All-optical nonlinear neural network system based on linear system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2025-06-23
- Publication Date
- 2026-07-03
AI Technical Summary
Existing optical neural networks suffer from high latency, high energy consumption, and system complexity when performing nonlinear operations. Furthermore, traditional solutions rely on photoelectric conversion or material nonlinear effects, resulting in high cost, poor stability, and limited signal compatibility.
An all-optical nonlinear neural network based on a linear system is adopted. By loading different physical quantities onto the input and output signals, nonlinear calculations are achieved using a ring resonant cavity array and coupling modules, avoiding photoelectric conversion and material nonlinear effects. The weight mapping of the neural network is realized by adjusting the coupling strength and resonant wavelength.
It enables nonlinear computation under low power consumption and low latency conditions, reduces hardware complexity, supports single-mode and multi-mode optical signal processing, adapts to diverse application scenarios, and has anti-electromagnetic interference capabilities, supporting online training and real-time adjustment.
Smart Images
Figure CN120745725B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of optical neural network technology, and relates to an all-optical nonlinear neural network system based on a linear system. Background Technology
[0002] With the widespread application of artificial neural networks in fields such as visual computing and natural language processing, the network size and computing power requirements have continued to surge, leading to a significant increase in energy consumption, hardware costs, and training time. Optical neural networks, with their advantages of ultra-high-speed parallel computing, low energy consumption, and resistance to electromagnetic interference, have emerged as a potential technological route to solve the computing power bottleneck. They utilize the physical properties of light to perform linear operations such as matrix multiplication at the speed of light, significantly improving computational efficiency.
[0003] However, the practical deployment of optical neural networks has long been constrained by the challenges of implementing nonlinear computations. Existing technologies suffer from two main drawbacks: First, relying on optoelectronic hybrid schemes to achieve nonlinear activation functions requires frequent photoelectric-optical conversions, leading to high latency and high energy consumption. Second, all-optical nonlinear schemes typically rely on the inherent optical nonlinear effects of the materials themselves, such as the nonlinear characteristics of silicon-based photonic devices. These schemes have stringent requirements for input optical power and mostly only support coherent optical signal processing. To meet power thresholds, the system is often forced to introduce optical amplifiers, which increases system complexity and cost.
[0004] Despite ongoing research into novel mechanisms of optical nonlinearity, the efficient construction of optical nonlinear computational units remains a core challenge due to obstacles such as high implementation costs, poor system stability, and limited signal compatibility. Therefore, a novel approach is urgently needed that eliminates the need for optical nonlinear materials and avoids photoelectric conversion, thereby fully leveraging the high-efficiency parallel capabilities of optical neural networks. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide an all-optical nonlinear neural network system based on a linear system. This system loads the input and output signals, which are loaded onto the same physical quantity (such as the electric field or power of an optical signal) in traditional optical neural networks, onto different physical quantities, thus realizing the function of a nonlinear neural network using a linear system.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] An all-optical nonlinear neural network system based on a linear system, comprising the following sequentially connected components:
[0008] The input module is used to apply the input signal of the neural network to the detuning of the resonant wavelength of the ring resonant cavity;
[0009] The neuron module includes a multi-layered ring resonant cavity array for linear transmission of optical signals;
[0010] The coupling module connects the input module and the neuron module, the layers of the neuron module, and the neuron module and the output module. It is used to map the weight coefficients of the neural network by adjusting the coupling strength.
[0011] The output module, connected to the last layer of the neuron module, is used to load the output signal of the neural network onto the output optical power;
[0012] The input and output signals are applied to different physical quantities, enabling the linear optical system to perform nonlinear calculations.
[0013] Furthermore, the input module includes:
[0014] The resonant wavelength modulation device array is directly connected to the ring resonant cavity array and is used to modulate the resonant wavelength;
[0015] A ring resonant cavity array, whose input terminal receives the input signal and whose output terminal is connected to a coupling module;
[0016] The coupling module includes:
[0017] A ring waveguide connects adjacent neuron layers;
[0018] The coupling region array is physically coupled to the ring waveguide and the ring resonant cavity of the adjacent layer, respectively.
