Photon convolutional neural network system based on polarization weighting and implementation method

Abstract

The invention discloses a photon convolutional neural network system based on polarization weighting and an implementation method, and the system generates light sources with different wavelengths through a tunable laser, adjusts the polarization state of an optical signal through a manual extrusion type polarization controller or an electric control polarization controller via a polarization maintaining optical fiber, combines the signals through a wavelength division multiplexer, and outputs the combined signals to the optical fiber. And an input signal generated by the high-speed digital-to-analog converter is loaded on the light intensity modulator. A modulated optical signal is amplified by an optical amplifier, one-bit information dislocation of different wavelength channels in a time domain is realized through a delay optical fiber, the optical signal is collected by a photoelectric detector and converted into a digital signal through an analog-to-digital converter, and then the digital signal is processed by a microprocessor, so that convolution operation formed by multiplication and addition is realized. The method is different from an existing convolution kernel loading mode, the angle of the polarization controller can be adjusted by directly utilizing the pre-trained convolution kernel weight, loading of the convolution kernel is achieved, the system cost is remarkably reduced, and the system complexity is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of photon convolutional neural networks, and in particular to a polarization-weighted photon convolutional neural network system and an implementation method thereof. Background Art

[0002] As an important model, convolutional neural network has shown great potential in image recognition, speech processing, natural language processing and other fields. However, traditional electronic computing devices face limitations in computing speed and energy consumption when processing large-scale data and complex operations, which has prompted people to seek new computing models and technical means to improve computing efficiency and processing capabilities.

[0003] As an emerging technology, optical computing uses the high-speed parallel transmission characteristics and large bandwidth advantages of optical signals to provide new possibilities for solving the bottleneck problems of traditional computing devices. As a cutting-edge technology that combines optical computing and deep learning, optical convolutional neural networks are attracting widespread attention and research in academia and industry. In optical convolutional neural networks, the application of optical devices is crucial. For example, devices such as tunable lasers, polarization controllers, wavelength division multiplexers, and optical intensity modulators can not only provide optical signal generation, modulation, and control functions, but also realize complex optical operations such as spectral processing, frequency selection, and wavelength modulation.

[0004] Although optical computing has the advantages of high speed and parallel processing in theory, it faces many challenges in practical applications, such as the stability of optical signals, the integration performance of devices, and the complexity of the system. Existing technologies have proposed a design based on an optical vector convolution accelerator, which realizes the simultaneous interleaving of time, wavelength, and space dimensions based on an integrated frequency comb light source (see [Xu XY, Tan MX, Corcoran B, et al. 11TOPS photonic convolutional accelerator for optical neural networks [J]. Nature 2021, 589 (7840): 44-51.]). An optimization method for the optical convolution kernel in the optoelectronic hybrid convolutional neural network (OHCNN) has also been proposed. The accuracy of the network architecture was tested using the Fashion-MNIST dataset, and the optimized optical convolution kernel increased the accuracy by 7.5%. The energy efficiency ratio (EER) of the dual-channel network is 46.7% higher than that of the single-channel network (see [Xu XF, Zhu LQ, Zhuang W, et al. Optimization of optical convolution kernel of optoelectronic hybrid convolution neural network[J]. Optoelectronics Letters, 2022, 18(3): 181-186.]).

[0005] With the advancement of optical technology and the continuous optimization of optical devices, optical convolutional neural networks have great application potential in the fields of large-scale data processing, high-efficiency computing, and real-time signal processing. They also have significant research significance in further reducing system complexity and lowering system costs. Summary of the invention

[0006] The problem to be solved by the present invention is to provide a polarization-weighted photon convolutional neural network system and an implementation method, which utilizes the high speed and parallel processing capabilities of the optical system to improve the computational efficiency of the convolutional neural network, is suitable for large-scale data processing and real-time signal processing scenarios, and can significantly reduce system costs and reduce system complexity.

