Optical convolution computing device

Abstract

The present application provides an optical convolution computing device, wherein one output of a light source module loads a first set of vector data to a target light source including a single-wavelength light source to output an intermediate light signal, and another output outputs an optical carrier without a modulated signal added thereto according to the target light source; a programmable FIR filter loads a second set of vector data as a weight vector obtained based on tap self-configuration to the intermediate light signal; and a detection module detects the optical power corresponding to the optical signal and the optical carrier without a modulated signal added thereto, which is used to represent the result of the vector convolution computing. The present application can only use a single-wavelength light source, which can reduce the difficulty of on-chip integration while ensuring the rate, and can reduce the manufacturing cost of the device; it can realize larger-scale convolution computing, and can improve the efficiency of image processing using a convolutional neural network, and can also effectively reduce the computing power consumption of the computer, and realize high-speed, low-power, and highly integrated chip-level neural network optical computing.

Description

Technical Field

[0001] The present application relates to the field of optical computing technology, and in particular to an optical convolution computing device. Background Art

[0002] Convolutional neural network (CNN) is a powerful type of artificial neural network used for image recognition, etc. It can significantly reduce the number of weight parameters while improving the accuracy of the neural network, reducing the difficulty of neural network training and enabling it to develop into a deeper neural network. The core operation of convolutional neural network, convolution, when implemented in a computer, is to decompose the convolution operation into multiple matrix multiplication operations. For the parallel operation structure of convolution operation and matrix multiplication, the computer's von Neumann architecture consumes a lot of computing power.

[0003] Optical computing uses the characteristics of light itself, with high parallelism, low power consumption, and fast computing speed, making it easy for optical chips to perform parallel computing. However, current implementation schemes, such as MZI grids and MRR weight libraries, require a calibration process before configuring weights, and then reconfigure. And because it uses multiple modulators or multi-wavelength light sources, it limits the chip integration and hinders the application of large-scale neural networks. Therefore, existing optical computing methods have the problems of low speed and high power consumption, which not only affects the efficiency of image processing using convolutional neural networks, but also greatly consumes the computing power of computers. Summary of the invention

[0004] In view of this, an embodiment of the present application provides an optical convolution computing device, including: a light source module, a programmable FIR filter and a detection module integrated on a chip;

[0005] The light source module includes two outputs, one for outputting an intermediate optical signal loaded with a first set of vector data according to a target light source, and the other for outputting an optical carrier without adding a modulation signal according to the target light source, wherein the target light source includes: a single wavelength light source;

[0006] The programmable FIR filter is used to receive the intermediate optical signal, load the second set of vector data obtained by self-configuration of the taps of the programmable FIR filter as a weight vector, output each discrete optical signal and generate a corresponding target optical signal, wherein each of the optical signals is used to represent the product between an element in the first set of vector data and an element in the second set of vector data;

[0007] The detection module is used to receive the target optical signal and the optical carrier without the modulated signal, and detect and obtain optical power data corresponding to the optical signal and the optical carrier without the modulated signal for representing the vector convolution calculation result.

[0008] Further, the light source module includes: a single wavelength light source module;

[0009] The single-wavelength light source module comprises: a laser, a first beam splitter and a first modulator connected in sequence;

[0010] Wherein, the laser is used to emit a single-channel single-wavelength optical carrier as the single-wavelength light source;

[0011] The first beam splitter is used to split the single-wavelength optical carrier of a single channel into two channels, so as to obtain one channel of the single-wavelength optical carrier directly output without adding a modulation signal and another channel of the single-wavelength optical carrier transmitted to the first modulator;

[0012] The first modulator is used to load a first set of vector data onto the single-wavelength optical carrier transmitted by the first beam splitter, and output an intermediate optical signal accordingly.

[0013] Further, the target light source further comprises: a multi-wavelength light source; correspondingly, the light source module comprises: a multi-wavelength light source module;

[0014] The multi-wavelength light source module comprises: an optical frequency comb source, a second beam splitter, a first wavelength division multiplexer, a modulator array and a wavelength division multiplexer connected in sequence;

[0015] Wherein, the optical frequency comb source is used to emit a single-channel multi-wavelength optical carrier as the multi-wavelength light source;

[0016] The second beam splitter is used to split the single-channel multi-wavelength optical carrier into two channels, so as to obtain one channel of the multi-wavelength optical carrier directly output without adding a modulation signal and another channel of the multi-wavelength optical carrier transmitted to the wavelength division multiplexer;

[0017] The first wavelength division multiplexer is used for dividing the multi-wavelength optical carrier transmitted by the second beam splitter into multiple single-wavelength optical carriers;

[0018] The modulator array is used to load each first group of vector data to multiple single-wavelength optical carriers respectively, and to output the single-wavelength intermediate optical signals corresponding to each of the single-wavelength optical carriers respectively;

[0019] The wavelength division multiplexer is used to multiplex the single-wavelength intermediate optical signals into one multi-wavelength intermediate optical signal.

[0020] Further, the modulator array includes: a plurality of second modulators, and the number of the second modulators is the same as the number of wavelengths of the multi-wavelength optical carrier;

[0021] Each of the second modulators is used to load one of the first groups of vector data onto a single wavelength optical carrier transmitted by the first wavelength division multiplexer.

[0022] Further, the programmable FIR filter comprises: a third beam splitter, a phase shifter array, a delay line array and a beam combiner group connected in sequence; the third beam splitter is connected to the light source module, and the beam combiner group is connected to the detection module;

[0023] The third beam splitter is used to split the intermediate optical signal into multiple paths to obtain multiple optical signals, wherein the power division ratio of each path is adjustable to self-configure the amplitude of the tap coefficient of the programmable FIR filter;

[0024] The phase shifter array is used to adjust the phase of each path of the multi-path optical signal output by the third beam splitter to self-configure the phase of the tap coefficient, thereby realizing the loading of the second set of vector data on the intermediate optical signal through the tap coefficient;

[0025] The delay line array is used to discretely delay the multi-path optical signals output by the phase shifter array to obtain multi-path delay-weighted optical signals;

[0026] The combiner group is used to combine the multi-path time-delay weighted optical signals output by the delay line array into one path to output a corresponding target optical signal, wherein the optical field of the target optical signal is the superposition result of the optical fields of the multi-path time-delay weighted optical signals.

[0027] Furthermore, the third beam splitter adopts a programmable MZI array;

[0028] The programmable MZI array includes a plurality of sequentially connected MZI units, and the first MZI unit includes one MZI; the number of MZIs in the other MZI units is twice the number of MZIs in the previous MZI unit;

[0029] Except for the last MZI unit, each of the MZIs in the other MZI units is connected to two MZIs in the next MZI unit.

