Opto-electric hybrid tensor convolution computing system, method, apparatus and readable storage medium

Abstract

The application relates to an optoelectronic hybrid tensor convolution calculation system, method, device and readable storage medium, which comprises the following steps: a laser source is used for generating an original optical frequency comb; an optical delay-spectroscopic device is used for splitting and adjusting the original optical frequency comb into two paths of the same frequency and phase-locked optical frequency combs; an electro-optical modulation array is used for separating the optical frequency comb on each path to a corresponding plurality of channels, loading input information onto the optical frequency comb of each channel, and performing frequency spectrum splicing on the output signals of the corresponding plurality of channels of each path to realize system noise reduction; an optical delay device is used for modulating the two frequency spectrum splicing results into optical signals with equal optical paths; an optical frequency mixer is used for mixing the two optical signals to generate two optical mixing signals with a phase difference of 0 and 180 degrees; an optoelectronic balanced detector is used for optoelectronic conversion of the two optical mixing signals to output electrical signals; and a data acquisition device is used for determining a convolution result according to the acquired electrical signals. The application can improve the signal-to-noise ratio and convolution calculation precision of the system.

Description

Technical Field

[0001] This application relates to the fields of optical communication technology and optical computing technology, and in particular to an optoelectronic hybrid tensor convolution computing system, method, device and readable storage medium. Background Technology

[0002] Artificial neural networks (ANNs) are mathematical or computational models in machine learning and cognitive science that mimic the structure and function of biological neural networks, used for estimating or approximating functions. However, current ANNs require a large number of tensor convolution operations, which account for 55% to 90% of the total computation. Therefore, accelerating tensor convolution operations is a crucial way to improve the computational power of neural networks. To enhance the computational power of neural networks, accelerating tensor convolution operations necessitates improving the computational capabilities of the underlying neuromorphic hardware. Optical neural networks (ONNs) have emerged as a trend in next-generation neuromorphic computing due to their potential to overcome the bandwidth bottleneck of electronic artificial neural networks and their high-speed, low-power capabilities.

[0003] Currently, there are various technical approaches to realize optical tensor convolution operations. One such approach utilizes IQ modulators, mixers, and balanced detectors to perform multiplication in the time domain and accelerate convolution in the frequency domain, creating a hybrid optoelectronic tensor convolution computation system. While this system offers advantages such as integrability, programmability, and scalability, its computational accuracy is reduced due to device non-ideals and noise during tensor convolution operations. This hinders the application of the hybrid optoelectronic tensor convolution acceleration system in AI applications. Therefore, noise reduction in hybrid optoelectronic tensor convolution computation systems is a pressing issue that needs to be addressed. Summary of the Invention

[0004] This application provides a photoelectric hybrid tensor convolution calculation system, method, device, and readable storage medium, with the aim of achieving noise reduction in the photoelectric hybrid tensor convolution calculation system, thereby improving the accuracy of convolution calculation.

[0005] Firstly, a photoelectric hybrid tensor convolution computation system is provided, comprising:

[0006] A laser source, used to generate the original optical frequency comb;

[0007] An optical delay-splitting device is used to split the original optical frequency comb and adjust it into two optical frequency combs with the same frequency and phase lock-in.

[0008] An electro-optic modulation array is used to separate the optical frequency comb of each path to multiple corresponding channels, load the input information onto the optical frequency comb of each channel, and splice the output signals of the multiple corresponding channels of each path to achieve system noise reduction.

[0009] An optical delay device is used to modulate the splicing result of two spectra into an optical signal with equal optical path length.

[0010] An optical mixer is used to mix two optical signals with equal optical path lengths to generate two optical mixed signals with a phase difference of 0° and 180°, respectively.

[0011] A photoelectric balance detector is used to perform photoelectric conversion on two optical mixing signals to output an electrical signal;

[0012] A data acquisition device used to acquire electrical signals and determine the convolution result based on the acquired electrical signals.

[0013] In some embodiments, the electro-optic modulation array includes:

[0014] The first modulation component includes a first dewavelength division multiplexer, a plurality of parallel first signal loading modulation units, and a first wavelength division multiplexer connected in sequence.

[0015] The second modulation component includes a second dewavelength division multiplexer, a plurality of parallel second signal loading modulation units, and a second wavelength division multiplexer connected in sequence.

[0016] In some embodiments, the first signal loading modulation unit includes a first waveform generator, a first modulator driver, a first 90° electrical phase shifter, and a first IQ modulator connected in sequence;

[0017] The second signal loading modulation unit includes a second waveform generator, a second modulator driver, a second 90° electrical phase shifter, and a second IQ modulator connected in sequence.

[0018] In some embodiments, the electrical signal I(t) output by the photoelectric balance detector is expressed as:

[0019]

[0020] In the formula, α represents the photoelectric conversion efficiency, C represents the amplitude of the light source, and A i B represents the data of the i-th input tensor. k This represents the data of the k-th weight tensor. This represents the initial frequency corresponding to the input tensor. The initial frequency corresponding to the weight tensor is represented by σ', where t represents time, ω0 represents the frequency interval of the electrical signal, and σ' represents the initial frequency of the weight tensor. ik This represents the current noise output from the photoelectric balance detector after noise reduction and filtering.