[0007] The present invention adopts the following technical scheme: a polarization-weighted photon convolutional neural network system, comprising: a laser module, a polarization module, a wavelength division multiplexer, a light intensity modulator, an optical amplifier, a high-speed digital-to-analog converter, a time-delay optical fiber, a photodetector, an analog-to-digital converter, and a microprocessor;

[0008] The laser module is constructed by a number of tunable lasers, wherein N tunable lasers emit N wavelengths with equal wavelength intervals, serving as the light source part of the system;

[0009] The polarization module sets a corresponding number of polarization controllers based on the number of tunable lasers, and the specific number N is determined by the size of the convolution kernel; the N polarization controllers respectively receive the output signal of each tunable laser, and adjust the polarization state of the optical signal through the corresponding polarization controller;

[0010] The wavelength division multiplexer is connected to the polarization controller, combines the laser signals of different wavelengths emitted by the tunable laser and sends them to the light intensity modulator;

[0011] The high-speed digital-to-analog converter is used to generate a waveform of the input signal, and after the waveform is generated according to the one-dimensional signal processed by the microprocessor, it is loaded onto the light intensity modulator;

[0012] The optical amplifier is arranged after the optical intensity modulator, and is used to amplify the optical signal that needs to be subjected to optical convolution calculation, and output it to the photodetector through the delay optical fiber; the delay optical fiber is used to generate a time delay of one bit of information between the time domain waveforms of each wavelength;

[0013] The photoelectric detector is used to detect the intensity of the optical signal amplified by the optical amplifier, and output the total power of all wavelengths of light in each time interval, convert it into an electrical signal, and input it into the analog-to-digital converter;

[0014] The analog-to-digital converter converts the analog signal obtained by the photodetector into a digital signal and outputs it to the microprocessor for processing;

[0015] The microprocessor processes the output signal, processes the image data into a one-dimensional signal, and sends it to a high-speed digital-to-analog converter.

[0016] Furthermore, the polarization controller is connected to the output of the tunable laser through a polarization-maintaining fiber. The polarization controller outputs elliptically polarized light by adjusting the pressure on the polarization-maintaining fiber, and changes the direction angle of the elliptically polarized light by adjusting the angle, thereby ensuring that the power of optical signals entering different channels of the modulator is consistent and maximum.

[0017] The polarization controller is also connected to the wavelength division multiplexer via a polarization-maintaining optical fiber to maintain the polarization state of the optical signal.

[0018] The wavelength division multiplexer, light intensity modulator, optical amplifier, delay optical fiber, photoelectric detector, analog-to-digital converter, microprocessor and high-speed digital-to-analog converter are connected in sequence through single-mode optical fiber.

[0019] Furthermore, the polarization controller is a manual squeezing polarization controller or an electrically controlled polarization controller. When it is an electrically controlled polarization controller, the polarization controller is also connected to a microprocessor, and the electrical signal sent by the microprocessor controls the stepping motor to change the angle of the polarization controller.

[0020] Furthermore, in the polarization module, there is at least one group of polarization controllers. When there are multiple groups of polarization controllers, the output signal of the first group of polarization controllers is further adjusted by several subsequent groups of polarization controllers, and then enters the wavelength division multiplexer for signal merging, so as to ensure that the optical signal of each wavelength reaches the optimal polarization state before merging, thereby improving the extinction ratio.

[0021] The technical solution of the present invention also provides: a method for implementing a photon convolutional neural network based on polarization weighting, which is applied to any of the above-mentioned systems, and the convolution kernel is loaded onto the light wave, comprising the following steps:

[0022] S1. Input light source:

[0023] S1.1, N tunable lasers respectively output N wavelength optical signals with equal wavelength intervals as the system input light source;

[0024] S1.2, after the output of the tunable laser is connected to the corresponding polarization controller, the polarization control signals from different tunable lasers are combined through a wavelength division multiplexer, and the combined optical signal is connected to the optical intensity modulator;

[0025] S1.3. Before performing optical convolution, use a polarization controller to adjust the optical signal power of different channels to be consistent and to maximize the power as much as possible. Connect a spectrometer to the output of the optical intensity modulator and obtain the optical signal power of each wavelength by observing the waveform.