[0030] Furthermore, the beam combiner group includes a plurality of beam combiner units connected in sequence, and the last beam combiner unit includes a first beam combiner; the number of first beam combiners in the other beam combiner units is twice the number of first beam combiners in the next beam combiner unit;

[0031] Except for the first beam combiner unit, every two first beam combiners in other beam combiner units are connected to a first beam combiner in the next beam combiner unit.

[0032] Further, the phase shifter array comprises: a plurality of first phase shifters, and the number of the first phase shifters is the same as the number of the multi-path optical signals output by the third beam splitter;

[0033] The delay line array includes: a plurality of delay lines, and the number of the delay lines is the same as the number of the multi-path optical signals output by the third beam splitter.

[0034] Further, the detection module includes: a single wavelength detection module;

[0035] The single wavelength detection module comprises: a second phase shifter, a second beam combiner and a first detector connected in sequence; the second beam combiner is connected to the light source module, and the second phase shifter is connected to the programmable FIR filter;

[0036] The second phase shifter is used to receive the target optical signal and change the phase difference of the target optical signal;

[0037] The second beam combiner is used to combine the target optical signal output by the second phase shifter and the optical carrier without the modulated signal output by the light source module into one path to obtain a corresponding composite signal;

[0038] The first detector is used to detect and obtain optical power data of the synthetic signal for representing a vector convolution calculation result.

[0039] Further, the target light source further comprises: a multi-wavelength light source; correspondingly, the detection module comprises: a multi-wavelength detection module;

[0040] The multi-wavelength detection module comprises: a third phase shifter, a third beam combiner, a second wavelength division multiplexer and a detector array connected in sequence; the third beam combiner is connected to the light source module, the third phase shifter is connected to the programmable FIR filter, and a plurality of second detectors are provided in the detector array;

[0041] The third phase shifter is used to receive the target optical signal and change the phase difference of the target optical signal;

[0042] The third beam combiner is used to combine the target optical signal output by the third phase shifter and the optical carrier without the modulated signal output by the light source module into one channel, so as to obtain a corresponding single-channel multi-wavelength synthetic signal;

[0043] The second wavelength demultiplexer performs demultiplexing processing on the single-channel multi-wavelength composite signal to obtain multiple groups of optical signals which are respectively vector convolution results;

[0044] The detector array is used to detect and obtain optical power data corresponding to each group of optical signals respectively serving as vector convolution results and used to represent the vector convolution calculation results.

[0045] The optical convolution computing device provided by the present application includes: a light source module, a programmable FIR filter and a detection module integrated on a chip; the light source module includes two outputs, one for loading an intermediate optical signal of a first group of vector data according to a target light source output, and the other for outputting an optical carrier without adding a modulation signal according to the target light source, wherein the target light source includes: a single-wavelength light source, and only a single-wavelength light source can be used, which can avoid the dependence on multi-wavelength light sources in the multi-modulators used in the MZI grid and the MRR wavelength division multiplexing, and can reduce the difficulty of on-chip integration on the basis of ensuring the rate, and can reduce the manufacturing cost of the optical convolution computing device; the programmable FIR filter is used to receive the intermediate optical signal, and load the second group of vector data obtained as a weight vector based on the tap self-configuration of the programmable FIR filter, output discrete optical signals and generate corresponding target optical signals, wherein each of the optical signals is used to represent the first group of vectors. The product of an element in the quantity data and an element in the second group of vector data is calculated. When a specified weight vector is implemented through an FIR filter, the FIR filter can be configured without pre-calibrating the FIR filter, and the taps of the FIR filter can be directly configured to implement the configuration of the weight vector. The configuration will not increase with the increase of the number of FIR filter taps (the scale of the weight vector), thereby reducing the complexity of the configuration process of the weight vector, so that it can implement larger-scale convolution calculations and facilitate more complex neural network applications. The detection module is used to receive the target optical signal and the optical carrier without the added modulation signal, and detect and obtain the optical power data corresponding to the optical signal and the optical carrier without the added modulation signal for representing the vector convolution calculation result. Matrix convolution can be implemented through vector convolution, and high-speed, low-power and highly integrated chip-level neural network optical calculations can be implemented, thereby effectively improving the efficiency of image processing by the convolution neural network implemented by the optical convolution calculation device.

[0046] Additional advantages, purposes, and features of the present application will be partially described in the following description, and will become partially apparent to those skilled in the art after studying the following, or may be learned from the practice of the present application. The purposes and other advantages of the present application can be achieved and obtained by the structures specifically pointed out in the specification and the drawings.

[0047] Those skilled in the art will understand that the purposes and advantages that can be achieved by the present application are not limited to the above specific description, and the above and other purposes that can be achieved by the present application will be more clearly understood based on the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the present application, constitute a part of the present application, and do not constitute a limitation of the present application. The components in the drawings are not drawn to scale, but are only for the purpose of illustrating the principles of the present application. In order to facilitate the illustration and description of some parts of the present application, the corresponding parts in the drawings may be enlarged, that is, they may become larger relative to other components in the exemplary device actually manufactured according to the present application. In the drawings:

[0049] Figure 1 Schematic diagram of the overall architecture and logic of the optical convolution computing device in the embodiment of the present application.

[0050] Figure 2 Schematic diagram of the architecture and logic of the single-wavelength light source module in the embodiment of the present application.

[0051] Figure 3 Schematic diagram of the architecture and logic of the multi-wavelength light source module in the embodiment of the present application.

[0052] Figure 4 Schematic diagram of the structure and logic of the programmable FIR filter in the embodiment of the present application.

[0053] Figure 5 Schematic diagram of the structure and logic of the third beam splitter in the embodiment of the present application.

[0054] Figure 6 The diagram is a schematic diagram of the architecture and logic of the combiner group in the embodiment of the present application.

[0055] Figure 7 Schematic diagram of the architecture and logic of the single wavelength detection module in the embodiment of the present application.

[0056] Figure 8 FIG. 1 is a schematic diagram of the architecture and logic of the multi-wavelength detection module in the embodiment of the present application.

[0057] Fig. 9 It is a flowchart of the method for self-configuring FIR tap coefficients in an embodiment of the present application.

[0058] Fig.10 Schematic diagram of the convolution calculation method in the embodiment of the present application.

[0059] Fig.11 Detailed architecture and logic diagram of a complete optical computing system device when a single wavelength light source is used in an embodiment of the present application.

[0060] Fig.12 The following is a schematic diagram of an example of the specific operation process of the convolution calculation in the embodiment of the present application.

[0061] Fig.13This is a first example schematic diagram of using one-dimensional vector convolution to implement two-dimensional matrix convolution in an embodiment of the present application.

[0062] Fig.14 This is a second example schematic diagram of using one-dimensional vector convolution to implement two-dimensional matrix convolution in an embodiment of the present application.