[0021] In some embodiments, the number of channels corresponding to each path is 2, and the spectrum splicing result of each output signal is obtained by the following calculation formula:

[0022] M1(t)=M s1 (t)+M s2 (t)

[0023] M2(t)=M s3 (t)+M s4 (t)

[0024] In the formula, M1(t) represents the spectrum splicing result of the upstream signal, M s1 (t) represents the output signal of the first channel corresponding to the upper path, M s2 M(t) represents the output signal of the second channel corresponding to the upper path, and M2(t) represents the spectrum splicing result of the lower path signal. s3 (t) represents the output signal of the third channel corresponding to the lower path, M s4 (t) represents the output signal of the fourth channel corresponding to the downlink, where the output signal of each channel contains its corresponding noise.

[0025] Secondly, a method for calculating optoelectronic hybrid tensor convolution is provided, including the following steps:

[0026] The acquired raw optical frequency comb is split and adjusted into two optical frequency combs with the same frequency and phase lock.

[0027] The optical frequency comb on each path is separated into multiple corresponding channels. The input information is loaded onto the optical frequency comb of each channel, and the output signals of the multiple corresponding channels of each path are spliced ​​together to achieve system noise reduction.

[0028] The spectrum splicing results of the two channels are modulated into optical signals with equal optical path lengths, and the two optical signals with equal optical path lengths are mixed to generate two optical mixed signals with a phase difference of 0° and 180° respectively.

[0029] The two optical mixing signals are converted into photoelectric signals, and the convolution result is determined based on the output electrical signal.

[0030] In some embodiments, the expression for the electrical signal I(t) output after photoelectric conversion is:

[0031]

[0032] In the formula, α represents the photoelectric conversion efficiency, C represents the amplitude of the light source, and A i B represents the data of the i-th input tensor. k This represents the data of the k-th weight tensor. This represents the initial frequency corresponding to the input tensor. The initial frequency corresponding to the weight tensor is represented by σ', where t represents time, ω0 represents the frequency interval of the electrical signal, and σ' represents the initial frequency of the weight tensor. ik This indicates the current noise output after noise reduction, filtering, and photoelectric conversion.

[0033] In some embodiments, the number of channels corresponding to each path is 2, and the spectrum splicing result of each output signal is obtained by the following calculation formula:

[0034] M1(t)=M s1 (t)+M s2 (t)

[0035] M2(t)=M s3 (t)+M s4 (t)

[0036] In the formula, M1(t) represents the spectrum splicing result of the upstream signal, M s1 (t) represents the output signal of the first channel corresponding to the upper path, M s2 M(t) represents the output signal of the second channel corresponding to the upper path, and M2(t) represents the spectrum splicing result of the lower path signal. s3 (t) represents the output signal of the third channel corresponding to the lower path, M s4 (t) represents the output signal of the fourth channel corresponding to the downlink, where the output signal of each channel contains its corresponding noise.

[0037] Thirdly, a photoelectric hybrid tensor convolution computing device is provided, comprising: a memory and a processor, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the aforementioned photoelectric hybrid tensor convolution computing method.

[0038] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned photoelectric hybrid tensor convolution calculation method.

[0039] This application provides a photoelectric hybrid tensor convolution calculation system, method, device, and readable storage medium, including a laser source for generating an original optical frequency comb; an optical delay-splitting device for splitting the original optical frequency comb and adjusting it into two phase-locked optical frequency combs of the same frequency; an electro-optic modulation array for separating the optical frequency combs on each path to multiple corresponding channels, loading input information onto the optical frequency combs of each channel respectively, and performing spectral splicing on the output signals of the multiple channels corresponding to each path to achieve system noise reduction; an optical delay device for modulating the spectral splicing result of the two paths into an optical signal with equal optical path length; an optical mixer for mixing the two optical signals with equal optical path length to generate two optical mixed signals with a phase difference of 0° and 180° respectively; a photoelectric balance detector for photoelectric conversion of the two optical mixed signals to output an electrical signal; and a data acquisition device for acquiring the electrical signal and determining the convolution result based on the acquired electrical signal. This application extends the convolution calculation unit in parallel by using dual optical frequency combs at the same frequency to reduce channel noise, thereby improving the signal-to-noise ratio of the system and thus improving the accuracy of convolution calculation. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A schematic diagram of the framework of a photoelectric hybrid tensor convolution calculation system provided in this application embodiment;

[0042] Figure 2 This is a schematic diagram of the structure of the optoelectronic hybrid tensor convolution calculation system provided in the embodiments of this application;

[0043] Figure 3 A schematic diagram of the structure of the undenoised optoelectronic hybrid tensor convolution calculation system provided in an embodiment of this application;

[0044] Figure 4 A flowchart illustrating a photoelectric hybrid tensor convolution calculation method provided in this application embodiment;

[0045] Figure 5 This is a schematic diagram of the structure of a photoelectric hybrid tensor convolution computing device provided in an embodiment of this application. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] This application provides an optoelectronic hybrid tensor convolution calculation system, method, device, and readable storage medium, with the aim of achieving noise reduction in the optoelectronic hybrid tensor convolution calculation system, thereby improving the accuracy of convolution calculation.

