A multi-layer all-optical high-dimensional neural network design method and system and application thereof

Abstract

A kind of multi-layer all-optical high-dimensional neural network design method and system and its application, system architecture includes input layer, linear calculation layer, non-linear calculation, output layer.Input layer uses optical amplifier, microcavity, waveguide, wavelength division multiplexer and other devices to realize multi-wavelength optical signal input.Linear layer is constituted by multiple programmable Mach-Zehnder interferometer (MZI) network, the network is modulated by changing the refractive index of waveguide Phase, to complete the calculation of matrix-vector multiplication.Nonlinear layer simulates common activation function (such as ReLU) by cascading MZI network.Output layer separates the calculation result by wavelength demultiplexer, and detects the output signal using photodetector.The application reduces the material and optical power dependence faced by the generation of nonlinear activation function, breaks through the bandwidth limitation and matching precision bottleneck of traditional all-optical neural network, and the system has higher integration and adaptability.

Description

Technical Field

[0001] This invention belongs to the field of photonic computing technology, specifically relating to a multi-layer all-optical high-dimensional neural network design method and system and its application, which is particularly suitable for realizing high-precision linear matrix operations and nonlinear activation function mapping in broadband, multi-polarization, and high-parallelism optical computing architectures. Background Technology

[0002] In recent years, with the continuous expansion of deep learning models, traditional electronic computing architectures have faced severe challenges in terms of computing power, power consumption, and latency. Photonic computing, with its advantages of high bandwidth, low power consumption, and strong parallel processing capabilities, has become a highly promising alternative in the post-Moore's Law era. In optical neural networks, the implementation of linear operations (such as matrix-vector multiplication) and nonlinear activation functions are two core issues.

[0003] Currently, existing research has proposed optical linear computation units based on Mach-Zehnder interferometer (MZI) networks, which can efficiently perform matrix operations. However, the optical implementation of nonlinear activation functions still faces many technical bottlenecks. In existing technologies, such as patent CN111860822B, a method for generating nonlinear activation functions based on the coherence of reference and signal light has been proposed. This scheme modulates the phase of the signal light to be processed and the reference light, inputs them into an optical interferometer module, and uses a cascaded MZI or a single-stage beam splitter to achieve nonlinear coupling operations in the optical domain, thereby outputting an activation signal. Although this method supports a certain degree of reconfigurability, it still has the following shortcomings:

[0004] Limited operating bandwidth: This scheme has high requirements for the coherence and wavelength stability of the light source, making it difficult to maintain consistent activation performance over a wide wavelength range (such as 1500nm-1580nm), which limits its application in wavelength division multiplexing (WDM) systems.

[0005] Insufficient polarization multiplexing capability: The system is quite sensitive to the polarization state of the input light, making it difficult to efficiently support multidimensional parallel computing architectures such as polarization multiplexing (PDM), which reduces the system's parallel processing capability.

[0006] The activation function has low accuracy: the generated activation curves deviate significantly from functions such as ReLU and Sigmoid, which are widely used in current neural networks, making it difficult to meet the requirements of high-precision training and inference.

[0007] In summary, existing technologies still have significant limitations in achieving broadband, polarization-insensitive, and high-precision reconfigurable nonlinear activation functions in all-optical neural networks. Therefore, there is an urgent need to propose a novel multilayer all-optical neural network architecture that can operate stably over a wide wavelength range, support multidimensional multiplexing and parallel computing, and possess high-precision nonlinear activation function fitting capabilities. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a design method and system for an all-optical neural network, which can operate efficiently over a wide wavelength range, support multidimensional multiplexing and parallel computing, achieve high-precision reconfigurable nonlinear activation, and possess real-time feedback calibration capabilities. This invention also presents a multilayer all-optical high-dimensional neural network design method and device, as well as its applications.

[0009] The technical solution of the present invention is as follows:

[0010] This invention provides a method for designing multilayer all-optical high-dimensional neural networks, the method comprising the following steps:

[0011] Step S1: Input code

[0012] Multiple sets of coherent optical signals covering the target's operating wavelength are generated using a broadband coherent light source (such as an optical frequency comb or a supercontinuum light source). The optical signals of each wavelength are combined by wavelength division multiplexing and then separated into at least two polarization branches (TE mode and TM mode) by polarization multiplexing. The input data is encoded onto the optical carriers of each wavelength and polarization state by an electro-optic modulator to obtain the encoded parallel optical signals.

