Optical diffraction neural network system capable of simultaneously multiplexing wavelength and polarization

Abstract

The invention provides an optical diffraction neural network system capable of simultaneously multiplexing wavelength and polarization, and the system is characterized in that the system comprises a light source module which is used for generating N parallel lights with different wavelengths; the polarized parallel light generation module is used for converting each parallel light into corresponding polarized light; the encoding module is used for performing polarization encoding according to each input image and the corresponding polarized light to obtain a corresponding encoding light path; the diffraction module is used for generating a light path according to all the converged coded light paths; the light splitting module is used for dividing the light path into N * 2 sub light paths corresponding to different calculation tasks according to the wavelength and the polarization direction; and the detection module is used for generating a calculation result corresponding to each sub light path. In a word, the method can simultaneously multiplex wavelength and polarization to process a plurality of identification tasks, and an accurate identification result is obtained.

Description

Technical Field

[0001] The present invention relates to an optical diffraction neural network, in particular to an optical diffraction neural network system that simultaneously multiplexes wavelength and polarization. Background Art

[0002] The optical diffraction neural network (D2NN) is a relatively new optical physics neural network that combines wavefront modulation and diffraction calculations. It offers advantages such as high speed, high throughput, and low power consumption. Compared to traditional neural networks based on electronic integrated platforms such as GPUs, D2NN uses variable optical modulation, such as phase modulation, on different pixels at different modulation layers, leveraging the complex amplitude weights of diffraction propagation between different layers to create a realistic optical physics network similar to a fully connected network. D2NN has achieved promising results in image recognition, scattering recovery, vortex light modulation, and parallel computing.

[0003] Traditional D2NN networks are forward-deduced based on the scalar approximation theory of light waves. Of the various degrees of freedom of light waves, only the intensity and phase dimensions are utilized. Research into applying more intervening dimensions that describe the properties of light wave objects, such as wavelength, polarization, and vortex states, to D2NNs is gaining attention.

[0004] Existing single-wavelength multiplexing diffractive optical networks (DONs) enable multi-channel parallel computing. However, the choice of wavelength significantly impacts network performance, and to ensure the performance of each channel, the spectral range that can be reused is limited. While single-polarization multiplexing diffractive optical networks offer unique advantages in multiple-input, multiple-output (MIMO) tasks, crosstalk between different channels can reduce the number of channels that can be reused.

[0005] In short, the parallel computing of existing diffraction optical networks still has the problem of low upper limit. Summary of the Invention

[0006] The present invention is made to solve the above problems, and its purpose is to provide an optical diffraction neural network system that simultaneously multiplexes wavelength and polarization.

[0007] The present invention provides an optical diffraction neural network system that simultaneously multiplexes wavelength and polarization, and is used to obtain corresponding calculation results based on N*2 input images corresponding to different calculation tasks. It has the following characteristics, including: a light source module, used to generate N parallel lights of different wavelengths; a polarized parallel light generation module, used to convert each parallel light into a first polarized light corresponding to a first polarization direction and a second polarized light corresponding to a second polarization direction, and use the first polarized light and the second polarized light as polarized light; an encoding module, including N*2 digital modulation plates, used to perform polarization encoding based on each input image and the corresponding polarized light, to obtain a corresponding encoded light path; a diffraction module, used to generate an optical path based on all the converged encoded light paths; a splitting module, used to divide the optical path into N*2 sub-optical paths corresponding to different calculation tasks based on wavelength and polarization direction; and a detection module, including a CMOS camera, used to generate the calculation results corresponding to each sub-optical path.

[0008] The optical diffraction neural network system for simultaneously multiplexing wavelength and polarization provided by the present invention may also have the following feature: the frequencies corresponding to all wavelengths are in the same band.

[0009] The optical diffraction neural network system for simultaneously multiplexing wavelength and polarization provided by the present invention may also have the following characteristics: the first polarization direction is along the x-axis direction, and the second polarization direction is along the y-axis direction.