[0048] Figure 1 This application provides an embodiment of a photoelectric hybrid tensor convolution calculation system, comprising: a laser source 1, an optical delay-splitter 2, an electro-optic modulation array 3, an optical delay device 4, an optical mixer 5, a photoelectric balanced detector 6, and a data acquisition device 7. It should be noted that the photoelectric hybrid tensor convolution calculation system may further include a computer control system 8, which is used to control the adjustment parameters in the optical delay-splitter 2, the electro-optic modulation array 3, and the optical delay device 4, and to provide feedback on the calculation results from the data acquisition device 7.

[0049] The system comprises: a laser source 1 for generating the original optical frequency comb; an optical delay-splitter 2 for splitting the original optical frequency comb and adjusting it into two phase-locked optical frequency combs of the same frequency; an electro-optic modulation array 3 for separating the optical frequency combs on each path into multiple corresponding channels, loading the input information onto the optical frequency combs of each channel, and splicing the output signals of the multiple channels corresponding to each path to achieve system noise reduction; an optical delay device 4 for modulating the spliced ​​results of the two paths into optical signals with equal optical path lengths; an optical mixer 5 for mixing the two optical signals with equal optical path lengths to generate two optical mixed signals with phase differences of 0° and 180° respectively; a photoelectric balance detector 6 for photoelectric conversion of the two optical mixed signals to output electrical signals; and a data acquisition device 7 for acquiring electrical signals and determining the convolution result based on the acquired electrical signals, and then transmitting the convolution result to a computer control system 8.

[0050] Specifically, the laser source 1 is preferably an optical frequency comb source, which is used to provide the system with a coherent phase-locked optical frequency comb, i.e., a dual optical frequency comb with the same frequency; the optical delay-splitter 2 is mainly used to amplify the power of the optical frequency comb, split the light, and adjust the optical path of the two paths to be consistent, so as to provide the system with two optical frequency combs with the same frequency and phase lock; the electro-optic modulation array 3 based on the IQ modulator is used to separate the comb teeth of different wavelengths in the optical frequency comb, and modulate the different optical frequency comb teeth by the orthogonal frequency signals corresponding to the input simulation data to be calculated; the optical delay device 4 is used to adjust the optical path between the two paths of light to be consistent and amplify the power of the two paths of light; the optical mixer 5 based on 180° mixes the two input paths of light to generate two signal lights, which are output by the photoelectric balance detector 6; the data acquisition device 7 is used to acquire the electrical signal output by the photoelectric balance detector 6 and determine the convolution result, and then output the convolution result to the computer control system 8.

[0051] See Figure 2 As shown, the optical delay-splitter 2 includes a 1-to-2 optical splitter 202 and a first adjustable optical delay line 203. The 1-to-2 optical splitter 202 is used to split the original optical frequency comb into two optical frequency combs, and the first adjustable optical delay line 203 is used to adjust the optical path of the optical frequency comb to produce two optical frequency combs with the same frequency and phase lock. The optical delay-splitter 2 also includes a first fiber amplifier 201, which is preferably a high-power polarization-maintaining erbium-doped fiber amplifier, to amplify the power of the optical frequency comb.

[0052] In some embodiments, the electro-optic modulation array 3 includes a first modulation component and a second modulation component. The first modulation component includes a first demultiplexer 301, a plurality of parallel first signal loading modulation units, and a first wavelength division multiplexer 319 connected in sequence. The second modulation component includes a second demultiplexer 302, a plurality of parallel second signal loading modulation units, and a second wavelength division multiplexer 320 connected in sequence. The demultiplexer is used to separate the optical frequency comb of each path to multiple corresponding channels. The signal loading modulation units are used to load the input information onto the optical frequency comb of each channel. The wavelength division multiplexer is used to perform spectral splicing on the output signals of each corresponding channel to achieve system noise reduction.

[0053] It should be noted that the number of channels corresponding to the upper and lower optical frequency combs can be determined according to actual needs; that is, in practical applications, the number of channels can be increased as needed. For example, see... Figure 2As shown, the number of parallel channels on each path is 2. Therefore, there are 2 parallel first signal loading modulation units A and B on the upper path and 2 parallel second signal loading modulation units C and D on the lower path. Assuming that one of the first signal loading modulation units A corresponding to the upper path includes a first waveform generator A303, a first modulator driver A307, a first 90° phase shifter A311, and a first IQ modulator A315 connected in sequence, then the other first signal loading modulation unit B includes a first waveform generator B304, a first modulator driver B308, a first 90° phase shifter B312, and a first IQ modulator B316 connected in sequence; similarly, one of the second signal loading modulation units C corresponding to the lower path includes a second waveform generator C305, a second modulator driver C309, a second 90° phase shifter C313, and a second IQ modulator C317 connected in sequence, then the other second signal loading modulation unit D includes a second waveform generator D306, a second modulator driver D310, a second 90° phase shifter D314, and a second IQ modulator D318 connected in sequence.

[0054] In some embodiments, the optical delay device 4 includes a second adjustable optical delay line 401, a second fiber amplifier 402, and a third fiber amplifier 403. The second fiber amplifier 402 and the third fiber amplifier 403 can both preferably be high-power polarization-maintaining erbium-doped fiber amplifiers to amplify the power of the optical signal and compensate for optical losses during system operation; the second adjustable optical delay line 401 is used to modulate the two optical paths into in-phase optical signals with equal optical path lengths.