[0019] An array of coupling strength control devices is directly connected to a ring waveguide and used to control the coupling strength.
[0020] The neuron module includes:
[0021] A multi-layer ring resonant cavity array, each layer of which is connected to a ring waveguide through a coupling region array;
[0022] The output module includes:
[0023] A ring resonant cavity array, with its input end connected to the terminal layer neurons via a coupling region array;
[0024] The probe waveguide array is coupled to the ring resonant cavity array through the coupling region array;
[0025] The coupling region array is connected to the ring resonant cavity array and the probe waveguide array, respectively.
[0026] Furthermore, the resonant wavelength modulation device is a thermo-optic modulation device or an electro-optic modulation device;
[0027] The ring resonant cavity is in the form of a circular, racetrack-shaped, or irregularly shaped closed waveguide, with a strip-shaped or ridge-shaped cross-section, transmitting single-mode or multi-mode optical signals.
[0028] Furthermore, the coupling strength adjustment device is a thermo-optic modulation device or an electro-optic modulation device;
[0029] The ring waveguide can be circular, racetrack-shaped, or irregularly shaped closed waveguide.
[0030] Furthermore, the coupling region is a coupling structure between a silicon-based waveguide and a ring resonant cavity, including any of the following:
[0031] Coupled between two straight waveguides;
[0032] Single-point or multi-point coupling between straight waveguides and curved waveguides;
[0033] Single-point or multi-point coupling between curved waveguides.
[0034] Furthermore, the input light of the probe waveguide is a single-mode or multi-mode optical signal, and the optical power at the output end is |a q,res | 2 Satisfying the formula:
[0035] |a q,res | 2 =|S q | 2 |a q,probe | 2
[0036]
[0037] Among them, S q For the transfer function κ q To detect the coupling strength between the waveguide and the ring resonator, G o,q For the transfer function of the output module, a q,probe This is the input optical complex amplitude.
[0038] A method for implementing an all-optical nonlinear neural network based on the system includes the following steps:
[0039] Input steps: Input signal x is processed through the input module. q The amount of detuning of the resonant wavelength applied to the ring resonant cavity;
[0040] Weight adjustment steps: The interlayer coupling strength J is adjusted through the coupling module to realize the weight mapping of the neural network;
[0041] Linear transmission steps: Optical signals are linearly transmitted in the neuron module via a ring resonant cavity array;
[0042] Nonlinear output steps: Measure the output optical power |a of the probe waveguide through the output module. q,res | 2 This is the result of nonlinear calculation.
[0043] Furthermore, in the nonlinear output step, the output optical power is derived from coupled-mode theory and satisfies:
[0044]
[0045] Among them, G i,q Pass functions to the input module, κ tot Let denot be the total loss of the resonant cavity, Δ be the detuning of the resonant cavity, and J be the coupling strength.
[0046] The beneficial effects of this invention are as follows:
[0047] (1) This invention achieves nonlinear calculation function in a completely linear optical system by loading the input signal and output signal of the neural network onto different physical quantities, thereby fundamentally avoiding the technical route of traditional optical neural networks that rely on material nonlinear effects or photoelectric conversion, and solving the problem of realizing optical nonlinear calculation.
[0048] (2) The fully linear optical structure is adopted to avoid the introduction of additional devices such as optical amplifiers and electro-optic modulators, eliminate the photoelectric conversion link, greatly reduce system energy consumption and signal delay, and at the same time reduce hardware integration complexity.
[0049] (3) Supports single-mode and multi-mode optical signal processing without the need for a strictly coherent light source; the ring resonator and silicon-based waveguide structure are compatible with existing photonic integration processes and can be directly deployed on general photonic chip platforms, reducing the threshold for industrialization.