[0010] The optical diffraction neural network system for simultaneously multiplexing wavelength and polarization provided by the present invention may also have the following features: wherein the diffraction module includes a first phase modulation layer, a second phase modulation layer, a third phase modulation layer, a linear polarization array layer, a fourth phase modulation layer, a fifth phase modulation layer and a sixth phase modulation layer arranged in sequence, and the linear polarization array layer is used to perform different operations on linearly polarized light incident at different angles.

[0011] The optical diffraction neural network system for simultaneously multiplexing wavelength and polarization provided by the present invention may also have the following features: wherein, corresponding training data sets are constructed according to different computing tasks, the optical diffraction neural network system is trained according to the training data sets, and the phase plates of the first phase modulation layer, the second phase modulation layer, the third phase modulation layer, the fourth phase modulation layer, the fifth phase modulation layer and the sixth phase modulation layer and the linear polarization array plate of the linear polarization array layer are manufactured according to the training results.

[0012] The optical diffraction neural network system for simultaneously multiplexing wavelength and polarization provided by the present invention may also have the following characteristics: wherein the expression of the loss function used in training is: Where N is the total number of channels, is the two-dimensional light intensity distribution of the optical network output surface on the i-th channel, is the two-dimensional distribution of the target true light intensity actually corresponding to the i-th channel during training.

[0013] The optical diffraction neural network system for simultaneously multiplexing wavelength and polarization provided by the present invention may also have the following features: wherein the optical splitting module includes a filter and an analyzer, and the optical path is divided into N*2 sub-optical paths through the filter and the analyzer in sequence.

[0014] Functions and effects of the invention

[0015] According to the optical diffraction neural network system for simultaneous wavelength and polarization multiplexing involved in the present invention, multiple polarized lights of different wavelengths and polarization directions are generated by the light source module and the polarized parallel light generation module, the encoding module performs polarization encoding based on the corresponding computing task in combination with the input image and polarized light to obtain an encoded light path, the diffraction module including the phase modulation layer and the linear polarization array layer performs corresponding calculations on each encoded light path to obtain an optical path, and the optical path is re-divided into sub-optical paths corresponding to each computing task by the optical splitting module and the detection module, and detection is performed to obtain the computing results. Therefore, the optical diffraction neural network system for simultaneous wavelength and polarization multiplexing of the present invention can simultaneously multiplex wavelength and polarization to process multiple computing tasks in parallel and obtain accurate computing results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 1 is a schematic structural diagram of an optical diffraction neural network system in an embodiment of the present invention;

[0017] Figure 2 is a schematic diagram of the results of matrix operations in an embodiment of the present invention;

[0018] Figure 3 1 is a schematic diagram of the process of handwritten digital image recognition in an embodiment of the present invention;

[0019] Figure 4 Schematic diagram of calculation results of a handwritten digital image in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments and the accompanying drawings specifically illustrate the optical diffraction neural network system of the present invention that simultaneously multiplexes wavelength and polarization.

[0021] This embodiment provides an optical diffraction neural network system that simultaneously multiplexes wavelength and polarization, and is used to obtain corresponding computational results based on N*2 input images corresponding to different computational tasks. In this embodiment, N=4, the input images are handwritten digit images, and the computational tasks are classification tasks, including a ten-digit classification task, a two-digit parity classification task, and a three-digit quality classification task, with the corresponding classification results being used as the computational results.

[0022] Figure 1 Schematic diagram of the structure of the optical diffraction neural network system in an embodiment of the present invention.

[0023] like Figure 1 As shown, the optical diffraction neural network system 100 includes a light source module 10, a polarized parallel light generation module 20, an encoding module 30, a diffraction module 40, a spectrometer module 50 and a detection module 60.