[0055] In some embodiments, the photoelectric balance detector 6 may preferably be a balance detector, the data acquisition device 7 includes an oscilloscope, and the oscilloscope can be controlled by the computer control system 8 and return the data to the computer control system 8.

[0056] It should be understood that all parameters of each waveform generator can be set by connecting to the computer control system 8; the I / Q modulator can be controlled by the computer control system 8; the tunable optical delay line can also be controlled by the computer control system 8 to adjust the optical path difference between the two optical paths so that the system works well.

[0057] The following combination Figure 2 The connection relationships between the various components in the hardware structure of the optoelectronic hybrid tensor convolution computation system are described. It should be noted that the various components in the optoelectronic hybrid tensor convolution computation system are connected via optical fibers and cables.

[0058] Preferably, the output port of the laser source 1 of the optical frequency comb source is connected to the input port of the first fiber amplifier 201, wherein the center frequency of the laser source 1 can be set to ω. cIts repetition frequency is Δω; the output port of the first fiber amplifier 201 is connected to the input port of the 1:1 beam splitter 202, and the two output ports of the 1:1 beam splitter 202 are respectively connected to the input port of the first adjustable optical delay line 203 and the optical input port of the first dewavelength division multiplexer 301; it can be understood that the optical path of the first adjustable optical delay line 203 can be set by the computer control system 8 to ensure that the two optical signals are in phase; the output port of the first adjustable optical delay line 203 is connected to the optical input port of the second dewavelength division multiplexer 302.

[0059] The two output ports of the first demultiplexer 301 are connected to the input ports of the first IQ modulator A315 and the first IQ modulator B316 in the IQ modulator array, respectively. Similarly, the two output ports of the second demultiplexer 302 are connected to the input ports of the second IQ modulator C317 and the second IQ modulator D318 in the IQ modulator array, respectively. (For simplification of the structural diagram...) Figure 2 The diagram only shows two channels for each of the upper and lower optical frequency combs. In practical applications, the number of channels can be increased in parallel as needed.

[0060] The signal output port of the first waveform generator A303 is connected to the input port of the first modulator driver A307; the signal output port of the first waveform generator B304 is connected to the input port of the first modulator driver B308; the signal output port of the second waveform generator C305 is connected to the input port of the second modulator driver C309; and the signal output port of the second waveform generator D306 is connected to the input port of the second modulator driver D310. The analog data input to the first waveform generator A303, first waveform generator B304, second waveform generator C305, and second waveform generator D306 can be directly set by the computer control system 8.

[0061] The output port of the first modulation driver A307 is connected to the input port of the first 90° phase shifter A311; the output port of the first modulation driver B308 is connected to the input port of the first 90° phase shifter B312; the output port of the second modulation driver C309 is connected to the input port of the second 90° phase shifter C313; and the output port of the second modulation driver D310 is connected to the input port of the second 90° phase shifter D314.

[0062] The two output ports of the first 90° phase shifter A311 are connected to the I and Q input ports of the first IQ modulator A315, respectively; the two output ports of the first 90° phase shifter B312 are connected to the I and Q input ports of the first IQ modulator B316, respectively; the two output ports of the second 90° phase shifter C313 are connected to the I and Q input ports of the second IQ modulator C317, respectively; and the two output ports of the second 90° phase shifter D314 are connected to the I and Q input ports of the second IQ modulator D318, respectively, so that the analog data input by the computer control system 8 is finally modulated onto the optical signal after passing through the modulation drive and the 90° phase shifters.

[0063] The optical output ports of the first IQ modulator A315 and the first IQ modulator B316 are connected to the input port of the first wavelength division multiplexer 319; the optical output ports of the second IQ modulator C317 and the second IQ modulator D318 are connected to the input port of the second wavelength division multiplexer 320. The optical output port of the first wavelength division multiplexer 319 is connected to the input port of the second fiber amplifier 402; the optical output port of the second wavelength division multiplexer 320 is connected to the second adjustable optical delay line 401. The second adjustable optical delay line 401 can be configured by the computer control system 8 to ensure that the two signal beams are in phase.

[0064] The output port of the second adjustable optical delay line 401 is connected to the input port of the third fiber amplifier 403; the output port of the second fiber amplifier 402 is connected to the input port ① of the 180° optical mixer 5; the output port of the third fiber amplifier 403 is connected to the input port ② of the 180° optical mixer 5. It can be understood that the 180° optical mixer 5 can be used to generate two optical signals with a phase difference of 180° that can be responded to by the optical detection system. Therefore, the two input optical signals corresponding to the second fiber amplifier 402 and the third fiber amplifier 403 will be mixed in the 180° optical mixer 5 to generate two optical signals, which will be output from the output ports ③ and ④ of the 180° optical mixer 5.

[0065] The output port ③ of the 180° optical mixer 5 is connected to the input port ① of the photoelectric balance detector 6; the output port ④ of the 180° optical mixer 5 is connected to the input port ② of the photoelectric balance detector 6, and the optical signal is converted into an electrical signal by the photoelectric balance detector 6; the output port of the photoelectric balance detector 6 is connected to the input port of the data acquisition device 7, and the output port of the data acquisition device 7 is connected to the computer control system 8.