[0026] S2. Load the convolution kernel:

[0027] S2.1. Change the angle of the polarization controller and record the optical signal power corresponding to each angle;

[0028] S2.2, converting the pre-trained weight value into a power value, calculating the angle of the corresponding polarization controller, and controlling each polarization controller to adjust to a predetermined angle so that the polarization state of each optical signal matches the corresponding weight value;

[0029] S2.3, after being adjusted by the polarization controller, the optical signals enter the wavelength division multiplexer in turn for signal merging. During the merging process, each optical signal of different wavelength carries the information of the corresponding convolution kernel weight;

[0030] S3, implementation of multiplication and addition:

[0031] S3.1, the combined optical signal is modulated by an optical intensity modulator, and the input signal waveform generated by the high-speed digital-to-analog converter is loaded onto the optical intensity modulator, so that the optical signal produces a corresponding modulation effect in the time domain, thereby realizing the multiplication of the convolution kernel and the image signal;

[0032] S3.2. The modulated optical signal is input into an optical amplifier for amplification. The optical intensity modulator loads the bit stream at a certain rate. The light of several wavelength channels is loaded with the same data after passing through the modulator. After passing through the delay optical fiber, a time delay difference is generated between them, so that the data information they carry is displaced by one bit information time in the time domain.

[0033] S3.3, the delayed optical signal enters the photodetector for conversion, the optical signals that are dislocated from each other are received by the photodetector, and the values ​​obtained by the product of step S3.1 are accumulated to realize the accumulation of the convolution operation; the photodetector converts the optical signal into an electrical signal, which is converted into a digital signal after passing through an analog-to-digital converter, and is processed by a microprocessor to complete nonlinear, pooling, and fully connected operations, thereby constructing a complete photonic convolutional neural network;

[0034] S4. Output and processing:

[0035] The microprocessor processes the collected electrical signals, converts the image data into a one-dimensional signal, obtains the restored feature map after optical calculation, and sends it to a high-speed digital-to-analog converter.

[0036] Furthermore, the polarization controller changes the direction angle of elliptically polarized light by adjusting the angle. When the polarization controller is a manual squeezing polarization controller, the angle is adjusted manually. When the polarization controller is an electrically controlled polarization controller, the angle is changed by controlling a stepping motor through an electrical signal sent by a microprocessor.

[0037] Furthermore, in step S3.3, the photonic convolutional neural network includes a multi-layer optical convolutional network, which includes multiple optical convolutional modules connected in series or reuses the same group of optical convolutional modules multiple times, each optical convolutional module includes a tunable laser, a polarization controller, a wavelength division multiplexer, a light intensity modulator, an optical amplifier, a delay fiber, a photodetector and an analog-to-digital converter; the output of each optical convolutional module is used as the input of the next optical convolutional module, and multi-layer convolution is realized through a multi-layer structure.

[0038] Furthermore, in step S3.4, the microprocessor processes the collected electrical signal, expands the one-dimensional data into a two-dimensional image, obtains a feature map restored after optical calculation, stores the result through the storage device of the microprocessor, and / or directly uses the result for deep learning network training and reasoning process, displays it through a visualization tool, and sends it to a high-speed digital-to-analog converter.

[0039] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0040] 1. The present invention is based on the polarization-weighted photon convolutional neural network implementation method, which utilizes the high speed and parallel processing capabilities of the optical system to improve the computational efficiency of the convolutional neural network. It is suitable for large-scale data processing and real-time signal processing scenarios, and can significantly reduce system costs and reduce system complexity.

[0041] 2. The present invention is based on a polarization-weighted photon convolutional neural network system and implementation method, which utilizes polarization to load the convolution kernel, and effectively implements optical convolution operations by introducing polarization control and optical coupling technology, thereby improving the performance ratio of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a structural diagram of a photon convolutional neural network of a hand-controlled polarization controller of the present invention;

[0043] Figure 2 is a graph showing the relationship between the angle and power of the manual polarization controller used in the present invention;

[0044] Figure 3 It is a schematic diagram of implementing convolution calculation in the present invention;

[0045] Figure 4 It is the photon convolutional neural network composition of the electrically controlled polarization controller of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0047] The present invention discloses a polarization-weighted photon convolutional neural network system, comprising: a laser module, a polarization module, a wavelength division multiplexer, a light intensity modulator, an optical amplifier, a high-speed digital-to-analog converter, a time-delay optical fiber, a photodetector, an analog-to-digital converter, and a microprocessor. Based on the system, polarization is further utilized to realize the loading of the convolution kernel, which can significantly reduce the system cost and reduce the system complexity.

[0048] Embodiment 1:

[0049] Polarization-weighted photon convolutional neural network systems, such as Figure 1As shown, it includes: a laser module, a polarization module, a wavelength division multiplexer, a light intensity modulator, an optical amplifier, a high-speed digital-to-analog converter, a delay optical fiber, a photodetector, and an analog-to-digital converter.