[0024] The light source module 10 includes N=4 monochromatic light sources, namely monochromatic light source 1, monochromatic light source 2, monochromatic light source 3 and monochromatic light source 4, which are used to generate corresponding parallel light. In this embodiment, the frequencies corresponding to the wavelengths of the four parallel lights are all in the terahertz band, which are 0.4e 12 Hz, 0.43e 12 Hz, 0.46e 12 Hz and 0.5e 12 Hz. Depending on the actual situation, other wavelengths can be used to perform the same multi-polarization channel calculation task.

[0025] The polarized parallel light generation module 20 is configured to convert each parallel light beam into a first polarized light beam corresponding to a first polarization direction and a second polarized light beam corresponding to a second polarization direction, with the first polarized light beam and the second polarized light beam being used as polarized light. In this embodiment, the polarized parallel light generation module 20 splits each parallel light beam into two beams, each of which then passes through a corresponding linear polarizer to generate the corresponding polarized light beam.

[0026] The first polarization direction is along the x-axis, and the second polarization direction is along the y-axis.

[0027] The encoding module 30 includes N*2=8 digital modulation plates, which are used to perform polarization encoding based on each input image and the corresponding polarized light, thereby generating a corresponding encoded optical path. In this embodiment, the corresponding input image is loaded through the digital modulation plates, and polarization encoding is performed based on the corresponding polarized light, thereby generating a corresponding encoded optical path. In this embodiment, there are a total of N*2=8 encoded optical paths, corresponding to 8 different input images.

[0028] The diffraction module 40 is used to generate a light path according to all the converged coded light paths.

[0029] In this embodiment, the diffraction module 40 includes a first phase modulation layer 401, a second phase modulation layer 402, a third phase modulation layer 403, a linear polarization array layer 404, a fourth phase modulation layer 405, a fifth phase modulation layer 406, and a sixth phase modulation layer 407. The linear polarization array layer 404 is used to perform different operations on linearly polarized light incident at different angles.

[0030] In this embodiment, corresponding training data sets are constructed according to different computing tasks, the optical diffraction neural network system 100 is trained according to the training data sets, and the phase plates of the first phase modulation layer 401, the second phase modulation layer 402, the third phase modulation layer 403, the fourth phase modulation layer 405, the fifth phase modulation layer 406 and the sixth phase modulation layer 407, as well as the linear polarizer array plate of the linear polarization array layer 404 are manufactured according to the training results.

[0031] Among them, the expression of the loss function used in training is:

[0032]

[0033] Where N is the total number of channels, is the two-dimensional light intensity distribution of the optical network output surface on the i-th channel, is the two-dimensional distribution of the target true light intensity actually corresponding to the i-th channel during training.

[0034] In this embodiment, based on the training results, the width of the phase plates in the first phase modulation layer 401, the second phase modulation layer 402, the third phase modulation layer 403, the fourth phase modulation layer 405, the fifth phase modulation layer 406 and the sixth phase modulation layer 407 are all set to 0.08m, and all contain 100×100 pixels. The phase plates are manufactured by processes such as photolithography printing.

[0035] In this embodiment, based on the training results, the linear polarizer array plate in the linear polarization array layer 404 is set to include 10×10 polarization angles, and the size of each local linear polarization is 8mm×8mm. The linear polarizer array plate is designed using a metamaterial design method.

[0036] The optical splitting module 50 is used to split the optical path into N*2 sub-optical paths corresponding to different computing tasks according to wavelength and polarization direction.

[0037] In this embodiment, the light splitting module 50 includes a filter 501 and an analyzer 502 , and the light path is sequentially divided into N*2 sub-light paths via the filter 501 and the analyzer 502 .

[0038] The detection module 60 includes a CMOS camera 601, which is used to generate calculation results corresponding to each sub-light path.