[0066] It is understood that in this embodiment, the computer control system 8 can control the analog data of the arbitrary waveform generator, the optical path of the adjustable optical delay line, the bias voltage of the IQ modulator, and the detection parameters of the oscilloscope, so as to ultimately realize the flexible adjustability of the optical computing system, reduce system noise, and improve calculation accuracy.

[0067] The following will illustrate the implementation steps of the optoelectronic hybrid tensor convolution calculation system based on co-frequency dual optical combs and orthogonal frequency signals provided in this embodiment.

[0068] (1) The center wavelength of the optical frequency comb laser source 1 is set to 193.1 THz, the repetition frequency Δω is set to 50 GHz, and the average power is set to 0 dBm. To ensure that the input power of the photoelectric balanced detector 6 is sufficiently strong, the power gain of the fiber amplifier can be set as high as possible, for example, to 23 dB. For simplicity, as follows... Figure 2 As shown, this embodiment only provides the optical paths of the two channels corresponding to the upper and lower optical comb teeth. The wavelengths of the upper and lower paths are 193.1THz and 193.1THz+50GHz. Of course, the number of optical comb teeth can be increased according to the actual application needs.

[0069] (2) The first adjustable optical delay line 203 is set using the computer control system 8 to ensure that the two optical paths are in phase.

[0070] (3) Set the center wavelengths of the filters of the first dewavelength division multiplexer 301 and the second dewavelength division multiplexer 302 to 193.1THz and 193.1THz+50GHz respectively, so as to separate the optical combs of different channels.

[0071] (4) The computer control system 8 inputs analog data into the first waveform generator A303, the first waveform generator B304, the second waveform generator C305, and the second waveform generator D306, so that the first waveform generator A303, the first waveform generator B304, the second waveform generator C305, and the second waveform generator D306 generate corresponding electrical signals according to the input analog data; the two electrical signals are respectively passed through the first modulator driver A307, the first modulator driver B308, and the second modulator driver C308. 09. The second modulator driver D310 amplifies the signal and inputs it to the first 90° phase shifter A311, the first 90° phase shifter B312, the second 90° phase shifter C313, and the second 90° phase shifter D314, respectively. The generated I and Q electrical signals are input to the electrical modulation input ports of the first IQ modulator A315, the first IQ modulator B316, the second IQ modulator C317, and the second IQ modulator D318, respectively, thus loading the input analog data (i.e., input information) onto the optical frequency comb.

[0072] The output ports of the first IQ modulator A315 and the first IQ modulator B316 are both connected to the input port of the first wavelength division multiplexer 319. The center wavelengths of different channels in the first wavelength division multiplexer 319 are set to 193.1THz and 193.1THz+50GHz to complete the spectrum splicing of the uplink signal and achieve system noise reduction, i.e., M1(t) = M s1 (t)+M s2 M(t), where M1(t) represents the spectrum splicing result of the upstream signal, M s1 (t) represents the output signal of the first channel (i.e., the first IQ modulator A315) corresponding to the upper path, M s2 (t) represents the output signal of the second channel (i.e., the first IQ modulator B316) corresponding to the upper path. The output signal of each channel contains the corresponding noise.

[0073] The output ports of the second IQ modulators C317 and D318 are connected to the input port of the second wavelength division multiplexer 320. The center wavelengths of different channels in the second wavelength division multiplexer 320 are set to 193.1THz and 193.1THz+50GHz to complete the spectrum splicing of the downstream signal, thereby achieving system noise reduction in the downstream path, i.e., M2(t)=M s3 (t)+M s4 M(t), where M2(t) represents the spectrum splicing result of the downstream signal, M s3 (t) represents the output signal of the third channel corresponding to the lower path, M s4 (t) represents the output signal of the fourth channel corresponding to the lower path. The output signal of each channel contains the corresponding noise.

[0074] (5) The computer control system 8 is used to set the second adjustable optical delay line 401 to ensure that the two optical signals are in phase. The obtained in-phase optical signals are input to the second fiber amplifier 402 and the third fiber amplifier 403 respectively. The computer control system 8 is used to set the output gain of the second fiber amplifier 402 and the third fiber amplifier 403 to amplify the optical signal and improve the signal-to-noise ratio. However, it is ensured that the optical power input to the photoelectric balance detector 6 does not exceed the linear receiving range of the detector.

[0075] (6) Then, the two amplified optical signals are input to the 180° optical mixer 5 through input ports ① and ②. After passing through the 180° optical mixer 5, the two optical signals generate optical mixing signals with a phase difference of 0° and 180° respectively at output ports ③ and ④.

[0076] (7) The obtained optical mixing signals with a phase difference of 0° and 180° are input into the photoelectric balance detector 6, and a current is generated after photoelectric conversion. The current generated by the photoelectric balance detector 6 is collected by the data acquisition device 7 and finally uploaded to the computer control system 8 for processing to obtain the final calculation result. The expression for the electrical signal I(t) output by the photoelectric balance detector 6 is:

[0077]

[0078] In the formula, α represents the photoelectric conversion efficiency, C represents the amplitude of the light source, and A i B represents the data of the i-th input tensor. k This represents the data of the k-th weight tensor. This represents the initial frequency corresponding to the input tensor. The initial frequency corresponding to the weight tensor is represented by σ', where t represents time, ω0 represents the frequency interval of the electrical signal, and σ' represents the initial frequency of the weight tensor. ik This indicates the current noise output from the photoelectric balance detector 6 after noise reduction and filtering.