[0063] The figures in the attached drawings are as follows:

[0064] 1. Light source module;

[0065] 11. Single wavelength light source module;

[0066] 111. Laser;

[0067] 112. a first beam splitter;

[0068] 113. A first modulator;

[0069] 12. Multi-wavelength light source module;

[0070] 121. Optical frequency comb source;

[0071] 122. a second beam splitter;

[0072] 123. Wave 1 demultiplexer;

[0073] 124. Modulator array;

[0074] 1241, second modulator;

[0075] 125. Wavelength division multiplexer; 2. Programmable FIR filter;

[0076] 21. The third beam splitter;

[0077] 211. Programmable MZI array;

[0078] 2111, MZI;

[0079] 22. Phase shifter array;

[0080] 221, first phase shifter;

[0081] 23. Delay line array;

[0082] 231. Delay line;

[0083] 24. Beam combiner group;

[0084] 241. a first beam combiner; 3. a detection module;

[0085] 31. Single wavelength detection module;

[0086] 311, second phase shifter;

[0087] 312, second beam combiner;

[0088] 313, the first detector;

[0089] 32. Multi-wavelength detection module;

[0090] 321, third phase shifter;

[0091] 322, the third beam combiner;

[0092] 323, second wave demultiplexer;

[0093] 324. Detector array;

[0094] 3241. The second detector. DETAILED DESCRIPTION

[0095] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the implementation modes and the accompanying drawings. Here, the illustrative implementation modes and descriptions of the present application are used to explain the present application, but are not intended to limit the present application.

[0096] It should also be noted here that in order to avoid obscuring the present application due to unnecessary details, only the structures and / or processing steps closely related to the scheme according to the present application are shown in the accompanying drawings, while other details that are not very relevant to the present application are omitted.

[0097] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0098] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0099] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0100] The MZI (Mach–Zehnder interferometer) grid space division multiplexing method refers to: using a single wavelength light source, utilizing coherent light interference, and combining the SVD algorithm, a certain structure and number of MZI grids, whose transmission matrix can realize any real matrix, and then matrix multiplication can be realized through multiple vector products. When realizing the convolution operation, the data needs to be preprocessed and the convolution operation is realized by matrix multiplication. Then, the MZI grid space division multiplexing method has the following problems: (1) The MZI grid structure contains multiple data input channels, and the number of channels matches the number of required modulators. However, the large size of the on-chip modulator will limit the number of input channels, which limits the integration scale of the MZI grid; (2) When the MZI grid realizes matrix multiplication, the MZI grid needs to be pre-calibrated and then configured. The calibration process will become more complicated as the scale of the MZI grid increases, further limiting the scale of the MZI grid.

[0101] MRR (micro-ring resonator) wavelength division multiplexing method: using a multi-wavelength light source, light sources of different wavelengths carry different data, using a micro-ring working in a resonant state to weight light of different wavelengths, and realizing matrix multiplication through multiple vector products. Matrix convolution can be realized through matrix multiplication, or by introducing a delay line in the MRR to directly realize matrix convolution, but this will greatly increase the complexity of the structure. Then, the problems of the MRR wavelength division multiplexing method are as follows: (1) the use of a multi-wavelength light source increases the complexity of the system and the difficulty of on-chip integration; (2) the MRR still needs to be calibrated before configuring the weight; (3) the MRR works in a resonant state. After the weight is configured, the resonant wavelength drifts due to thermal crosstalk and other problems, and the configured weight parameters change.

[0102] Among them, convolution is the core operation in convolutional neural networks. When implementing multidimensional discrete convolution, that is, the input is a multidimensional array, the convolution kernel is also a multidimensional array, and it is discrete in time. This is also the most common case in convolutional neural networks. Multidimensional discrete convolution can be expressed as:

[0103] y(i, j) = ∑ m ∑ n H(im,jn)X(m,n) (1)

[0104] In formula (1), H is the weight matrix, X is the input matrix, m and n represent m rows and n columns; i and j are sliding steps; y(i, j) represents the output under the sliding step; H(im, jn) represents the value of the weight matrix corresponding to the im row and jn column; X(m, n) represents the value of the input matrix corresponding to the m row and n column.

[0105] In the AI ​​field, the implementation of convolution omits the operation of flipping the convolution kernel. In fact, it implements the cross-correlation operation. The cross-correlation operation and convolution operation are collectively referred to as convolution. One-dimensional convolution can be expressed as:

[0106] y(k)=∑ m h(mk)x(m) (2)

[0107] In formula (2), k is the sliding step size; h(mk) represents the value of the weight vector at the mkth position.

[0108] Integrated optical circuit refers to an integrated optical system formed by planarizing and miniaturizing traditional discrete optical devices such as light sources, optical couplers, modulators, optical switches, etc. Compared with integrated circuits, the optical wave carrier frequency of integrated optical circuits is more than 1,000 times higher than that of electronics. In addition to processing information in one-dimensional time sequence, integrated optical circuits also have the ability to process information in parallel in space. The high information capacity and multi-dimensional information processing capabilities of integrated optical circuits make them more suitable for high-speed parallel computing than integrated circuits. However, the current schemes for implementing parallel convolution computing with integrated optical circuits do not fully utilize the high-speed multi-dimensional information processing capabilities of integrated optical circuits, and are limited by the current manufacturing process of integrated optical circuits, such as the difficulty in realizing integrated multi-wavelength light sources, the large size of some optical devices such as modulators, and the chip calibration problems caused by the impact of the manufacturing tolerance of integrated optical circuits on the performance of optical chips.

[0109] Based on this, the embodiment of the present application provides an optical convolution computing device, which is a self-configuring optical chip for implementing the convolution computing method of the convolutional neural network, which can avoid the dependence on multi-wavelength light sources in the multi-modulators and MRR wavelength division multiplexing used in the MZI grid, and can realize matrix convolution through vector convolution. An algorithm based on this structure is proposed, which can be used to self-configure weight parameters, avoid the process of first calibrating and then configuring the MZI grid and MRR weight configuration process, and directly configure the weight parameters. Only a single-wavelength light source and a single modulator can be used to reduce the difficulty of on-chip integration and reduce costs while ensuring the rate. Therefore, the core part of the optical convolution computing device provided in the embodiment of the present application is a programmable FIR filter. When the specified weight vector is implemented through the FIR filter, the taps of the FIR filter can be directly configured to implement the configuration of the weight vector without pre-calibrating the FIR filter. The complexity of the configuration process will not increase with the increase in the number of FIR filter taps (the scale of the weight vector), so that it can realize larger-scale convolution calculations, which is convenient for more complex neural network applications. In practical applications, multi-wavelength light sources can also be used to simultaneously calculate the convolution operations of multiple groups of vectors to further improve the calculation speed.