[0013] Step S2: Linear matrix operations

[0014] The encoded parallel optical signal is input into a linear computing network (which can employ either the Clements or Reck architecture) consisting of multiple programmable Mach-Zehnder interferometers (MZIs). The phase of each MZI is adjusted through thermo-optical or carrier injection effects, enabling the linear computing network to perform a preset matrix-vector multiplication operation on the input optical signal and output a linearly transformed optical signal.

[0015] Step S3: High-precision nonlinear activation

[0016] The linearly transformed optical signal is input to a nonlinear mapping unit (N≥4, preferably N≥10) composed of N cascaded MZI stages, where each MZI stage contains a programmable phase shifter. By optimizing the phase parameters of each stage of the phase shifter, the relationship between the input optical power and the output optical power of the nonlinear mapping unit is made to approximate a preset neural network activation function curve with a relative error of less than 5%. The activation function is at least one of ReLU, Sigmoid, Tanh, or GELU.

[0017] The specific method for optimizing the phase parameters of each stage of the phase shifter is as follows: K sampling points (K≥50) are selected at equal intervals within the working range of the input optical power; for each candidate phase combination, the mean square error between the output optical power of the cascaded MZI corresponding to each sampling point and the target activation function value is calculated; the phase combination is iteratively updated using gradient descent, genetic algorithm, or extreme value search algorithm until the mean square error is less than [a certain value]. The phase combination that minimizes the mean square error is recorded as the fixed configuration of the nonlinear mapping unit.

[0018] Step S4: Real-time feedback calibration involves monitoring the output optical power in real time at the output of the nonlinear mapping unit and comparing the monitored value with the theoretical value of the preset activation function at that input optical power. When the relative deviation exceeds a preset threshold (e.g., 5%), the phase parameters of each stage of the phase shifter are automatically adjusted to restore the output optical power to the theoretical value. This calibration can be performed periodically or triggered when a temperature sensor detects a chip temperature change exceeding 0.5°C.

[0019] Step S5: Output detection separates the optical signal after nonlinear activation into output channels by wavelength demultiplexing and polarization demultiplexing in sequence. Then, the optical signal is converted into an electrical signal by a photodetector array, which serves as the final output of the neural network.

[0020] The present invention also provides a multilayer all-optical high-dimensional neural network device for implementing the above method, the device comprising:

[0021] The input layer includes: a broadband multi-wavelength light source for generating coherent optical signals covering the target operating wavelength; a wavelength division multiplexer, whose input is connected to the optical output of the multi-wavelength light source for combining multi-wavelength optical signals; a polarization multiplexer, whose input is connected to the output of the wavelength division multiplexer for separating the combined optical signal into at least two polarization branches; and an electro-optic modulator array, respectively disposed on each polarization branch and each wavelength channel, for encoding external input data onto an optical carrier.

[0022] A linear computing layer, whose optical input end is connected to the optical output end of the input layer, is composed of an MZI network consisting of multiple programmable Mach-Zehnder interferometers (MZIs) arranged according to the Clements architecture or Reck architecture, and is used to perform matrix-vector multiplication operations on the input optical signal.

[0023] A nonlinear computing layer, whose optical input is connected to the optical output of the linear computing layer, comprises at least one cascaded MZI nonlinear mapping unit. Each nonlinear mapping unit is composed of N cascaded MZIs (N≥10), and each MZI contains a programmable phase shifter (thermo-optic phase shifter or carrier injection type phase shifter). The nonlinear mapping unit is used to perform nonlinear activation function mapping on the input optical signal.

[0024] The output layer has its optical input end connected to the optical output end of the nonlinear computing layer. The output layer includes: a wavelength demultiplexer (such as a cascaded microring resonator array) for separating multi-wavelength signals; a polarization demultiplexer for separating signals with different polarization states; and a photodetector array for converting the separated optical signals of each channel into electrical signals.