[0079] The following will explain the mathematical principles behind the convolution calculation implemented by the optoelectronic hybrid tensor convolution calculation system.

[0080] Let the tensors to be convolved be [A1, A2, ..., A... N ] and [B1,B2,...,B M The waveforms are input from the first waveform generator A303 and the first waveform generator B304 respectively. The waveform, where A i The data representing the input tensor is input through the second waveform generator C305 and the second waveform generator D306, respectively. The waveform, where B k The data representing the weight tensor.

[0081] Let the optical wavelengths of the optical frequency comb channel be Ccos(ω) c t), Ccos(ω c +Δωt), where C is the amplitude of the light source, ω c Let ω be the frequency of the center wavelength of the optical frequency comb, and Δω be the repetition frequency of the optical frequency comb. The output signals of the first IQ modulator A315, the first IQ modulator B316, the second IQ modulator C317, and the second IQ modulator D318 can be expressed as follows:

[0082]

[0083]

[0084]

[0085]

[0086] Among them, M s1 (t) represents the output signal of the first IQ modulator A315, M s2 (t) represents the output signal of the first IQ modulator B316, M s3 (t) represents the output signal of the second IQ modulator C317, M s4 (t) represents the output signal of the second IQ modulator D318. as well as These represent the current noise on the corresponding channel before noise reduction.

[0087] The four signals mentioned above are combined by a wavelength division multiplexer to achieve system noise reduction:

[0088] M1(t)=M s1 (t)+M s2 (t) (5)

[0089] M2(t)=M s3 (t)+M s4 (t) (6)

[0090] It should be understood that the noise in M1(t) and M2(t) are respectively and noise and Independent and uncorrelated, noise and Since they are independent and uncorrelated, they can partially cancel each other out after addition, so the signal-to-noise ratio of M1(t) will be higher than that of M1(t). and The signal-to-noise ratio of M2(t) will be higher than and Therefore, it is evident that the more parallel signal-loaded modulation units there are in the electro-optic modulation array structure, the higher the signal-to-noise ratio and the lower the impact of noise. Thus, this embodiment can effectively improve the system's signal-to-noise ratio and achieve noise reduction.

[0091] After the two combined optical signals are received by the input terminals ① and ② of the 180° optical mixer 5, respectively, optical mixing signals E1(t) and E2(t) with a phase difference of 0° and 180° are generated at the output ports ③ and ④.

[0092] E1(t)=M1(t)+M2(t) (7)

[0093] E2(t)=M1(t)-M2(t) (8)

[0094] Subsequently, the optical mixing signals E1(t) and E2(t) generated by the output ports ③ and ④ of the 180° optical mixer 5 are received and detected by the input ports ① and ② of the photoelectric balance detector 6, respectively, and the output electrical signals of the photoelectric balance detector 6 are I1(t) and I2(t), respectively. It should be noted that noise exists in the system, and the noise differs between different channels; see [reference needed]. Figure 3 As shown, before system noise reduction was performed, that is, before using... Figure 2 Before the physical structure shown is used for noise reduction, the current detected by the photoelectric balance detector 6 is:

[0095]

[0096] σ ik Indicates origin Figure 3 The current noise in the optical path where the first waveform generator 303 is located and the optical path where the second waveform generator 305 is located are shown.

[0097] In this embodiment, after noise reduction using formulas (5) and (6), the actual current detected by the photoelectric balance detector 6 is:

[0098]

[0099] Where, σ' 1,ik This indicates that, without considering the operating bandwidth limitation of the photoelectric balanced detector 6 (i.e., not considering noise due to bandwidth reduction), the noise originating from... Figure 2 The current noise, σ', is shown in the optical path where the first waveform generator A303 is located and the optical path where the second waveform generator C305 is located. 2,ik This indicates that, without considering the operating bandwidth limitation of the photoelectric balance detector 6, the source is... Figure 2 The current noise in the optical path where the first waveform generator B304 is located and the optical path where the second waveform generator D306 is located are shown. Compared with formula (9), the signal-to-noise ratio is improved.

[0100] However, due to the limited operating bandwidth of the photoelectric balance detector 6, the final actual current output of the photoelectric balance detector 6 after filtering is:

[0101]

[0102] Where α is the photoelectric conversion efficiency, determined by the performance of the photoelectric balance detector, and σ' ik This represents the final output current noise on the photoelectric balance detector 6. Therefore, it can be seen that the signal-to-noise ratio of formula (11) is higher than that of formula (9), i.e., using... Figure 2 The signal-to-noise ratio of the output signal of the architecture shown is higher than that of the unused architecture. Figure 2The signal-to-noise ratio of the output signal of the architecture shown is high, therefore the architecture proposed in this embodiment can effectively improve the calculation accuracy.