[0050] The laser module includes N tunable lasers, and the polarization module includes N manual squeezing fiber polarization controllers, where the size of N is specifically determined by the size of the convolution kernel.

[0051] Based on the above-mentioned polarization-weighted photon convolutional neural network implementation method, it includes the following parts:

[0052] 1. Input light source part:

[0053] Step 1.1: Build a laser module consisting of N tunable lasers.

[0054] In particular, in this embodiment, N is four, and four tunable lasers are constructed, each using four wavelengths with equal wavelength intervals, corresponding to a 2×2 convolution kernel, as the light source part of the entire system;

[0055] Step 1.2: The laser output is connected to a manual squeezed fiber polarization controller. The wavelength division multiplexer is used to combine the polarization control signals from different lasers, and the combined optical signal is connected to the modulator.

[0056] Step 1.3: Since the modulator is sensitive to polarization, before performing the multiplication and addition of optical convolution, use a manual squeezed fiber polarization controller to adjust the optical signal power of different channels to be consistent and as large as possible. It is mainly used to adjust the elliptical polarization. Different powers will affect the subsequent addition and multiplication results. The power of each wavelength can be observed by connecting a spectrometer to the modulator output to observe the waveform;

[0057] 2. After adjusting the input light source, load the convolution kernel. The specific steps are as follows:

[0058] Step 2.1: Change the angle of the polarization controller and record the optical signal power corresponding to each angle;

[0059] Step 2.2: Convert the pre-trained weight values ​​into power values ​​and calculate the corresponding angle of the manual squeezed fiber polarization controller.

[0060] The angle and power relationship of the manual squeezed optical fiber polarization controller used in this embodiment is as follows: Figure 2 As shown, when the manual squeezed fiber polarization controller angle is 48° and 88°, the minimum power -14.2dbm can be obtained, and when the manual squeezed fiber polarization controller angle is 67°, the maximum power -1.7dbm can be obtained. Each manual squeezed fiber polarization controller is controlled to adjust to a predetermined angle so that the polarization state of each optical signal matches the corresponding weight value;

[0061] Step 2.3: Two or more sets of manual squeezed fiber polarization controllers can be used, and steps 2.1 and 2.2 are repeated. For example, the output signal of the first set of manual squeezed fiber polarization controllers is further adjusted by the second set of manual squeezed fiber polarization controllers, and finally enters the wavelength division multiplexer for signal merging;

[0062] Step 2.4: After the optical signal is adjusted by the manual squeeze fiber polarization controller, it enters the wavelength division multiplexer for signal merging. During the merging process, each optical signal of different wavelength will carry the information of the corresponding convolution kernel weight;

[0063] 3. After the convolution kernel is loaded, perform multiplication and addition operations, such as Figure 3 As shown, the specific steps are as follows:

[0064] Step 3.1: The combined optical signal is modulated by an optical intensity modulator, and the input signal waveform generated by the high-speed digital-to-analog converter is loaded onto the modulator.

[0065] In this embodiment, the rate of the high-speed digital-to-analog converter is 12.5 Gbps, the vertical resolution is 10 bits, the bandwidth is 5.3 GHz, and the corresponding time for sending one bit is 80 ps, ​​so that the optical signal produces a corresponding modulation effect in the time domain, and finally realizes the multiplication of the convolution kernel and the one-dimensional image signal.

[0066] The method of expanding a two-dimensional image signal into a one-dimensional signal is as follows:

[0067] First, the input image is divided into blocks, and each convolution kernel is used to slide the window according to the step size to extract the covered area, and each area is expanded into a row vector; finally, all the expanded column vectors are arranged in order into a large row vector; the convolution kernel is expanded into a row vector;

[0068] Step 3.2: The modulated optical signal enters the optical bit information amplifier for amplification to ensure that the signal strength is sufficient for subsequent processing. The modulator loads the bit stream at a certain rate, and the light of the four wavelength channels is loaded with the same data after passing through the modulator. After passing through the delay fiber, a time delay difference is generated between them, which just causes the data information they carry to be misaligned by one bit in the time domain;

[0069] Step 3.3: The delayed optical signal enters the photodetector for conversion.