[0110] The details are described in detail through the following multiple embodiments.

[0111] See also Figure 1 The present application embodiment provides an optical convolution computing device, which specifically includes the following contents:

[0112] A light source module 1, a programmable FIR filter 2 and a detection module 3 are integrated on the chip;

[0113] The light source module 1 includes two outputs, one for loading a first set of vector data on a target light source to output an intermediate optical signal, and the other for outputting an optical carrier without adding a modulation signal according to the target light source, wherein the target light source includes: a single wavelength light source.

[0114] That is to say, the light source module 1 comprises two outputs, one for outputting the intermediate optical signal loaded with the first set of data, and the other for outputting the optical carrier without adding the modulation signal.

[0115] The programmable FIR filter 2 is used to receive the intermediate optical signal, and load a second group of vector data as a weight vector obtained based on the tap self-configuration of the programmable FIR filter 2, output discrete optical signals and generate corresponding target optical signals, wherein each of the optical signals is used to represent the product between an element in the first group of vector data and an element in the second group of vector data.

[0116] That is, the programmable FIR filter 2 is used to receive the intermediate optical signal output by the light source module 1, and load the second set of data to be calculated (generally a weight vector), and the loading of the second set of data is completed by the tap of the self-configured FIR filter, and the FIR filter is not calibrated. Each optical signal output by the FIR filter indicates the product of an element of the first set of data and an element of the second set of data.

[0117] The detection module 3 is used to receive the target optical signal and the optical carrier without the modulated signal, and detect and obtain the optical power data corresponding to the optical signal and the optical carrier without the modulated signal for representing the vector convolution calculation result.

[0118] That is, the detection module 3 is used to detect the optical signal output by the programmable FIR filter 2, and the detected optical power result indicates the convolution calculation result information.

[0119] From the above description, it can be seen that the optical convolution computing device provided in the embodiment of the present application can only use a single-wavelength light source, which can reduce the difficulty of on-chip integration while ensuring the rate, and can reduce the manufacturing cost of the device; it can achieve larger-scale convolution calculations, and can improve the efficiency of image processing using convolutional neural networks, and can also effectively reduce the computing power consumption of the computer, and realize high-speed, low-power and highly integrated chip-level neural network optical computing.

[0120] In order to further improve the application reliability and effectiveness of the light source module 1 of the single-wavelength light source, in the optical convolution calculation device provided in the embodiment of the present application, see Figure 2 The light source module 1 in the optical convolution computing device specifically includes the following contents:

[0121] Single wavelength light source module 11; The single wavelength light source module 11 comprises: a laser 111, a first beam splitter 112 and a first modulator 113 connected in sequence;

[0122] Wherein, the laser 111 is used to emit a single-channel single-wavelength optical carrier as the single-wavelength light source;

[0123] The first beam splitter 112 is used to split the single-wavelength optical carrier of a single channel into two channels, so as to obtain one channel of the single-wavelength optical carrier directly output without adding a modulation signal and another channel of the single-wavelength optical carrier transmitted to the first modulator 113;

[0124] The first modulator 113 is used to load a first set of vector data onto the single-wavelength optical carrier transmitted by the first beam splitter 112 and output an intermediate optical signal accordingly.

[0125] Specifically, the single wavelength light source module 11 includes a laser 111, a first beam splitter 112 and a first modulator 113. The laser 111 is used to send a single-channel single-wavelength optical carrier, the first beam splitter 112 is used to split the single-channel optical carrier into two channels, and the first modulator 113 is used to load a first set of vector data on one of the optical carriers.

[0126] In the actual application of the optical convolution computing device provided in the embodiment of the present application, a multi-wavelength light source can also be used to simultaneously calculate the convolution operation of multiple groups of vectors to further improve the operation speed. That is, the target light source also includes a multi-wavelength light source; see Figure 3 The light source module 1 in the optical convolution computing device may also specifically include the following contents:

[0127] A multi-wavelength light source module 12; the multi-wavelength light source module 12 comprises: an optical frequency comb source 121, a second beam splitter 122, a first wavelength division multiplexer 123, a modulator array 124 and a wavelength division multiplexer 125 connected in sequence;

[0128] Wherein, the optical frequency comb source 121 is used to emit a single-channel multi-wavelength optical carrier as the multi-wavelength light source;

[0129] The second beam splitter 122 is used to split the single-channel multi-wavelength optical carrier into two channels, so as to obtain one channel of the multi-wavelength optical carrier directly output without adding a modulation signal and another channel of the multi-wavelength optical carrier transmitted to the wavelength division multiplexer;

[0130] The first wavelength division multiplexer 123 is used to divide the multi-wavelength optical carrier transmitted by the second beam splitter 122 into multiple single-wavelength optical carriers;

[0131] The modulator array 124 is used to load each first set of vector data to multiple single-wavelength optical carriers respectively, and to output the single-wavelength intermediate optical signals corresponding to each of the single-wavelength optical carriers respectively;

[0132] The wavelength division multiplexer 125 is used to multiplex the single-wavelength intermediate optical signals into one multi-wavelength intermediate optical signal.

[0133] In order to further improve the application effectiveness of the modulator array 124, the modulator array 124 specifically includes: multiple second modulators 1241, and the number of the second modulators 1241 is the same as the number of wavelengths of the multi-wavelength optical carrier; each of the second modulators 1241 is used to load one of the first group of vector data on a single wavelength optical carrier transmitted by the first wavelength division multiplexer 123.

[0134] Specifically, the optical frequency comb source 121 is a spectrum composed of a series of frequency components that are evenly spaced and have a coherent and stable phase relationship in the spectrum, and is used to send a multi-wavelength optical carrier. The second beam splitter 122 is used to divide the multi-wavelength optical carrier into two paths, one for direct output and one for loading multiple sets of vector data. In practical applications, the beam splitter can be a planar optical waveguide. Because the wavelengths are independent of each other during the calculation process, when the beam splitter splitting ratio meets the conditions of the detection module 3, the beam splitter is not required to have the same splitting ratio for each wavelength. The first wavelength division multiplexer 123 is a device that divides a multi-wavelength optical signal into multiple single-wavelength optical signals. For example, when the multi-wavelength optical carrier input to the wavelength division multiplexer 125 is λ1, λ2, ..., λn, the corresponding multiple output wavelengths are λ1, λ2, ..., λn respectively. The modulator array 124 is used to load multiple sets of vector data on the multi-wavelength optical carrier output by the wavelength division multiplexer. Each second modulator 1241 loads a set of vector data on one of the wavelength optical carriers. The number of second modulators 1241 is the same as the number of wavelengths. The second modulator 1241 only needs to modulate the amplitude of the optical carrier, and the phase of the optical carrier remains unchanged. In actual implementation, a push-pull modulator can be used, and voltages with opposite signs are applied to the two arms of the Mach-Zehnder interferometer. The wavelength division multiplexer 125 has the opposite function to the wavelength division multiplexer, and is used to multiplex the multi-channel multi-wavelength intermediate optical signals output by the second modulator 1241 into one multi-wavelength intermediate optical signal output.