[0103] A bandpass filter can be used to filter frequencies of 10 ... According to formulas (1) to (6), the noise of the array based on the optical frequency comb is reduced after averaging compared to the unit structure based on a single wavelength light source. When reading the amplitude, the amplitude with less noise can be obtained.

[0104] It should be noted that the photoelectric hybrid tensor convolution calculation system in this embodiment needs to meet the following two conditions in terms of frequency setting, otherwise it will cause spectral errors and overlap, resulting in invalid results: (1) If z∈Z, then the mirror frequencies will be superimposed, i.e. (1) The amplitude of the frequency waveform is superimposed; (2) Δω must be large enough to ensure the frequency band. and frequency band and There is no overlap.

[0105] The electrical signal is acquired by the data acquisition device 7 and output to the computer control system 8.

[0106] Thus, the convolution result can be given by formula (11).

[0107] Where the input tensor [A1,A2,...A] N The length N and the weight tensor [B1, B2, ..., B] M The length M of the convolution operation is adjustable, thus enabling arbitrary convolution operations at the 8-level of the computer control system. This allows for adjustable configuration, making the system not only simple to operate and easy to adjust, but also more flexible and with greater throughput.

[0108] In summary, the optoelectronic hybrid tensor convolution calculation system based on orthogonal frequency signals provided in this embodiment employs dual optical frequency combs at the same frequency to extend the convolution calculation units, mainly composed of IQ modulators, mixers, and balanced detectors, in parallel. Channel averaging is performed using the results from multiple optical frequency comb teeth to reduce channel noise, and the final result is output using an oscilloscope. Therefore, this embodiment not only effectively reduces channel noise and improves the system's signal-to-noise ratio and convolution calculation accuracy, but also allows all modulation to be implemented optically, resulting in a large bandwidth. Furthermore, since the system uses all-fiber devices, it can be integrated, offering the advantage of easy on-chip integration, which is of great significance in the field of optical computing.

[0109] See Figure 4As shown in the embodiments of this application, a method for calculating optoelectronic hybrid tensor convolution is also provided, including the following steps:

[0110] Step S10: Split the obtained raw optical frequency comb and adjust it into two optical frequency combs with the same frequency and phase lock.

[0111] Step S20: Separate the optical frequency comb of each path to the corresponding multiple channels, load the input information onto the optical frequency comb of each channel respectively, and perform spectrum splicing on the output signals of the corresponding multiple channels of each path to achieve system noise reduction;

[0112] Step S30: Modulate the spectrum splicing results of the two channels into optical signals with equal optical path lengths, and mix the two optical signals with equal optical path lengths to generate two optical mixed signals with phase differences of 0° and 180° respectively.

[0113] Step S40: Perform photoelectric conversion on the two optical mixing signals, and determine the convolution result based on the output electrical signal.

[0114] Furthermore, the expression for the electrical signal I(t) output after photoelectric conversion is:

[0115]

[0116] In the formula, α represents the photoelectric conversion efficiency, C represents the amplitude of the light source, and A i B represents the data of the i-th input tensor. k This represents the data of the k-th weight tensor. This represents the initial frequency corresponding to the input tensor. The initial frequency corresponding to the weight tensor is represented by σ', where t represents the period, ω0 represents the frequency interval of the electrical signal, and σ' represents the frequency of the weight tensor. ik This indicates the current noise output after noise reduction, filtering, and photoelectric conversion.

[0117] Furthermore, each path corresponds to 2 channels, and the spectrum splicing result of each output signal is obtained using the following formula:

[0118] M1(t)=M s1 (t)+M s2 (t)

[0119] M2(t)=M s3 (t)+M s4 (t)

[0120] In the formula, M1(t) represents the spectrum splicing result of the upstream signal, M s1 (t) represents the output signal of the first channel corresponding to the upper path, M s2 M(t) represents the output signal of the second channel corresponding to the upper path, and M2(t) represents the spectrum splicing result of the lower path signal.s3 (t) represents the output signal of the third channel corresponding to the lower path, M s4 (t) represents the output signal of the fourth channel corresponding to the downlink, where the output signal of each channel contains its corresponding noise.

[0121] It should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the methods and steps described above can be referred to the corresponding process in the aforementioned embodiment of the optoelectronic hybrid tensor convolution calculation system, and will not be repeated here.

[0122] It should be noted that the step numbers in the embodiments of this application do not limit the order of operations in the technical solution of this application.

[0123] This application also provides an optoelectronic hybrid tensor convolution calculation device, including: a memory, a processor, and a network interface connected via a system bus. The memory stores at least one instruction, which is loaded and executed by the processor to implement all or part of the steps of the aforementioned optoelectronic hybrid tensor convolution calculation method.

[0124] The network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0125] A processor can be a CPU, or other general-purpose processors, DSPs (Digital Signal Processors), ASICs (Application Specific Integrated Circuits), FPGAs (Field Programmable Gate Arrays), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor, or any conventional processor. The processor is the control center of a computer device, connecting all parts of the computer device through various interfaces and lines.

[0126] Memory can be used to store computer programs and / or modules. The processor performs various functions of the computer device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can primarily include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on system usage, etc. Furthermore, memory may include high-speed random access memory (RAM), and may also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, SMC (Smart Media Card), SD (Secure Digital) cards, flash memory cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0127] This application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements all or part of the steps of the aforementioned photoelectric hybrid tensor convolution calculation method.

