CN-122021760-A - Multi-wavelength channel diffraction neural network system based on threshold screening and processing method

CN122021760ACN 122021760 ACN122021760 ACN 122021760ACN-122021760-A

Abstract

The invention discloses a multi-wavelength channel diffraction neural network system based on threshold screening and a processing method thereof, belonging to the technical field of artificial intelligence and optical computing. The optical diffraction module is used for receiving the task channel light fields and realizing diffraction modulation and threshold screening of the task channel light fields by changing super-surface structures on different modulation layers, the task channel light fields share the same set of diffraction modulation parameters, namely parameters such as phase, mask generation and threshold, the threshold screening is used for comparing mask generation parameters with threshold parameters, performing binarization selection on the phase parameters according to comparison results to screen out target phase parameters, and the output detection module is used for detecting output light intensity. The invention expands the task channel and improves the system integration level.

Inventors

LI XIANGPING
HE MINYI
XU JILIAN
MOU ZHEN

Assignees

暨南大学

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (10)

1. The multi-wavelength channel diffraction neural network system based on threshold screening is characterized by comprising a light source module, a coding module, an optical diffraction module and an output detection module which are connected in sequence; The light source module is used for generating N light beams with different working wavelengths; the coding module is used for respectively coding task information corresponding to M different tasks to be processed onto the N beams with different working wavelengths to form a multi-channel input light field with the working wavelengths corresponding to the tasks to be processed one by one, wherein the multi-channel input light field comprises N task channel light fields; The optical diffraction module comprises a first modulation layer, a third modulation layer and a fourth modulation layer which are sequentially arranged along an optical axis, wherein the surface of each modulation layer is provided with a super-surface structure, the optical diffraction module is used for receiving the N task channel light fields and realizing diffraction modulation and threshold screening of each task channel light field by changing the super-surface structures on different modulation layers, N, M, L are positive integers; On each modulation layer, the optical fields of all task channels share the same set of diffraction modulation parameters, wherein the diffraction modulation parameters comprise phase parameters, mask generation parameters and threshold parameters; the threshold screening is to compare the mask generation parameter with the threshold parameter, perform binarization selection on the phase parameter according to the comparison result, and finally screen out the target phase parameter; And the output detection module is used for detecting the output light intensity of the light field of each task channel modulated by the optical diffraction module and obtaining the processing result corresponding to each task information.
2. The threshold-screening-based multi-wavelength channel diffractive neural network system of claim 1, wherein said optical diffraction module further comprises a free propagation space between two adjacent modulation layers; each modulation layer is provided with a modulation plane on the light incident side; In the optical diffraction module, a first super-surface structure is arranged on a modulation plane of at least one modulation layer, and the first super-surface structure comprises a plurality of phase modulation units positioned on the modulation plane; On each of the modulation layers except the L-th modulation layer, continuous and independent spatial phase modulation is applied to incident light by rotating each of the phase modulation units on the modulation plane.
3. The multi-wavelength channel diffraction neural network system based on threshold value screening according to claim 2, wherein in the optical diffraction module, the first super-surface structure is arranged on each modulation layer except the L-th modulation layer, and the L-1 modulation layers perform optical field regulation by adopting continuous phase modulation without threshold value screening operation; the L-th modulation layer is provided with a second super-surface structure, the second super-surface structure comprises a plurality of threshold screening units, and the threshold screening units are used for screening out phase modulation units corresponding to the target phase parameters.
4. A multi-wavelength channel diffraction neural network system based on threshold screening according to claim 3, wherein each of said threshold screening units is a hole structure provided on the surface of said L-th modulation layer; Each phase modulation unit is a bump structure arranged on the surface of the corresponding modulation layer.
5. The threshold-screening-based multi-wavelength channel diffraction neural network system according to claim 1, wherein the light source module comprises N single-color light sources which are independently arranged and are used for generating N light beams with different discrete working wavelengths; The coding module comprises M wavelength channel coding units for coding the first wavelength channel Task information of each task to be processed is loaded to the working wavelength as follows To form and carry the first Coding light beam of each task information to realize working wavelength And the first The tasks to be processed are in one-to-one correspondence, wherein, The output detection module comprises a plurality of detection units.
6. An optical multiplexing processing method, which is applied to a multi-wavelength channel diffraction neural network system based on threshold value screening according to any one of claims 1 to 5, and comprises the following steps: S1, pre-constructing a trainable model comprising L modulation layers, wherein each modulation layer is provided with a geometric phase super surface, each super surface in the trainable model is regarded as a trainable phase layer, and the phase distribution of each phase layer is a parameter to be optimized; s2, the coding module independently codes task information of a plurality of tasks to be processed on light beams with corresponding working wavelengths through M wavelength channel coding units to form a plurality of independent coded light beams; The method comprises the steps of S3, receiving and processing a plurality of coded light beams from a coding module by an optical diffraction module, carrying out optical propagation on the coded light beams with different wavelengths in a free propagation space at the same time, carrying out diffraction modulation through L modulation layers in sequence, and completing calculation of a plurality of tasks in the optical propagation process, wherein the first modulation layer to the L-1 modulation layer apply continuous and independent phase regulation on the incident coded light beams through rotating a phase modulation unit without threshold screening operation; S4, after the coded light beams are sequentially transmitted through L modulation layers and the interlayer free transmission spaces, the coded light beams reach the detection plane of the output detection module through the last free transmission space; Step S5, obtaining task losses of each wavelength by calculating the mean square error between the output light intensity distribution of each working wavelength and the corresponding target light intensity distribution, and carrying out weighted summation on the losses of all task channels to obtain a composite total loss function; Step S6, calculating gradients of phase parameters, mask generation parameters and threshold parameters of each modulation layer under all working wavelengths by using an error back propagation algorithm in the training process, and iteratively updating the parameters by using a gradient descent algorithm; And S7, repeatedly executing the steps S1 to S6 to enable the network learning to convert the light field of each task channel into the expected output light intensity distribution until the network optimization training reaches convergence or reaches the preset iteration times, and completing the training.
7. The optical multiplexing processing method according to claim 6, wherein in the step S3, each of the phase modulation units uses its corresponding mask generation parameter in the L-th modulation layer, and compares the mask generation parameter with the corresponding threshold parameter after mapping by the normalization function to obtain a mask having a value of 0 or 1; When the mask value is 1, the corresponding phase parameter is reserved and participates in the phase modulation in the training process; When the mask value is 0, the corresponding phase parameter is suppressed in the training process.
8. The optical multitasking method of claim 7, wherein said normalization function is a Sigmoid function for mapping said mask generation parameter to a continuous value range between 0 and 1 and comparing the mapped parameter with a corresponding threshold parameter by the normalization function to determine the mask value; When the value of the result mapped by the normalization function is larger than or equal to the threshold parameter, the mask at the corresponding position is set to be 1, and when the value of the result mapped by the normalization function is smaller than the threshold parameter, the mask at the corresponding position is set to be 0, so that binarization selection is realized.
9. The optical multitasking method according to claim 6, wherein in the step S4, the detection plane of the output detection module is divided into a physical sub-areas, each physical sub-area corresponds to a preset target class, and has B detection units for detecting output light intensities of N different wavelength channels, respectively, for each wavelength channel, by determining a maximum value of the output light intensities in the a physical sub-areas, and outputting a target class corresponding to the physical sub-area where the maximum value is located as a prediction result of the wavelength channel, wherein A, B is a positive integer.
10. The optical multiplexing processing method according to claim 6, wherein in said step S5, said composite total loss function Loss of task channel corresponding to each working wavelength According to the weight coefficient And (3) carrying out weighted summation to obtain the formula: ; Wherein, the As a total number of wavelength channels, Represent the first The individual operating wavelengths correspond to the loss of the task channel, Loss of channels for the task The corresponding weight coefficient.