[0050] (4) Utilizing the natural parallelism of optical systems, linear operations between neural network layers are completed at the speed of light. At the same time, equivalent nonlinear activation is achieved by decoupling input and output physical quantities, thus balancing high-speed computing and complex task processing capabilities.
[0051] (5) Linear optical elements have strong anti-electromagnetic interference characteristics, and parameters such as coupling strength and resonant wavelength can be dynamically reconstructed through thermo-optic or electro-optic modulation, supporting online training and real-time adjustment of neural network weights, and adapting to diverse application scenarios.
[0052] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:
[0054] Figure 1 This is a schematic diagram illustrating the principle of the present invention;
[0055] Figure 2 This is a schematic diagram of the structure of the present invention;
[0056] Figure 3 This is a flowchart illustrating the signal propagation process in this invention.
[0057] Figure 4 This is a schematic diagram of the network structure for implementing MNIST handwritten digit classification and recognition in this invention;
[0058] Figure 5 This is a performance graph of the present invention when implementing MNIST handwritten digit classification and recognition; Figure 5 (a) shows the test accuracy of the neural network and the number of training iterations. Figure 5 (b) is a confusion matrix diagram of the test accuracy. Detailed Implementation
[0059] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0060] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0061] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0062] In the all-optical nonlinear neural network based on a linear system, input and output signals of different physical quantities per unit are adopted, rather than the traditional method of the same physical quantity per unit.
[0063] Figure 1 This is a schematic diagram of the principle of the present invention, showing the overall architecture of the present invention, including an input module, a coupling module, a neuron module, and an output module. Figure 2 This is a schematic diagram of the structure of the present invention. Its working principle is as follows: The input module consists of a circular silicon-based ring resonator array and a resonant wavelength tuning device array. The resonant wavelength tuning device is realized by fabricating a thermal electrode on the ring resonator. By adjusting the voltage applied across the thermal electrode, the resonant wavelength of the ring resonator can be changed, thereby introducing a series of detuning amounts into the ring resonators of the input module. This detuning amount is the input signal of the network. The input signal is transmitted to the next-layer neuron module through the coupling module. The coupling module consists of a racetrack-shaped ring waveguide, several two-point coupling regions, and several coupling strength tuning devices. The coupling strength tuning device is also realized by fabricating a thermal electrode on the ring waveguide. By adjusting the voltage applied across the thermal electrode, the coupling strength between the upper / lower-layer ring resonators and the ring waveguide of this layer is regulated. The squared modulus of this coupling strength serves as the weight coefficient of the neural network. The neuron module consists of a circular silicon-based ring resonator array, which serves as the carrier for signal transmission to realize the signal transfer between layers. The output module consists of a circular ring resonator array, a coupling region array, and a detection waveguide array. An optical signal is input at one end of the detection waveguide, and the output optical power at the other end of the waveguide is the output signal of this neural network. Thus, it can be seen that the input and output signals of the neural network are respectively loaded on two different physical quantities, namely the detuning amount of the micro-ring resonator and the output optical signal power. Therefore, the function of a nonlinear neural network can be realized using a linear network, and the theoretical proof is as follows.
