ1), and one or more optical modulators 203a to m.

[0151] The optical modulator 203 modulates the laser light input from the light source 201 with the input signal x, for example, and transmits the WDM optical signal ui to the optical processing unit 30-1. The WDM optical signal ui undergoes a Depthwise convolution operation using variable splitters and one or more delay interference systems arranged in the delay interference optical circuits 80a to m, and outputs the resulting output light to the Pointwise convolution operation optical processing unit 30-2.

[0152] The Pointwise convolution optical computing unit 30-2 may be configured using the Mach-Zehnder interferometer (MZI) array 80-1 described in Reference 6. The optical receiving unit 40 includes at least one or more PDs 401a to m. Optical signals output from the optical computing unit 30-2 can be input to the PDs 401a to m.

[0153] Both configurations offer significant advantages, such as the ability to integrate convolutional arithmetic mechanisms, which were previously difficult to implement, within a single chip.

[0154] (Additional Considerations) The foregoing description of embodiments of the present invention is presented for illustrative purposes only and is not intended to be exhaustive or to limit to the exact forms disclosed. Those skilled in the art will understand that many modifications and changes are possible in light of the above disclosure.

[0155] Finally, the language used herein has been selected primarily for readability and instructional purposes, and may not be selected to describe or limit the subject matter of the invention. Therefore, the scope of the invention is intended to be limited by the appended claims, not by this detailed description. Accordingly, the disclosure of embodiments of the invention is intended to illustrate, not limit, the scope of the invention as described in the claims.

[0156] This disclosure is useful, for example, for neural networks based on convolutional operations.

[0157] 10, 10A-D, 90, 90A-B Optical computing unit for convolution operations 20 Optical modulation unit 30-1, 30-2 Optical computing unit 40 Optical receiving unit 50 Digital circuit 60 DAC 70a-m MRR array 80a-m Delay line interference optical circuit 80-1 MZI array 100A, 100B Optical computing unit module 201 Light source 203a-m Optical modulator 205a-m, 305a-m Wavelength demultiplexer (WDM-DEMUX) 207a-m, 307a-m Wavelength multiplexer (WDM-MUX) 209, 209-1 Optical fiber array 211 Optical waveguide 301a-m Optical computing circuit 303a-m WDM type matrix calculator 303-1 Spatial Light Modulator (SLM) 305-1 Diffraction Grating 308a, b First Lens 309a, b Second Lens 311 Optical Elements 401a-m Optical Receiver (PD) 403 ADC 405 Digital Circuits 801, 803, 805 Phase Shifter 811 Variable Splitter 833, 833-1-L Delay Interferometer

Claims

1. An optical computing device comprising: a light source that emits laser light of one or more wavelengths; a first digital circuit that converts an electrical input signal into a matrix signal; one or more optical modulators that convert the matrix signal into an optically modulated signal; one or more optical computing circuits that perform an optical convolution operation on the output light of the optical modulator, wherein the optical computing circuit is one of the following: an optical computing circuit including a diffraction grating and a spatial light modulator; an optical computing circuit including a microring resonator (MRR array); an optical computing circuit including a variable splitter and a plurality of cascaded delay interference systems; or an optical computing circuit including a Mach-Zehnder interferometer (MZI) array; and one or more optical receivers that receive the output light of the optical computing circuit and convert it into an electrical signal.

2. The optical computing device according to claim 1, wherein the first digital circuit performs a rearrangement operation on the input signal to convert it into the matrix signal.

3. The optical computing apparatus according to claim 1, comprising a wavelength-dispersing element disposed between the optical modulator and the optical computing circuit, which linearly delays the output light of the optical modulator with respect to wavelength.

4. The optical computing device according to claim 1, further comprising a second digital circuit that receives the electrical signals from one or more optical receivers and performs subtraction processing.

5. The optical computing device according to claim 1, comprising a connection between a first optical computing circuit which is any one of the optical computing circuits and a second optical computing circuit which is any one of the optical computing circuits.

Images