[0135] In summary, the light source module 1 includes two outputs, one for outputting an optical carrier without a modulated signal, and one for outputting an intermediate optical signal loaded with the first set of vector data to be calculated. In the embodiment of the present application, the light source module 1 can adopt a single-wavelength light source or a multi-wavelength light source. When a single-wavelength light source is adopted, the two outputs are respectively: a single-wavelength optical carrier, and a single-wavelength carrier loads an intermediate optical signal of a set of vector data. When a multi-wavelength light source is adopted, the two outputs are respectively: a single-wavelength optical carrier, and a multi-wavelength optical carrier loads an intermediate optical signal of multiple sets of vector data, wherein the number of vectors is equal to the number of wavelengths. The light source module 1 can send multiple optical signals according to each element of the first set of vector data to be calculated, and can load the first set of vector data to be calculated by modulating the amplitude of the optical carrier. At this time, the phase of the optical carrier needs to remain unchanged, and different elements of the first set of vector data correspond to different amplitudes of the optical carrier. When implementing different convolution operations, the wavelength of the optical carrier needs to satisfy a certain relationship with the FSR of the programmable FIR filter 2. Only a single-wavelength light source can be used, which can avoid the dependence on multi-wavelength light sources in the multi-modulators and MRR wavelength division multiplexing used in the MZI 2111 grid, reduce the difficulty of on-chip integration while ensuring the rate, and reduce the manufacturing cost of the optical convolution computing device; in practical applications, multi-wavelength light sources can also be used to simultaneously calculate the convolution operations of multiple groups of vectors, further improving the computing speed.

[0136] In order to further improve the application reliability and effectiveness of the programmable FIR filter 2 of the single-wavelength light source, in the optical convolution calculation device provided in the embodiment of the present application, see Figure 4 , the programmable FIR filter 2 in the optical convolution computing device specifically includes the following contents:

[0137] A third beam splitter 21, a phase shifter array 22, a delay line array 23 and a beam combiner group 24 connected in sequence; the third beam splitter 21 is connected to the light source module 1, and the beam combiner group 24 is connected to the detection module 3;

[0138] The third beam splitter 21 is used to split the intermediate optical signal into multiple paths to obtain multiple optical signals, wherein the power division ratio of each path is adjustable to self-configure the amplitude of the tap coefficient of the programmable FIR filter 2;

[0139] The phase shifter array 22 is used to adjust the phase of each path of the multi-path optical signal output by the third beam splitter 21 to self-configure the phase of the tap coefficient, thereby realizing the loading of the second set of vector data on the intermediate optical signal through the tap coefficient;

[0140] The delay line array 23 is used to discretely delay the multi-path optical signals output by the phase shifter array 22 to obtain multi-path delay-weighted optical signals;

[0141] The beam combiner group 24 is used to combine the multi-path time-delay weighted optical signals output by the delay line array 23 into one path to output a corresponding target optical signal, wherein the optical field of the target optical signal is the superposition result of the optical fields of the multi-path time-delay weighted optical signals.

[0142] The implementation of the programmable FIR filter 2 can be roughly divided into two structures, cascade type and parallel type, and there are many variations for these two structures. The third beam splitter 21 is used to divide the intermediate optical signal output by the light source module 1 into multiple paths, and the power division ratio of each path is adjustable, which is equivalent to adjusting the amplitude of the tap coefficient of the programmable FIR filter 2.

[0143] The phase shifter array 22 is used to adjust the phase of each path of the multi-path signal output by the third beam splitter 21, which is equivalent to adjusting the phase of the tap coefficient of the programmable FIR filter 2. The delay line array 23 is used to discretely delay the multi-path optical signals, and optical waveguides of different lengths can be used in specific implementation. The beam combiner group 24 can also be called a multi-path beam combiner, which is used to combine multiple paths of delayed weighted optical signals into one path, and the output optical field is the superposition of the optical fields of each path.

[0144] In order to further improve the application reliability and effectiveness of the third beam splitter 21, in the optical convolution computing device provided in the embodiment of the present application, see Figure 5 , the third beam splitter 21 in the optical convolution computing device may specifically adopt a programmable MZI array 211;

[0145] The programmable MZI array 211 includes a plurality of sequentially connected MZI 2111 units, and the first MZI 2111 unit includes one MZI 2111; the number of MZI 2111 in the other MZI 2111 units is twice the number of MZI 2111 in the previous MZI 2111 unit;

[0146] Except for the last MZI 2111 unit, each of the MZI 2111 in the other MZI 2111 units is connected to two MZI 2111 in the next MZI 2111 unit.

[0147] The power division ratio of each MZI 2111 is adjustable, and a multi-channel power divider with adjustable power division ratio can be realized by using this structure.

[0148] In order to further improve the application reliability and effectiveness of the beam combiner group 24, in the optical convolution computing device provided in the embodiment of the present application, see Figure 6, the beam combiner group 24 in the optical convolution computing device includes a plurality of beam combiner units connected in sequence, and the last beam combiner unit includes a first beam combiner 241; the number of first beam combiners 241 in other beam combiner units is twice the number of first beam combiners 241 in the next beam combiner unit;

[0149] Except for the first beam combiner unit, every two first beam combiners 241 in other beam combiner units are connected to one first beam combiner 241 in the next beam combiner unit. The 2*2 beam combiner is cascaded, and the 2*2 beam combiner can be a directional coupler, a multi-mode interferometer, etc.

[0150] In order to further improve the application reliability and effectiveness of the phase shifter array 22 and the delay line array 23, in the optical convolution computing device provided in the embodiment of the present application, the phase shifter array 22 in the optical convolution computing device specifically includes: a plurality of first phase shifters 221, and the number of the first phase shifters 221 is the same as the number of the multi-path optical signals output by the third beam splitter 21;

[0151] The delay line array 23 specifically includes: a plurality of delay lines 231 , and the number of the delay lines 231 is the same as the number of the multi-path optical signals output by the third beam splitter 21 .