[0070] In this embodiment, the bandwidth of the photodetector is 40 GHz, and the photodetector converts the optical signal into an electrical signal. The collected electrical signal is processed by a microprocessor to complete nonlinear, pooling, full connection and other operations to build a complete photonic convolutional neural network.

[0071] After completing the addition and multiplication in the convolution operation, the result is finally output and processed:

[0072] The microprocessor processes the collected electrical signals and re-expands the one-dimensional data into a two-dimensional image, that is, the restored feature map after optical calculation is obtained. The processing results can be stored through the storage device of the microprocessor, or directly used in the subsequent deep learning network training and reasoning process. The calculation results can be displayed through visualization tools to verify the implementation effect and performance of the convolutional neural network;

[0073] In addition, it can also be extended to the implementation of multi-layer convolution. The specific steps are as follows:

[0074] Repeating the above steps can realize a multi-layer optical convolution network structure, which includes multiple optical convolution modules connected in series, each module including a tunable laser, a polarization controller, a wavelength division multiplexer, an optical intensity modulator, a delay fiber, a photodetector and an analog-to-digital converter. The output of each optical convolution module is used as the input of the next module, and a deeper convolution operation is realized through a multi-layer structure. In addition, the same group of modules can be reused multiple times to realize multi-layer convolution.

[0075] Embodiment 2:

[0076] The above-mentioned embodiment 1 is based on the optical convolutional neural network implementation method of the traditional manual squeezed fiber polarization controller weighting, but the manual adjustment of the angle is relatively rough. Therefore, the manual squeezed fiber polarization controller can be replaced by an electrically controlled polarization controller to achieve weighting;

[0077] In this embodiment, the electrically controlled polarization controller has a fast response speed and high precision, and can achieve more flexible polarization state control. The electrically controlled polarization controller implements a weighted alternative solution, such as Figure 4 As shown, the specific steps are as follows:

[0078] Step 1: Construct a laser module consisting of N tunable lasers, for example, construct four tunable lasers, using four wavelengths with equal wavelength intervals, corresponding to a 2×2 convolution kernel, as the light source part of the entire system;

[0079] Step 2: The laser output is connected to the electrically controlled polarization controller, and the wavelength division multiplexer is used to combine the polarization control signals from different lasers, and the combined optical signal is connected to the modulator;

[0080] Step 3: Before performing the multiplication and addition of the optical convolution, the electrical signal sent by the microprocessor controls the stepper motor to change the angle of the polarization controller, and the electrically controlled polarization controller is used to adjust the optical signal power of different channels to be consistent and to the maximum value as much as possible. The power of each wavelength can be observed by connecting the spectrometer to the modulator output waveform;

[0081] Step 4: Change the angle of the polarization controller and record the optical signal power corresponding to each angle;

[0082] Step 5: Convert the pre-trained weight value into a power value, calculate the corresponding angle of the polarization controller, and control the stepper motor to change the angle of the polarization controller through the electrical signal sent by the microprocessor, and adjust it to a predetermined angle so that the polarization state of each optical signal matches the corresponding weight value;

[0083] Step 6: Two or more polarization controllers can be used, and the above steps 4 and 5 can be repeated. For example, the output signal of the first set of polarization controllers is further adjusted by the second set of polarization controllers, and finally enters the wavelength division multiplexer for signal merging. Through this multi-level adjustment, it can be ensured that the optical signal of each wavelength reaches the optimal polarization state before merging, thereby improving the extinction ratio.

[0084] Step 7: The optical signal adjusted by the electrically controlled polarization controller enters the wavelength division multiplexer for signal merging. The merged optical signal carries the information of the corresponding convolution kernel weight.

[0085] Step 8: The combined optical signal is modulated by an optical intensity modulator, and the input signal waveform generated by the high-speed digital-to-analog converter is loaded onto the modulator, so that the optical signal produces a corresponding modulation effect in the time domain.

[0086] Step 9: The modulated optical signal enters the optical amplifier for amplification to ensure that the signal strength is sufficient for subsequent processing. The amplified optical signal passes through the delay fiber, and a 1-bit delay is generated between each wavelength channel to achieve the addition operation.

[0087] Step 10: The delayed optical signal enters the photodetector for conversion, and the photodetector converts the optical signal into an electrical signal. The collected electrical signal is processed by a microprocessor, and the one-dimensional signal can be restored to a feature map on the microprocessor for subsequent neural network training.