Description

Multi-wavelength channel diffraction neural network system based on threshold screening and processing method Technical Field The invention belongs to the technical field of artificial intelligence and optical computing, and particularly relates to a multi-wavelength channel diffraction neural network system based on threshold screening and a processing method thereof. Background Optical diffraction neural networks are a particular computational model, the essence of which is to use optical elements to model and implement neural network reasoning functions. From the physical mechanism level, the calculation model relies on physical processes such as interference, diffraction and the like generated when light waves propagate in a medium, so that parallel matrix operation is realized. It is based on such an operation mode that the optical diffraction neural network exhibits advantageous characteristics of high speed, low power consumption, natural parallelism, and the like. Compared with the traditional artificial neural network based on electronic hardware, the optical diffraction neural network regulates and controls the light field (such as phase modulation) through pixel units on different modulation layers, and specifically, the optical diffraction neural network utilizes complex amplitude weight distribution formed in the interlayer diffraction propagation process to construct an optical physical computing network equivalent to a fully-connected neural network, and the architecture breaks through the bottleneck of traditional electronic computing in energy efficiency optimization and parallel processing capacity. At present, optical diffraction neural networks have shown good performance in single-task processing scenarios such as image classification, target recognition, and the like. As in the prior art, the patent application with publication number CN117040623a discloses an angular momentum classification and identification system based on a diffraction neural network, which constructs an optical diffraction neural network. Meanwhile, the super surface can flexibly regulate and control the amplitude, phase, polarization and other multidimensional parameters of incident light waves by means of the characteristic of the sub-wavelength structure of the super surface and through designing the geometric shape and the spatial arrangement of the nano structure units, and realizes the complex light field regulation and control function in a compact plane shape, thereby providing key support for the miniaturization and the high integration level of the optical diffraction neural network. Although optical diffraction neural networks have advanced in single-task processing, practical application scenarios are more complex, and often require their ability to simultaneously cope with multiple tasks. Most current computing systems are limited by a fixed physical architecture, which lacks flexible dynamic reconstruction or multiple parallel computing capabilities when faced with multitasking. To solve the above problems, current related studies are mainly explored from the following three directions: Firstly, part of computing systems adopt a mechanical reconfigurable scheme, namely task switching is realized by physically replacing optical elements, but the scheme mainly relies on mechanical reconfigurable operation, so that the requirements on alignment accuracy are high, the operation flow is complex, and the system structure becomes more complex due to reconfiguration replacement; secondly, adopting a polarization multiplexing scheme, wherein different tasks are encoded to different polarization state channels so as to realize parallel computation in the same optical system, but the polarization multiplexing scheme is limited by the polarization freedom degree of light, the number of the multiplexing channels is extremely limited, crosstalk exists between the polarization channels, and the expansion capacity of the computing system under the multi-task channels is further limited; Thirdly, the multi-task learning problem of the optical diffraction neural network is solved on the algorithm level, and the catastrophic forgetting is avoided by introducing an elastic weight holding mechanism, so that a single optical diffraction neural network can sequentially execute a plurality of tasks, however, the computing system can only process a single task at any time, retraining and updating are needed once every time one task is added, the training cost can linearly increase along with the number of the tasks, the training complexity index is increased due to the increase of the number of the tasks, and the channel expansion capability is still limited. Thus, the bottleneck encountered by existing optical multi-tasking systems in channel expansion, system integration and mechanical reconstruction has become a major technical challenge in the art. Disclosure of Invention Aiming at the problems in the related art, the in