[0152] In summary, the programmable FIR filter 2 is based on coherent light interference. From the frequency domain point of view, it is a module that can filter the input data. Its transmission spectrum is a periodic spectrum. The transmission spectrum of FIR can be changed by changing the tap coefficients of the programmable FIR filter 2, and the free spectral range (FSR) of the programmable FIR filter 2 does not change. From the time domain point of view, it is equivalent to performing delayed weighted summation on the input signal, that is, the convolution operation of the signal is implemented. The discrete form of the programmable FIR filter 2 is expressed as the above formula (6), and the obtained output discrete signal is the one-dimensional discrete convolution result of the input signal x(n) and the tap coefficient. By programming the tap coefficients of the FIR filter, the second set of vector data is loaded on the output intermediate light signal of the light source module 1.

[0153] When the programmable FIR filter 2 receives a single-channel single-wavelength optical signal output by the light source module 1, the programmable FIR filter 2 loads a second set of vector data for the single-channel single-wavelength optical signal. When the programmable FIR filter 2 receives a single-channel multi-wavelength optical signal output by the light source module 1, the programmable FIR filter 2 simultaneously loads a second set of vector data for multiple wavelength intermediate optical signals. The output result of the programmable FIR filter 2 indicates the convolution result of two or more sets of vector data.

[0154] The self-configuration process of the programmable FIR filter 2 needs to meet the minimum phase condition, namely: (1) the reference path of the optical convolution calculation device needs to have the shortest path delay relative to the programmable FIR filter 2; (2) assuming that the tap coefficient of the reference path is a0, the tap coefficients corresponding to each delay path of the programmable FIR filter 2 are h(n), which needs to satisfy |a0|≥

[0155] ∑ n |k(n)|.

[0156] In order to further improve the application reliability and effectiveness of the detection module 3 applicable to a single-wavelength light source, in the optical convolution computing device provided in the embodiment of the present application, see Figure 7 , the detection module 3 in the optical convolution computing device specifically includes the following contents:

[0157] Single wavelength detection module 31; the single wavelength detection module 31 comprises: a second phase shifter 311, a second beam combiner 312 and a first detector 313 connected in sequence; the second beam combiner 312 is connected to the light source module 1, and the second phase shifter 311 is connected to the programmable FIR filter 2;

[0158] The second phase shifter 311 is used to receive the target optical signal and change the phase difference of the target optical signal;

[0159] The second beam combiner 312 is used to combine the target optical signal output by the second phase shifter 311 and the optical carrier without the modulated signal output by the light source module 1 into one path to obtain a corresponding composite signal;

[0160] The first detector 313 is used to detect and obtain optical power data of the synthetic signal for representing a vector convolution calculation result.

[0161] The light source module 1 can use a single-wavelength optical carrier or a multi-wavelength optical carrier, and different detection modules 3 correspond to different optical carriers. When a single-wavelength optical carrier is selected, the corresponding single-wavelength detection module 31 realizes the coherent detection function. The second phase shifter 311 is used to change the phase difference of the two optical signals input by the single-wavelength detection module 31, the second combiner 312 is used to combine the two optical signals into one output, and the first detector 313 (also written as a PD detector) is used to detect the optical power after the two signals are combined. The second combiner 312 can be a multi-mode interferometer or a directional coupler, etc.

[0162] In order to further improve the application reliability and effectiveness of the detection module 3 applicable to the multi-wavelength light source, in the optical convolution computing device provided in the embodiment of the present application, see Figure 8 , the detection module 3 in the optical convolution computing device further specifically includes the following contents:

[0163] A multi-wavelength detection module 32; the multi-wavelength detection module 32 comprises: a third phase shifter 321, a third beam combiner 322, a second wavelength division multiplexer 323 and a detector array 324 connected in sequence; the third beam combiner 322 is connected to the light source module 1, the third phase shifter 321 is connected to the programmable FIR filter 2, and the detector array 324 is provided with a plurality of second detectors 3241;

[0164] The third phase shifter 321 is used to receive the target optical signal and change the phase difference of the target optical signal;

[0165] The third beam combiner 322 is used to combine the target optical signal output by the third phase shifter 321 and the optical carrier without the modulated signal output by the light source module 1 into one channel, so as to obtain a corresponding single-channel multi-wavelength synthetic signal;

[0166] The second wavelength demultiplexer 323 performs demultiplexing processing on the single-channel multi-wavelength composite signal to obtain multiple groups of optical signals which are respectively vector convolution results;

[0167] The detector array 324 is used to detect and obtain optical power data corresponding to each group of optical signals serving as vector convolution results, which is used to represent the vector convolution calculation result.

[0168] The functions of the third phase shifter 321 and the third beam combiner 322 in the multi-wavelength detection module 32 are as described above, and have the same functions as the second phase shifter 311 and the second beam combiner 312 in the single-wavelength detection module 31. The second wavelength demultiplexer 323 is used to demultiplex a single-channel multi-wavelength optical signal. The third beam combiner 322 outputs an optical signal indicating the result of convolution of multiple groups of first vector data loaded by the multi-wavelength carrier of the light source module 1 and the second vector data loaded by the programmable FIR filter 2. After wavelength demultiplexing, it is equivalent to separating the convolution results of multiple groups of first vector data and second vector data. Each wavelength indicates the convolution result of a group of first vector data and second vector data. The number of output channels of the second wavelength demultiplexer 323 is the same as the number of wavelengths of the optical carrier. The detector array 324 is used to detect multiple groups of vector convolution results. The number of second detectors 3241 in the detector array 324 is the same as the number of output channels of the second wavelength division multiplexer 323. Each second detector 3241 corresponds to the convolution result of the first vector data loaded by a wavelength optical carrier and the second vector data loaded by the programmable FIR filter 2.

[0169] In summary, the detection module 3 is used to detect the light intensity after the optical signal output by the programmable FIR filter 2 is combined with the optical carrier output by the light source module 1, and the detected power indicates the convolution information of the vector. When the light source module 1 adopts a single-wavelength optical carrier, the optical power detected by the detection module 3 indicates the convolution calculation result of the first group of vectors loaded by the light source module 1 and the second group of vectors loaded by the programmable FIR filter 2. When the light source module 1 adopts a multi-wavelength carrier, it is equivalent to that the light source module 1 loads multiple groups of vectors at the same time. At this time, the detection module 3 outputs multiple vectors at the same time, corresponding to the convolution result of the multiple groups of vectors loaded by the light source module 1 and the vectors loaded by the programmable FIR filter 2.

[0170] That is to say, the optical convolution computing device provided in the embodiment of the present application is an optical convolution computing system based on an on-chip integrated programmable FIR filter 2. The optical convolution computing device simultaneously utilizes time division multiplexing and space division multiplexing technologies, and can further improve the computing speed by utilizing the dimension of wavelength, thereby maximizing the multi-dimensional information processing capability of the optical chip while ensuring the chip integration, and the core device of the system, the programmable FIR filter 2, can be self-configured, avoiding the subsequent calibration problem of the optical chip.