[0025] A real-time feedback calibration system includes: a miniature optical power monitor (such as a Ge-on-Si PIN photodiode) located at the output of the nonlinear computing layer for real-time detection of the output optical power of each channel; a control circuit (FPGA or microcontroller) whose input is electrically connected to the output of the optical power monitor and whose output is electrically connected to the drive terminals of each phase shifter in the nonlinear computing layer; the control circuit is used to compare the monitored value with a preset activation function target value and automatically adjust the drive parameters of the phase shifters according to the deviation.

[0026] In the above device, both the linear computing layer and the nonlinear computing layer employ polarization-independent broadband directional couplers. These directional couplers exhibit a coupling ratio of 50:50±2% for both TE and TM modes within the wavelength range of 1500nm to 1580nm, with an insertion loss of less than 0.2dB.

[0027] The device supports parallel computation of at least two of the following multiplexing methods: wavelength multiplexing, polarization multiplexing, and mode multiplexing. The number of parallel channels M ≥ 32, and the MZI network size of the linear computation layer is M × M.

[0028] Based on the above-mentioned device, the present invention also provides an online reconstruction method for nonlinear activation functions: a first set of phase parameters is loaded into the cascaded MZI phase shifters in the nonlinear computation layer by a control circuit to realize a first activation function; when it is necessary to switch to a second activation function, the control circuit reads the second set of phase parameters (obtained through offline training and pre-stored in a lookup table) from local memory or external input, and writes them sequentially into each stage of the phase shifters; the real-time feedback calibration system automatically performs a calibration after switching, measures the deviation between the actual output and the target value of the second activation function, and if the deviation exceeds a threshold, performs fine-tuning until the accuracy requirements are met.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] Polarization insensitivity: Through innovative polarization-insensitive optical device design, the system can consistently process signals with different polarization states, reducing the number of redundant optical networks required due to polarization issues and improving the chip's effective parallel processing capability.

[0031] Wavelength insensitivity: Through innovatively designed wavelength-insensitive optical devices, the system can process wavelength division multiplexing signals in the range of 1500nm to 1580nm, reducing the problem of low multiplexing efficiency caused by narrow wavelength range and improving the chip's broadband processing capability.

[0032] High precision: The number of cascaded MZIs determines the fitting accuracy of the nonlinear function. Taking the ReLU function as an example, when the input light power is less than 1mW and the initial power is zero, high-precision fitting can be achieved through 10 levels of phase modulation. Figure 5 As shown. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall architecture of the multilayer all-optical high-dimensional neural network device according to Embodiment 1 of the present invention.

[0034] Figure 2 This is a detailed diagram of the architecture of Embodiment 1 of the present invention.

[0035] Figure 3 This is a schematic diagram of the cascaded MZI structure in the nonlinear computing layer of this invention.

[0036] Figure 4 This is a flowchart of the algorithm used in this invention to determine the optimal phase combination of cascaded MZIs.

[0037] Figure 5 The figure shows the fitting result of the ReLU function implemented using 10-level MZI in Embodiment 1 of the present invention.

[0038] Figure 6 The diagram shows the structure and simulation results of the polarization-independent broadband directional coupler in this invention; where (a) is the structural diagram, and (b) and (c) are the simulated transmittance curves for TE polarization and TM polarization, respectively. Detailed Implementation

[0039] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1: A 32-channel all-optical high-dimensional neural network processor based on 4-wavelength × 2 polarization multiplexing

[0041] Please see Figure 1 , Figure 1 The figure shows a schematic diagram of the overall architecture of the multilayer all-optical high-dimensional neural network device according to Embodiment 1 of the present invention. As shown in the figure, the device includes: an input layer 100, a linear calculation layer 200, a nonlinear calculation layer 300, an output layer 400, and a real-time feedback calibration system 500. Figure 2 The architecture details are shown. Figure 3 The detailed structure of the cascaded MZI units in the nonlinear computing layer is shown.