[0088] Step 11: Repeating the above steps can realize a multi-layer optical convolution network structure. The multi-layer optical convolution network includes multiple optical convolution modules connected in series, each module includes a tunable laser, a polarization controller, a wavelength division multiplexer, an optical intensity modulator, an optical amplifier, a delay fiber, a photodetector, and an analog-to-digital converter. The output of each optical convolution module is used as the input of the next module, and a deeper convolution operation is realized through a multi-layer structure. In addition, the same set of modules can be reused multiple times to play the role of multi-layer convolution.

[0089] In summary, the present invention aims to propose a method for realizing a photonic convolutional neural network based on polarization weighting, which significantly improves the computational efficiency of the convolutional neural network by utilizing the high speed and parallel processing capabilities of the optical system, and is suitable for large-scale data processing and real-time signal processing scenarios. By introducing polarization control and optical coupling technology, the present invention can effectively realize optical convolution operations, improve the performance ratio of the system, and provide a new idea for the realization of photonic convolutional neural networks.

[0090] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A polarization-weighted photon convolutional neural network system, characterized in that: include: Laser module, polarization module, wavelength division multiplexer, light intensity modulator, optical amplifier, high-speed digital-to-analog converter, delay fiber, photodetector, analog-to-digital converter, microprocessor, The laser module is composed of a number of tunable lasers and serves as the light source part of the system; The polarization module is provided with a corresponding number of polarization controllers based on the number of tunable lasers, receives the output signal of each tunable laser respectively, and adjusts the polarization state of the optical signal through the corresponding polarization controller; The wavelength division multiplexer is connected to the polarization controller, combines the laser signals of different wavelengths emitted by the tunable laser and sends them to the light intensity modulator; The high-speed digital-to-analog converter is used to generate a waveform of the input signal, and after the waveform is generated according to the one-dimensional signal processed by the microprocessor, it is loaded onto the light intensity modulator; The optical amplifier is arranged after the light intensity modulator, and is used to amplify the optical signal that needs to be subjected to optical convolution calculation, and output it to the photodetector through the delay optical fiber; The delay optical fiber is used to generate a time delay of one bit of information between the time domain waveforms of each wavelength; The photoelectric detector is used to detect the intensity of the optical signal amplified by the optical amplifier, and output the total power of all wavelengths of light in each time interval, convert it into an electrical signal, and input it into the analog-to-digital converter; The analog-to-digital converter converts the analog signal obtained by the photodetector into a digital signal and outputs it to the microprocessor for processing; The microprocessor processes the output signal, processes the image data into a one-dimensional signal, and sends it to a high-speed digital-to-analog converter.

2. The polarization-weighted photon convolutional neural network system according to claim 1, characterized in that: In the laser module, the number N of tunable lasers is determined by the size of the convolution kernel, and the N tunable lasers emit N wavelengths with equal wavelength intervals, which are respectively input into each polarization controller in the polarization module.

3. The polarization-weighted photon convolutional neural network system according to claim 2, characterized in that: The polarization controller is connected to the output of the tunable laser through a polarization-maintaining fiber. The polarization controller outputs elliptically polarized light by adjusting the pressure on the polarization-maintaining fiber, and changes the direction angle of the elliptically polarized light by adjusting the angle, so as to ensure that the power of the optical signals entering different channels of the modulator is consistent and maximum. The polarization controller is also connected to the wavelength division multiplexer via a polarization-maintaining optical fiber to maintain the polarization state of the optical signal.

4. The polarization-weighted photon convolutional neural network system according to claim 1, characterized in that: The polarization controller is a manual squeezing polarization controller or an electrically controlled polarization controller. When it is an electrically controlled polarization controller, the polarization controller is also connected to a microprocessor, and the electrical signal sent by the microprocessor controls the stepping motor to change the angle of the polarization controller.

5. The polarization-weighted photon convolutional neural network system according to claim 1, characterized in that: In the polarization module, there is at least one group of polarization controllers. When there are multiple groups of polarization controllers, the output signal of the first group of polarization controllers is further adjusted by several subsequent groups of polarization controllers, and then enters the wavelength division multiplexer for signal merging, so as to ensure that the optical signal of each wavelength reaches the optimal polarization state before merging, thereby improving the extinction ratio.