[0171] The above embodiment describes the structure of the optical convolution computing device provided in the embodiment of the present application. For ease of understanding, the following will be combined with Fig. 9 The method of self-configuring FIR tap coefficients and Fig.10 The convolution calculation method shown is an exemplary description of how the optical convolution calculation device provided in an embodiment of the present application performs convolution calculation when the light source module 1 adopts a single-wavelength light source.

[0172] The self-configured FIR tap coefficient method can be based on Fig.11 The optical convolution computing device shown is implemented. At this time, it is required that the spectrum of the modulator has no effect on the power spectrum measured by the detection module 3. A wide-bandwidth modulator can be used. The modulator in the light source module 1 can also be directly removed to measure the system power spectrum. Specifically, Fig. 9The frequency sweeping light output in step S201 can be realized by the tunable laser 111 of the light source module 1, and the detection module 3 of the optical convolution calculation device outputs the power of the system at the corresponding frequency, that is, the system power spectrum (the output power of the system at each wavelength) is obtained. In step S202, it can be known from the Kramers-Kronig relationship that when the system meets the minimum phase condition, the system phase response can be restored by the system amplitude response, thereby obtaining the complete spectrum response of the system. In step S203, the complex spectrum response obtained in step S202 is a periodic spectrum. By intercepting one period of the periodic spectrum and performing an inverse Fourier transform, the tap coefficients of the FIR filter and the proportion of the reference light (optical carrier) input by the detection module 3 can be obtained simultaneously. Based on the actual FIR tap coefficients measured in the first three steps, different training algorithms can be adopted for different structures adopted by FIR, so as to obtain the complex spectrum response obtained in step S203. Figure 4 As an example, the third beam splitter 21 and the beam combiner group 24 respectively adopt Figure 5 and Figure 6 The binary tree structure of the third beam splitter 21 can be obtained from the tap coefficient of the FIR. Taking the 4-tap FIR filter as an example, the tap coefficient amplitudes are [a1, a2, a3, a4] from top to bottom. The beam splitter requires 2 levels of MZI 2111, and the splitting ratio of the first level MZI 2111 is The splitting ratios of the secondary MZI2111 are and The angle value of the tap coefficient is obtained Figure 4 The phase of the phase shift array 1022 is used as a priori information to update the variable parameters in the programmable FIR filter 2 using the gradient descent method.

[0173] Fig.10 This is a flow chart of a method for performing convolution calculation by an optical convolution calculation device. For the sake of clarity, the present application embodiment takes the first vector data X=[a1, a2, a3, a4, a5, a6, a7, a8, a9] and the second vector data H=[h1, h2, h3, h4] as an example. Fig.10 In step S301, the light source module 1 loads the first set of vector data through the push-pull modulator, and the output intermediate light signal is loaded with the second set of vector data through the programmable FIR filter 2, that is, step S302. The output discrete time series is the specific operation process of the convolution calculation as follows Fig.12 As shown in the figure, the input signal x(n) undergoes discrete increasing delays, different delays correspond to different h(n), and the output signal at each moment is the convolution calculation result of the corresponding two vectors X and H.

[0174] When using one-dimensional vector convolution to implement two-dimensional matrix convolution, the specific implementation method is as follows Fig.13 and Fig.14As shown, in Fig.13 In the above example, the two input matrices A and B are expanded into two vectors according to certain rules. Fig.14 In this case, the above vectors are subjected to vector convolution operation. Fig.14 The vector convolution output results y4, y8, y12, y16 and Fig.13 The matrix convolution operation results c1, c2, c3, and c4 are the same, that is, matrix convolution can be achieved through vector convolution operation.

[0175] exist Figure 12 to Figure 14 In the above, a represents the input vector or input matrix; a followed by a is used to distinguish different elements in a; b represents the convolution kernel matrix; b followed by a number is used to distinguish different elements in b; h represents the convolution kernel vector; h followed by a number is used to distinguish different elements of the convolution kernel vector; y represents the vector convolution result; c represents the two-dimensional convolution output result; c followed by a number is used to distinguish different elements of the two-dimensional convolution result; such as "a1b1", "a+number+b+number" is used to represent the product of an element in a and an element in b, and the multiplication sign is omitted.

[0176] Based on this, the optical convolution computing device provided by the present application uses a programmable FIR filter to implement convolution operations. The programmability of the FIR filter is reflected in the adjustable amplitude and phase of its taps. The FIR filter input optical signal can use a single-wavelength carrier, and the FIR filter supports single-channel input, thereby reducing the requirements for the light source and modulator, and is more suitable for on-chip integration. The tap coefficients of the programmable FIR filter 2 of the above-mentioned optical convolution computing device can be self-configured under uncalibrated conditions by adding a reference path to the FIR. The added reference path is required to meet the conditions of the Kramers-Kronig algorithm, namely (1) the reference path has the shortest delay compared to the programmable FIR filter 2; (2) assuming that the reference path tap coefficient is a0, the tap coefficients corresponding to each delay path of the programmable FIR filter 2102 are h(n), which need to satisfy |a0|>∑ n |h(n)|. The programmable FIR filter 2 supports operation under multi-wavelength carrier conditions at the same time, so the wavelength division multiplexing method can be used to increase the parallelism and further improve the operation speed.

[0177] The optical convolution computing device provided in this application has at least the following beneficial effects:

[0178] (1) By using both time division multiplexing and space division multiplexing technologies, and making full use of the high parallelism of light, only a single wavelength light source and a single modulator can be used, thus getting rid of its dependence on multi-wavelength light sources and multiple modulators. On the basis of maintaining high computing speed, the integration density is improved, and it is easy to realize on-chip integration;

[0179] (2) A method for self-configuring neural network weight parameters without structural pre-calibration is proposed;

[0180] (3) The entire system works in a stable state, and there is no problem of unstable jitter of weight parameters. The system is compatible with wavelength division multiplexing method and can use the dimension of wavelength to further improve the computing speed.

[0181] In the present application, the expressions "first", "second", "third", etc. are only used to distinguish the functions of components in different scenarios or modules. The expressions "first", "second", "third", etc. are only used to distinguish that the functions of components in different scenarios or modules are different, but can be implemented by the same type or the same devices or components. Of course, they can also be implemented by different types or different devices or components.

[0182] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is specifically performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0183] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0184] In this application, the features described and / or illustrated for one embodiment may be used in the same manner or in a similar manner in one or more other embodiments, and / or combined with or replace the features of other embodiments.