[0064] The present invention realizes the nonlinear calculation function of the neural network by using the nonlinear relationship between the detuning amount of the ring resonator in the input module and the output optical signal power of the output module. Next, the coupled-mode theory (CMT) will be used to prove the nonlinear relationship between the input and output signals. Suppose there is an n-layer network. According to CMT, the relationships of the optical signals between layers are shown in Equations (1) - (4):
[0065]
[0066] where a i,q represents the complex amplitude of the optical signal in the q-th ring resonator in the input module, is the total loss of the q-th ring resonator in the input module, Δ (p,q) is the detuning amount of the q-th ring resonator of the neuron module in the p-th layer (1 < p < n - 2), x qis the detuning amount input to the q-th ring resonator of the input module, is the coupling strength between the q-th ring resonator of the output module and the l-th ring resonator of the first layer of the neuron module, represents the complex conjugate. a p,q represents the complex amplitude of the optical signal in the q-th ring resonator of the neuron module in the p-th layer (1 < p < n - 2), is the total loss of the q-th ring resonator of the neuron module in the p-th layer, is the coupling strength between the l-th ring resonator of the neuron module in the (p - 1)-th layer and the q-th ring resonator in the p-th layer, is the coupling strength between the q-th ring resonator of the neuron module in the p-th layer and the l-th ring resonator of the neuron module in the (p + 1)-th layer. a o,q represents the complex amplitude of the optical signal in the q-th ring resonator of the input module, is the total loss of the q-th ring resonator of the output module, Δ (o,q) is the detuning amount of the q-th ring resonator of the output module, is the coupling strength between the l-th ring resonator of the neuron module in the (n - 2)-th layer and the q-th ring resonator of the output module, κ q is the coupling strength between the q-th detection waveguide of the output module and the ring resonator, a q,probe and a q,res are the complex amplitudes of the input optical signal and the corresponding output optical signal of this detection waveguide. The above equations can be derived layer by layer starting from the input module to obtain the relationship between the response output a q,res and x q . After derivation, the following steady-state solution can be obtained:
[0067]
[0068] |a q,res | 2 = |S q | 2 |a q,probe | 2 (7)
[0069]
[0070] G in equations (5) and (6) i,q and G o,q represent the transfer process of the detuning amount x q in the input and output modules. Its imaginary part is the detuning amount introduced to the next layer. S q is the output optical signal a q,res on the q-th detection waveguide of the output module q,probeThe ratio is the transfer function of the optical signal in the system. From equation (5), we know that S... q The magnitude of the optical signal is independent of the intensity of each layer, indicating that this system is a linear system. However, the input signal x of the neural network... q Included in G o,q In the middle, with the output optical signal power |a q,probe | 2 The existence of nonlinear relationships gives the network the ability to perform nonlinear calculations.
[0071] Figure 3 This demonstrates the signal propagation process in a neural network, with a detuning quantity x input to the ring resonant cavity of the input module. q , to obtain G i,q After passing through the coupling module, the coupling strength is adjusted by the coupling strength control device to update the weights, and then G is... i,q G is obtained on the ring resonator of the neuronal module in the first layer. 1,q This pattern continues until the output module is reached, where G is obtained. o,q Finally, the intensity of the output optical signal |a| is obtained by probing the waveguide. q,res | 2 As can be seen from equation (5), the output signal |a q,res | 2 With input signal x q The existence of a nonlinear relationship between and proves that nonlinear computation can be achieved on linear systems.
[0072] Figure 4 This paper demonstrates an example of applying the present invention to MNIST handwritten digit classification and recognition. In this example, a two-layer neural network structure with a 20-neuron input layer and a 10-neuron output layer is used. Figure 4 As shown, the 784-pixel handwritten digit image is first convolutionally encoded to obtain 20 pieces of information, which are then used as the detuning factor x. q The signals are sequentially input into the input layer, and finally, the light intensity |a| of the output signal is obtained in the output layer. q,res | 2 The strength of the output light signal is used as the basis for the final classification judgment. For example... Figure 4 As shown in the bar chart on the right, the value of |S| is obtained by comparing the intensity of the output light with the intensity of the input probe light. q | 2 At the number 1, |S q | 2 The maximum value corresponds to the maximum output signal light intensity, and the image is ultimately determined to be the number 1.