[0128] The embodiments of this application can implement all or part of the aforementioned processes, or they can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various methods described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, ROM (Read-Only memory), RAM (Random Access memory), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0129] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, servers, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0130] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0131] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0132] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A photoelectric hybrid tensor convolution calculation system, characterized in that, include: A laser source, used to generate the original optical frequency comb; An optical delay-splitting device is used to split the original optical frequency comb and adjust it into two optical frequency combs with the same frequency and phase lock-in. An electro-optic modulation array is used to separate the optical frequency comb of each path to multiple corresponding channels, load the input information onto the optical frequency comb of each channel, and splice the output signals of the multiple corresponding channels of each path to achieve system noise reduction. An optical delay device is used to modulate the splicing result of two spectra into an optical signal with equal optical path length. An optical mixer is used to mix two optical signals with equal optical path lengths to generate two optical mixed signals with a phase difference of 0° and 180°, respectively. A photoelectric balance detector is used to perform photoelectric conversion on two optical mixing signals to output an electrical signal; A data acquisition device used to acquire output electrical signals and determine the convolution result based on the acquired electrical signals; Each path corresponds to 2 channels. The spectrum splicing result of each output signal is obtained using the following formula: In the formula, This indicates the result of spectrum splicing of the upstream signal. This indicates the output signal of the first channel corresponding to the upper path. This indicates the output signal of the second channel corresponding to the upper path. This indicates the result of spectrum splicing of the downstream signal. This indicates the output signal of the third channel corresponding to the downstream path. This represents the output signal of the fourth channel corresponding to the downlink, where the output signal of each channel contains its corresponding noise.

2. The photoelectric hybrid tensor convolution calculation system as described in claim 1, characterized in that, The electro-optic modulation array includes: The first modulation component includes a first dewavelength division multiplexer, a plurality of parallel first signal loading modulation units, and a first wavelength division multiplexer connected in sequence. The second modulation component includes a second dewavelength division multiplexer, a plurality of parallel second signal loading modulation units, and a second wavelength division multiplexer connected in sequence.

3. The photoelectric hybrid tensor convolution calculation system as described in claim 2, characterized in that: The first signal loading modulation unit includes a first waveform generator, a first modulator driver, a first 90° electrical phase shifter, and a first IQ modulator connected in sequence. The second signal loading modulation unit includes a second waveform generator, a second modulator driver, a second 90° electrical phase shifter, and a second IQ modulator connected in sequence.

4. The photoelectric hybrid tensor convolution calculation system as described in claim 2, characterized in that, The electrical signal output by the photoelectric balance detector The expression is: In the formula, Indicates photoelectric conversion efficiency. Indicates the amplitude of the light source. Indicates the first i The data of the input tensor Indicates the first k Data for a weighted tensor This represents the initial frequency corresponding to the input tensor. This represents the initial frequency corresponding to the weight tensor. t Indicates time, Indicates the frequency interval of an electrical signal. This indicates the current noise output from the photoelectric balance detector 6 after noise reduction and filtering.

5. A method for calculating optoelectronic hybrid tensor convolution, characterized in that, Includes the following steps: The acquired raw optical frequency comb is split and adjusted into two optical frequency combs with the same frequency and phase lock. The optical frequency comb on each path is separated into multiple corresponding channels. The input information is loaded onto the optical frequency comb of each channel, and the output signals of the multiple corresponding channels of each path are spliced ​​together to achieve system noise reduction. The spectrum splicing results of the two channels are modulated into optical signals with equal optical path lengths, and the two optical signals with equal optical path lengths are mixed to generate two optical mixed signals with a phase difference of 0° and 180° respectively. The two optical mixing signals are converted into photoelectric signals, and the convolution result is determined based on the output electrical signal. Each path corresponds to 2 channels. The spectrum splicing result of each output signal is obtained using the following formula: In the formula, This indicates the result of spectrum splicing of the upstream signal. This indicates the output signal of the first channel corresponding to the upper path. This indicates the output signal of the second channel corresponding to the upper path. This indicates the result of spectrum splicing of the downstream signal. This indicates the output signal of the third channel corresponding to the downstream path. This represents the output signal of the fourth channel corresponding to the downlink, where the output signal of each channel contains its corresponding noise.

6. The photoelectric hybrid tensor convolution calculation method as described in claim 5, characterized in that, The electrical signal output after photoelectric conversion The expression is: In the formula, Indicates photoelectric conversion efficiency. Indicates the amplitude of the light source. Indicates the first i The data of the input tensor Indicates the first k Data for a weighted tensor This represents the initial frequency corresponding to the input tensor. This represents the initial frequency corresponding to the weight tensor. t Indicates time, Indicates the frequency interval of an electrical signal. This indicates the current noise output after noise reduction, filtering, and photoelectric conversion.

7. A photoelectric hybrid tensor convolution computing device, characterized in that, include: A memory and a processor, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the photoelectric hybrid tensor convolution calculation method of claim 5 or 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the photoelectric hybrid tensor convolution calculation method as described in claim 5 or 6.

Images