[0185] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An optical convolution computing device, characterized in that: include: On-chip integrated light source module, programmable FIR filter and detection module; The light source module includes two outputs, one for outputting an intermediate optical signal loaded with a first set of vector data according to a target light source, and the other for outputting an optical carrier without adding a modulation signal according to the target light source, wherein the target light source includes: a single wavelength light source; The programmable FIR filter is used to receive the intermediate optical signal, load the second set of vector data obtained by self-configuration of the taps of the programmable FIR filter as a weight vector, output each discrete optical signal and generate a corresponding target optical signal, wherein each of the optical signals is used to represent the product between an element in the first set of vector data and an element in the second set of vector data; The detection module is used to receive the target optical signal and the optical carrier without the modulated signal, and detect and obtain optical power data corresponding to the optical signal and the optical carrier without the modulated signal for representing the vector convolution calculation result.

2. The optical convolution computing device according to claim 1, characterized in that: The light source module comprises: a single wavelength light source module; The single-wavelength light source module comprises: a laser, a first beam splitter and a first modulator connected in sequence; Wherein, the laser is used to emit a single-channel single-wavelength optical carrier as the single-wavelength light source; The first beam splitter is used to split the single-wavelength optical carrier of a single channel into two channels, so as to obtain one channel of the single-wavelength optical carrier directly output without adding a modulation signal and another channel of the single-wavelength optical carrier transmitted to the first modulator; The first modulator is used to load a first set of vector data onto the single-wavelength optical carrier transmitted by the first beam splitter, and output an intermediate optical signal accordingly.

3. The optical convolution computing device according to claim 1, characterized in that: The target light source further comprises: a multi-wavelength light source; correspondingly, the light source module comprises: a multi-wavelength light source module; The multi-wavelength light source module comprises: an optical frequency comb source, a second beam splitter, a first wavelength division multiplexer, a modulator array and a wavelength division multiplexer connected in sequence; Wherein, the optical frequency comb source is used to emit a single-channel multi-wavelength optical carrier as the multi-wavelength light source; The second beam splitter is used to split the single-channel multi-wavelength optical carrier into two channels, so as to obtain one channel of the multi-wavelength optical carrier directly output without adding a modulation signal and another channel of the multi-wavelength optical carrier transmitted to the wavelength division multiplexer; The first wavelength division multiplexer is used for dividing the multi-wavelength optical carrier transmitted by the second beam splitter into multiple single-wavelength optical carriers; The modulator array is used to load each first group of vector data to multiple single-wavelength optical carriers respectively, and to output the single-wavelength intermediate optical signals corresponding to each of the single-wavelength optical carriers respectively; The wavelength division multiplexer is used to multiplex the single-wavelength intermediate optical signals into one multi-wavelength intermediate optical signal.

4. The optical convolution computing device according to claim 3, characterized in that: The modulator array includes: a plurality of second modulators, and the number of the second modulators is the same as the number of wavelengths of the multi-wavelength optical carrier; Each of the second modulators is used to load one of the first groups of vector data onto a single wavelength optical carrier transmitted by the first wavelength division multiplexer.

5. The optical convolution computing device according to claim 1, characterized in that: The programmable FIR filter comprises: a third beam splitter, a phase shifter array, a delay line array and a beam combiner group connected in sequence; the third beam splitter is connected to the light source module, and the beam combiner group is connected to the detection module; The third beam splitter is used to split the intermediate optical signal into multiple paths to obtain multiple optical signals, wherein the power division ratio of each path is adjustable to self-configure the amplitude of the tap coefficient of the programmable FIR filter; The phase shifter array is used to adjust the phase of each path of the multi-path optical signal output by the third beam splitter to self-configure the phase of the tap coefficient, thereby realizing the loading of the second set of vector data on the intermediate optical signal through the tap coefficient; The delay line array is used to discretely delay the multi-path optical signals output by the phase shifter array to obtain multi-path delay-weighted optical signals; The combiner group is used to combine the multi-path time-delay weighted optical signals output by the delay line array into one path to output a corresponding target optical signal, wherein the optical field of the target optical signal is the superposition result of the optical fields of the multi-path time-delay weighted optical signals.

6. The optical convolution computing device according to claim 5, characterized in that: The third beam splitter adopts a programmable MZI array; The programmable MZI array includes a plurality of sequentially connected MZI units, and the first MZI unit includes one MZI; the number of MZIs in the other MZI units is twice the number of MZIs in the previous MZI unit; Except for the last MZI unit, each of the MZIs in the other MZI units is connected to two MZIs in the next MZI unit.

7. The optical convolution computing device according to claim 5, characterized in that: The beam combiner group includes a plurality of beam combiner units connected in sequence, and the last beam combiner unit includes a first beam combiner; the number of first beam combiners in the other beam combiner units is twice the number of first beam combiners in the next beam combiner unit; Except for the first beam combiner unit, every two first beam combiners in other beam combiner units are connected to a first beam combiner in the next beam combiner unit.

8. The optical convolution computing device according to claim 5, characterized in that: The phase shifter array comprises: a plurality of first phase shifters, and the number of the first phase shifters is the same as the number of the multi-path optical signals output by the third beam splitter; The delay line array includes: a plurality of delay lines, and the number of the delay lines is the same as the number of the multi-path optical signals output by the third beam splitter.

9. The optical convolution computing device according to claim 1, characterized in that: The detection module comprises: a single wavelength detection module; The single wavelength detection module comprises: a second phase shifter, a second beam combiner and a first detector connected in sequence; the second beam combiner is connected to the light source module, and the second phase shifter is connected to the programmable FIR filter; The second phase shifter is used to receive the target optical signal and change the phase difference of the target optical signal; The second beam combiner is used to combine the target optical signal output by the second phase shifter and the optical carrier without the modulated signal output by the light source module into one path to obtain a corresponding composite signal; The first detector is used to detect and obtain optical power data of the synthetic signal for representing a vector convolution calculation result.

10. The optical convolution computing device according to claim 1, characterized in that: The target light source further includes: a multi-wavelength light source; correspondingly, the detection module includes: a multi-wavelength detection module; The multi-wavelength detection module comprises: a third phase shifter, a third beam combiner, a second wavelength division multiplexer and a detector array connected in sequence; the third beam combiner is connected to the light source module, the third phase shifter is connected to the programmable FIR filter, and a plurality of second detectors are provided in the detector array; The third phase shifter is used to receive the target optical signal and change the phase difference of the target optical signal; The third beam combiner is used to combine the target optical signal output by the third phase shifter and the optical carrier without the modulated signal output by the light source module into one channel, so as to obtain a corresponding single-channel multi-wavelength synthetic signal; The second wavelength demultiplexer performs demultiplexing processing on the single-channel multi-wavelength composite signal to obtain multiple groups of optical signals which are respectively vector convolution results; The detector array is used to detect and obtain optical power data corresponding to each group of optical signals respectively serving as vector convolution results and used to represent the vector convolution calculation results.

Images