[0039] In this embodiment, the eight channels of the optical diffraction neural network system 100 each perform a different matrix operation task, randomly generate an input matrix 3*3 as an input vector, set eight transformation matrices 9*9 as different transformation processes of the matrix operation of each channel, and finally take the calculation result of the transformation matrix 9*9 and the input matrix 3*3 as the ideal output 3*3 and compare it with the network output 3*3 generated by the optical diffraction neural network system 100 according to the input matrix 3*3, thereby verifying the effectiveness of the optical diffraction neural network system 100 in processing multiple tasks in parallel.

[0040] The expression of the input matrix 3*3 is:

[0041]

[0042] Where X k The input matrix 3*3 corresponding to the kth channel.

[0043] The calculation expression of the ideal output 3*3 is:

[0044] Y k =[W k ]X k ,

[0045] Where [W k ] is the transformation matrix 9*9 corresponding to the kth channel, Y k is the ideal output 3*3 corresponding to the kth channel.

[0046] Figure 2 Schematic diagram of the result of matrix operation in an embodiment of the present invention.

[0047] like Figure 2 As shown, the first column is the different wavelengths used in the optical diffraction neural network system 100, namely λ1, λ2, λ3 and λ4, the second column is the different polarization directions in the optical diffraction neural network system 100, namely the horizontal x-axis direction and the vertical y-axis direction. The four wavelengths and two polarization directions are combined to generate 8 different channels. The third column is the input matrix 3*3 corresponding to each channel, the fourth column is the transformation matrix 9*9 corresponding to each channel, the fifth column is the ideal output 3*3 corresponding to each input matrix 3*3 and the corresponding transformation matrix 9*9, and the sixth column is the network output 3*3 corresponding to each channel. It can be seen that the ideal output 3*3 is basically consistent with the network output 3*3, so the optical diffraction neural network system 100 can effectively realize the parallel calculation of 8 different tasks.

[0048] The following describes the process of performing handwritten digital image recognition using the optical diffraction neural network system that simultaneously multiplexes wavelength and polarization according to this embodiment with reference to the accompanying drawings.

[0049] Figure 3 1 is a flow chart of handwritten digital image recognition in an embodiment of the present invention.

[0050] like Figure 3 As shown, the optical diffraction neural network system 100 is used to complete eight different computing tasks on a handwritten digital image, including the following steps:

[0051] In step S1 , the light source module 10 is used to generate parallel lights of four different wavelengths.

[0052] In step S2 , the polarized parallel light generating module 20 is used to convert each parallel light into two polarized lights with different polarization directions.

[0053] In step S3 , the encoding module 30 is used to perform polarization encoding according to each input image and the corresponding polarized light to obtain a corresponding encoding light path.

[0054] Step S4: using the diffraction module 40 to generate a light path according to all the converged coded light paths.

[0055] In step S5 , the optical path is divided into eight sub-optical paths corresponding to different computing tasks according to wavelength and polarization direction using the optical splitting module 50 .

[0056] Step S6: using the detection module 60 to generate calculation results corresponding to each sub-light path.

[0057] Figure 4 Schematic diagram of calculation results of a handwritten digital image in an embodiment of the present invention.

[0058] like Figure 4 As shown, (a) is a task of performing a binary classification of digital odd and even on a handwritten digital image. (a1) is the handwritten digital number 2 input to the optical diffraction neural network system 100, which is encoded according to polarized light with a wavelength λ and a polarization direction along the x-axis. (a2) is the corresponding calculation result, including two light spots on the left and right. The left light spot indicates that it is an odd number, and the right light spot indicates that it is an even number. The light spot with the highest light intensity is selected as the final calculation result. (b) is a task of performing a binary classification of digital prime and composite on a handwritten digital image. (b1) is the handwritten digital number 5 input to the optical diffraction neural network system 100, which is encoded according to polarized light with a wavelength λ and a polarization direction along the y-axis. (b2) is the corresponding calculation result, including three light spots c, d, and e, representing prime numbers, composite numbers, and 0, respectively. The light spot with the highest light intensity is selected as the final calculation result. In this embodiment, the point with the highest light intensity is a white light spot. It can be seen that the optical diffraction neural network system 100 obtains accurate calculation results based on the input handwritten digital image and the corresponding classification task.