[0042] Input Layer 100: This embodiment employs an on-chip multi-wavelength light source, including components such as optical amplifiers, microcavities, and waveguides. A wavelength division multiplexer combines optical signals of different wavelengths. It not only supports multi-wavelength parallel computing but also exhibits polarization insensitivity, making it suitable for input signals with different polarizations. To encode the input signals, the system uses an optical modulator, which can employ thermo-optical modulation, electro-optical modulation, liquid crystal modulation, or mechanical modulation methods depending on specific requirements. The optical modulator enables adjustable signal strength encoding, allowing input optical signals of different wavelengths to carry rich information into the network computing layer.

[0043] Linear Computation Layer 200: The linear layer consists of multiple programmable MZI networks. These MZI networks achieve phase modulation by changing the refractive index of the waveguide, thereby completing matrix-vector multiplication calculations. By programming and controlling the phase of the MZI networks, the system can dynamically adjust the parameters of matrix operations to meet the needs of different computational tasks.

[0044] Nonlinear computation layer 300: Common activation functions (such as ReLU) are simulated through a cascaded MZI network. The synergistic effect of multiple phase modulators enables the system to accurately approximate common nonlinear activation functions and ensures stable operation of the network under wide wavelength and multi-polarization conditions. Compared with nonlinear activation layers composed of traditional interferometers or microring resonators, the design of this invention has higher nonlinear matching accuracy.

[0045] Output Layer 400: The output layer separates the calculation results using a wavelength demultiplexer and detects the output signal using a photodetector. The system also incorporates a polarization beam splitter to support parallel output of multi-polarization signals. The detection results are fed back to the control module for training the neural network or outputting the calculation results.

[0046] like Figure 2 As shown, this implementation example demonstrates a multilayer all-optical high-dimensional neural network device that supports 4 wavelength channels, 2 polarization states (TE / TM), and a total of 32 parallel computing channels.

[0047] 1. Input layer

[0048] An on-chip integrated optical frequency comb is used as an on-chip multi-wavelength light source, with each wavelength corresponding to a signal channel. The optical frequency comb system mainly consists of a highly stable pump source and a high-Q microcavity, capable of generating high-precision multi-wavelength optical signals. These wavelengths can be tuned within narrow intervals (e.g., 60 GHz) to meet the requirements of high-density wavelength division multiplexing. After these multi-wavelength signals enter the system, they are split onto different channels by polarization multiplexers and wavelength multiplexers. In the data modulation stage, the electrical signal is intensity-encoded for the optical signal of each channel by a modulator.

[0049] 2. Design of Optical Computing Core (Neural Network)

[0050] Linear computational layer: MZI (Mach-Zehnder interferometer) network

[0051] The linear layer in the optical computing core is composed of a programmable MZI network. The phase of the input optical signal is adjusted by changing the waveguide refractive index through thermo-optical effects, enabling matrix-vector multiplication. This network is used to execute the linear operations in the deep neural network. The design and simulation results of the polarization-independent broadband directional coupler constituting the MZI are as follows: Figure 3 As shown, the device exhibits polarization insensitivity over a wide wavelength range and is capable of 50:50 beam splitting in TE and TM modes.

[0052] Nonlinear computation layer: Implementation of nonlinear activation function

[0053] The signal light and probe light are input into a cascaded MZI network, such as an N-stage MZI network. Figure 3 As shown. The transmission matrix of a single MZI matrix is:

[0054]

[0055] The transmission matrix for cascading N MZIs is:

[0056]

[0057] By optimizing the phase of each matrix The algorithm aims to find the optimal phase combination so that the relationship between input and output optical power approximates the ReLU function. The specific algorithm for finding the optimal phase combination is shown in the block diagram. Figure 4 ).

[0058] To realize a polarization-insensitive beam splitter under broadband conditions, a designed 2x2 directional coupler and its performance are as follows: Figure 6 As shown.

[0059] 3. Output Layer Design

[0060] The wavelength demultiplexer in the output layer is based on a microring resonator design. This device can separate signals of multiple wavelengths from the same waveguide and transmit them to their respective output ports. Then, a polarization demultiplexer transmits the TE and TM mode signals to different detectors. This allows the system to utilize data streams with different polarization states for parallel computation. Finally, each output channel is equipped with a photodetector to convert the optical signal into an electrical signal. The detection results can be fed back to the control module for model training or to output the final result.