6. The polarization-weighted photon convolutional neural network system according to claim 1, characterized in that: The wavelength division multiplexer, light intensity modulator, optical amplifier, delay optical fiber, photoelectric detector, analog-to-digital converter, microprocessor and high-speed digital-to-analog converter are connected in sequence through single-mode optical fiber.

7. A method for implementing a polarization-weighted photon convolutional neural network, applied to the system according to any one of claims 1 to 6, wherein the convolution kernel is loaded onto a light wave, characterized in that: The steps include: S1. Input light source: S1.1, N tunable lasers respectively output N wavelength optical signals with equal wavelength intervals as the system input light source; S1.2, after the output of the tunable laser is connected to the corresponding polarization controller, the polarization control signals from different tunable lasers are combined through a wavelength division multiplexer, and the combined optical signal is connected to the optical intensity modulator; S1.

3. Before performing optical convolution, use a polarization controller to adjust the optical signal power of different channels to be consistent and to maximize the power as much as possible. Connect a spectrometer to the output of the optical intensity modulator and obtain the optical signal power of each wavelength by observing the waveform. S2. Load the convolution kernel: S2.

1. Change the angle of the polarization controller and record the optical signal power corresponding to each angle; S2.2, converting the pre-trained weight value into a power value, calculating the angle of the corresponding polarization controller, and controlling each polarization controller to adjust to a predetermined angle so that the polarization state of each optical signal matches the corresponding weight value; S2.3, after being adjusted by the polarization controller, the optical signals enter the wavelength division multiplexer in turn for signal merging. During the merging process, each optical signal of different wavelength carries the information of the corresponding convolution kernel weight; S3, implementation of multiplication and addition: S3.1, the combined optical signal is modulated by an optical intensity modulator, and the input signal waveform generated by the high-speed digital-to-analog converter is loaded onto the optical intensity modulator, so that the optical signal produces a corresponding modulation effect in the time domain, thereby realizing the multiplication of the convolution kernel and the image signal; S3.

2. The modulated optical signal is input into an optical amplifier for amplification. The optical intensity modulator loads the bit stream at a certain rate. The light of several wavelength channels is loaded with the same data after passing through the modulator. After passing through the delay optical fiber, a time delay difference is generated between them, so that the data information carried by the light of each wavelength channel is displaced by one bit information time in the time domain. S3.3, the delayed optical signal enters the photodetector for conversion, the optical signals that are dislocated from each other are received by the photodetector, and the values ​​obtained by the product of step S3.1 are accumulated to realize the accumulation of the convolution operation; the photodetector converts the optical signal into an electrical signal, which is converted into a digital signal after passing through an analog-to-digital converter, and is processed by a microprocessor to complete nonlinear, pooling, and fully connected operations, thereby constructing a complete photonic convolutional neural network; S4. Output and processing: The microprocessor processes the collected electrical signals, converts the image data into a one-dimensional signal, obtains the restored feature map after optical calculation, and sends it to a high-speed digital-to-analog converter.

8. The method for implementing a polarization-weighted photon convolutional neural network according to claim 7, characterized in that: The polarization controller changes the direction angle of elliptically polarized light by adjusting the angle. When the polarization controller is a manual squeezing polarization controller, the angle is adjusted manually. When the polarization controller is an electrically controlled polarization controller, the angle is changed by controlling a stepping motor through an electrical signal sent by a microprocessor.

9. The method for implementing a polarization-weighted photon convolutional neural network according to claim 7, characterized in that: In step S3.3, the photonic convolutional neural network includes a multi-layer optical convolutional network, which includes multiple optical convolutional modules connected in series or the same group of optical convolutional modules is reused multiple times, each optical convolution module includes a tunable laser, a polarization controller, a wavelength division multiplexer, a light intensity modulator, an optical amplifier, a delay fiber, a photodetector and an analog-to-digital converter; the output of each optical convolution module is used as the input of the next optical convolution module, and multi-layer convolution is realized through a multi-layer structure.

10. The method for implementing a polarization-weighted photon convolutional neural network according to claim 7, characterized in that: In step S3.4, the microprocessor processes the collected electrical signal, expands the one-dimensional data into a two-dimensional image, obtains a feature map restored after optical calculation, stores the result through the storage device of the microprocessor, and / or directly uses the result for deep learning network training and reasoning process, displays it through a visualization tool, and sends it to a high-speed digital-to-analog converter.

Images