[0059] Functions and Effects of the Embodiments

[0060] According to the optical diffraction neural network system for simultaneous wavelength and polarization multiplexing involved in this embodiment, multiple polarized lights of different wavelengths and polarization directions are generated by the light source module and the polarized parallel light generation module. The encoding module performs polarization encoding based on the corresponding computing task in combination with the input image and polarized light to obtain an encoded light path. The diffraction module including the phase modulation layer and the linear polarization array layer performs corresponding calculations on each encoded light path to obtain an optical path. The optical path is re-divided into sub-optical paths corresponding to each computing task by the splitting module and the detection module, and detection is performed to obtain the computing results. In short, this method can simultaneously multiplex wavelength and polarization to process multiple computing tasks in parallel and obtain accurate computing results.

[0061] Those skilled in the art will appreciate that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An optical diffraction neural network system that simultaneously multiplexes wavelength and polarization, used to obtain corresponding calculation results based on N*2 input images corresponding to different calculation tasks, characterized in that: include: A light source module, used to generate N parallel lights of different wavelengths; a polarized parallel light generating module, configured to convert each of the parallel lights into a first polarized light corresponding to a first polarization direction and a second polarized light corresponding to a second polarization direction, and use the first polarized light and the second polarized light as polarized light; An encoding module, comprising N*2 digital modulation boards, configured to perform polarization encoding according to each of the input images and the corresponding polarized light, to obtain a corresponding encoding optical path; a diffraction module, configured to generate a light path according to all the converged coded light paths; A light splitting module, configured to split the light path into N*2 sub-light paths corresponding to the different computing tasks according to the wavelength and polarization direction; The detection module includes a CMOS camera and is used to generate the calculation results corresponding to each of the sub-light paths.

2. The optical diffraction neural network system for simultaneous wavelength and polarization multiplexing according to claim 1, characterized in that: in, The frequencies corresponding to all the wavelengths are in the same band.

3. The optical diffraction neural network system for simultaneous wavelength and polarization multiplexing according to claim 1, characterized in that: in, The first polarization direction is along the x-axis, and the second polarization direction is along the y-axis.

4. The optical diffraction neural network system for simultaneous wavelength and polarization multiplexing according to claim 1, characterized in that: in, The diffraction module includes a first phase modulation layer, a second phase modulation layer, a third phase modulation layer, a linear polarization array layer, a fourth phase modulation layer, a fifth phase modulation layer and a sixth phase modulation layer, which are arranged in sequence. The linear polarization array layer is used to perform different operations on linearly polarized light incident at different angles.

5. The optical diffraction neural network system for simultaneous wavelength and polarization multiplexing according to claim 4, characterized in that: in, Corresponding training data sets are constructed according to the different computing tasks, the optical diffraction neural network system is trained according to the training data sets, and the phase plates of the first phase modulation layer, the second phase modulation layer, the third phase modulation layer, the fourth phase modulation layer, the fifth phase modulation layer and the sixth phase modulation layer, and the linear polarizer array plate of the linear polarization array layer are manufactured according to the training results.

6. The optical diffraction neural network system for simultaneous wavelength and polarization multiplexing according to claim 5, characterized in that: in, The expression of the loss function used in the training is: Where N is the total number of channels, is the two-dimensional light intensity distribution of the optical network output surface on the i-th channel, is the two-dimensional distribution of the target true light intensity actually corresponding to the i-th channel during training.

7. The optical diffraction neural network system for simultaneous wavelength and polarization multiplexing according to claim 1, characterized in that: in, The light splitting module includes a filter and an analyzer. The optical path is divided into N*2 sub-optical paths via the filter and the analyzer in sequence.

Images