[0061] Taking a forward inference as an example, the workflow is as follows:

[0062] ① The external electrical signal (input vector) is encoded onto optical carriers of various wavelengths and polarizations by a modulator.

[0063] ② The input optical signal enters the linear computing layer 200, and the MZI mesh performs matrix multiplication operations.

[0064] ③ The linear calculation results enter the nonlinear calculation layer 300, and the activation function is mapped through cascaded MZI units.

[0065] ④ If a deeper network is needed, the nonlinear output can be fed back to the linear computation layer (through an optical switch or waveguide loop) to achieve a multi-layer network.

[0066] ⑤ The final result is converted into an electrical signal output by the output layer 400.

[0067] ⑥ The calibration system 500 periodically or automatically calibrates the transmission characteristics of the nonlinear unit when the temperature changes.

[0068] This invention systematically solves the technical challenges of existing all-optical neural networks in terms of bandwidth, polarization adaptability, accuracy, and stability by using cascaded MZI nonlinear mapping units, polarization-independent broadband directional couplers, and a real-time feedback calibration system, providing a complete technical solution for large-scale, high-reliability photonic computing chips.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art can make various modifications, equivalent substitutions, and improvements without departing from the principles of the present invention, and these should also be considered within the scope of protection of the present invention.

Claims

1. A method for designing a multi-layer all-optical high-dimensional neural network, characterized in that, Includes the following steps: Step S1: Input encoding: Use a multi-wavelength light source to generate multiple sets of coherent optical signals covering the target working wavelength. After the optical signals of each wavelength are combined by wavelength division multiplexing, they are separated into at least two polarization branches by polarization multiplexing. The input data is encoded onto the optical carriers of each wavelength and each polarization state by an electro-optic modulator to obtain the encoded parallel optical signal. Step S2: Linear computation: The encoded parallel optical signal is input into a linear computation network composed of multiple programmable Mach-Zehnder interferometers (MZIs). The phase of each MZI is adjusted through thermo-optical or carrier injection effects to realize matrix-vector multiplication and output the linearly transformed optical signal. Step S3: Nonlinear activation: The optical signal after the transformation is input into a nonlinear mapping unit composed of cascaded MZIs. The cascaded MZIs are composed of N MZIs connected in series, and each MZI contains a phase shifter. By optimizing the phase parameters of each phase shifter, the relationship between the input optical power and the output optical power of the nonlinear mapping unit is made to approximate the preset activation function curve with a relative error of less than 5%. Step S4: Output detection: The optical signal after nonlinear activation is sequentially separated into output channels by wavelength demultiplexing and polarization demultiplexing, and then converted into electrical signals by a photodetector array as the final output of the neural network.

2. The multi-layer all-optical high-dimensional neural network design method of claim 1, wherein, In step S3, the number of cascaded MZI stages N≥4, preferably N≥10; the activation function is selected from one or more of ReLU, Sigmoid, Tanh or GELU. 3.The multi-layer all-optical high-dimensional neural network design method of claim 1, wherein, In step S3, the method for optimizing the phase parameters of each stage of the phase shifter specifically includes: K sampling points are selected at equal intervals within the working range of the input optical power, where K ≥ 50; For each candidate phase combination, calculate the mean square error between the output optical power of the cascaded MZI and the target activation function value at each sampling point; The phase value is iteratively updated using gradient descent, genetic algorithm, or extreme value search algorithm until the mean square error is less than a preset threshold. The phase combination that minimizes the mean square error is recorded as the fixed configuration of this nonlinear mapping unit.

4. The multilayer all-optical high-dimensional neural network design method according to claim 1, characterized in that, It also includes a real-time feedback calibration step: the output optical power is monitored in real time at the output end of the nonlinear mapping unit, and the monitored value is compared with the target value of the preset activation function under the input optical power; when the relative deviation exceeds the preset threshold, the phase parameters of the phase shifters at each stage are automatically adjusted so that the output optical power is restored to the target value.

5. A multilayer all-optical high-dimensional neural network system for implementing the method of any one of claims 1 to 4, characterized in that, include: Input layer: It contains: - A broadband multi-wavelength light source used to generate coherent optical signals covering the target's operating wavelength; - A wavelength division multiplexer, whose input is connected to the optical output of the multi-wavelength light source, is used to combine multi-wavelength optical signals; - A polarization multiplexer, the input of which is connected to the output of the wavelength division multiplexer, is used to separate the combined optical signal into at least two polarization branches; - An electro-optic modulator array, respectively set on each polarization state branch and each wavelength channel, is used to encode external input data onto an optical carrier; Linear computing layer: Its optical input end is connected to the optical output end of the input layer, and it is composed of multiple programmable Mach-Zehnder interferometer (MZI) networks, used to perform matrix-vector multiplication operations on the input optical signal; Nonlinear computing layer: Its optical input end is connected to the optical output end of the linear computing layer, and it contains at least one cascaded MZI nonlinear mapping unit. Each nonlinear mapping unit is composed of N MZIs cascaded in sequence, and each MZI contains a phase shifter. The nonlinear mapping unit is used to perform nonlinear activation function mapping on the input optical signal. Output layer: Its optical input terminal is connected to the optical output terminal of the nonlinear computing layer, and the output layer includes: - Wavelength demultiplexer, used to separate multi-wavelength signals; - Polarization demultiplexer, used to separate signals with different polarization states; - A photodetector array is used to convert the separated optical signals from each channel into electrical signals.

6. The multilayer all-optical high-dimensional neural network system according to claim 5, characterized in that, It also includes a real-time feedback calibration module, which contains: A miniature optical power monitor is installed at the output end of the nonlinear computing layer to detect the output optical power of each channel in real time. The control circuit has its input terminal electrically connected to the output terminal of the optical power monitor, and its output terminal electrically connected to the drive terminal of each phase shifter in the nonlinear calculation layer. The control circuit is used to compare the monitored value with the preset activation function target value, and automatically adjust the drive parameters of the phase shifter according to the deviation.

7. The multilayer all-optical high-dimensional neural network system according to claim 5, characterized in that, The MZI in the linear and nonlinear computing layers uses a polarization-independent broadband directional coupler, and the coupling ratio of the TE mode and TM mode in the wavelength range of 1500nm to 1580nm is 50:50±2%.

8. The multilayer all-optical high-dimensional neural network system according to claim 5, characterized in that, In the nonlinear computing layer, each nonlinear mapping unit contains N≥4 levels of MZI, preferably N≥10; the phase shifter is a thermo-optical phase shifter or a carrier injection type phase shifter.

9. The multilayer all-optical high-dimensional neural network system according to claim 5, characterized in that, The multi-wavelength light source is an optical frequency comb light source or a supercontinuum light source, the wavelength division multiplexer is an arrayed waveguide grating or a micro-ring resonator array, and the electro-optic modulator is selected from one or more of thermo-optic modulators, electro-optic modulators, liquid crystal modulators, or mechanical modulators.

10. The multilayer all-optical high-dimensional neural network system according to claim 5, characterized in that, The programmable MZI network in the linear computing layer adopts the Clements architecture or the Reck architecture, with a network size of M×M, where M is the number of parallel channels, and M≥32.

11. A method for online reconstruction of nonlinear activation functions in a multilayer all-optical high-dimensional neural network, characterized in that, The system based on any one of claims 5 to 10 includes the following steps: The first set of phase parameters is loaded into the cascaded MZI phase shifters in the nonlinear computing layer by the control circuit to realize the first activation function. When it is necessary to switch to the second activation function, the control circuit reads the second set of phase parameters from the local memory or external input and writes them into each stage of the phase shifter in sequence; The real-time feedback calibration system automatically performs a calibration after switching, measuring the deviation between the actual output and the target value of the second activation function. If the deviation exceeds the threshold, it performs fine-tuning until the accuracy requirements are met.

12. The online reconstruction method for nonlinear activation functions according to claim 11, characterized in that: The first set of phase parameters and the second set of phase parameters are obtained through offline training. Each set of parameters corresponds to an activation function and is pre-stored in the lookup table of the control circuit.

Images