[0073] Figure 5 They were shown respectively Figure 4 The performance of the cases in the training sample. Figure 5(a) A graph showing the number of training iterations versus accuracy for this example, which ultimately achieved an accuracy of 93.15%; Figure 5 (b) is the confusion matrix diagram for this case; taking the number 9 as an example, there are 1009 images of the number 9. In the recognition, the network recognized 9 as 9 920 times and recognized 9 as 4 31 times. Recognizing 9 as 4 was the most frequent error. In fact, 9 and 4 are the closest in similarity in MNIST handwritten characters. The occurrence of this recognition error is reasonable and in line with the cognitive rules of human vision. It also proves that the present invention has a certain degree of authenticity and reliability.
[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. An all-optical nonlinear neural network system based on a linear system, characterized in that: Including those connected sequentially: The input module is used to apply the input signal of the neural network to the detuning of the resonant wavelength of the ring resonant cavity; The neuron module includes a multi-layered ring resonant cavity array for linear transmission of optical signals; The coupling module connects the input module and the neuron module, the layers of the neuron module, and the neuron module and the output module. It is used to map the weight coefficients of the neural network by adjusting the coupling strength. The output module, connected to the last layer of the neuron module, is used to load the output signal of the neural network onto the output optical power; In this system, the input and output signals are applied to different physical quantities, enabling the linear optical system to perform nonlinear calculations. The output optical power in the nonlinear calculations is derived from coupled-mode theory and satisfies the following conditions: ， in, Pass functions to the input module. This represents the transfer function of the output module. This represents the input signal of the neural network. This represents the total loss of the resonant cavity. This is the detuning of the resonant cavity. The coupling strength is denoted as .
2. The linear system based all-optical nonlinear neural network system of claim 1, wherein: The input module includes: The resonant wavelength modulation device array is directly connected to the ring resonant cavity array and is used to modulate the resonant wavelength; A ring resonant cavity array, whose input terminal receives the input signal and whose output terminal is connected to a coupling module; The coupling module includes: A ring waveguide connects adjacent neuron layers; The coupling region array is physically coupled to the ring waveguide and the ring resonant cavity of the adjacent layer, respectively. An array of coupling strength control devices is directly connected to a ring waveguide and used to control the coupling strength. The neuron module includes: A multi-layer ring resonant cavity array, each layer of which is connected to a ring waveguide through a coupling region array; The output module includes: A ring resonant cavity array, with its input end connected to the terminal layer neurons via a coupling region array; The probe waveguide array is coupled to the ring resonant cavity array through the coupling region array; The coupling region array is connected to the ring resonant cavity array and the probe waveguide array, respectively.
3. The all-optical nonlinear neural network system based on a linear system according to claim 2, characterized in that: The resonant wavelength modulation device is a thermo-optic modulation device or an electro-optic modulation device; The ring resonant cavity is in the form of a circular, racetrack-shaped, or irregularly shaped closed waveguide, with a strip-shaped or ridge-shaped cross-section, transmitting single-mode or multi-mode optical signals.
4. The linear system based all-optical nonlinear neural network system of claim 2, wherein: The coupling strength adjustment device is a thermo-optic modulation device or an electro-optic modulation device; The ring waveguide can be circular, racetrack-shaped, or irregularly shaped closed waveguide.
5. The linear system based all-optical nonlinear neural network system of claim 2, wherein: The coupling region is a coupling structure between a silicon-based waveguide and a ring resonant cavity, including any of the following: Coupled between two straight waveguides; Single-point or multi-point coupling between straight waveguides and curved waveguides; Single-point or multi-point coupling between curved waveguides.
6. The all-optical nonlinear neural network system based on a linear system according to claim 2, characterized in that: The input light of the probe waveguide is a single-mode or multi-mode optical signal, and the output light power satisfies the formula: ， in, For transfer functions, To detect the coupling strength between the waveguide and the ring resonator, This is the transfer function for the output module. This is the input optical complex amplitude.
7. A method for implementing an all-optical nonlinear neural network based on the system described in any one of claims 1 to 6, characterized in that: Includes the following steps: Input step: inputting an input signal through an input module an amount of detuning of a resonant wavelength of the ring resonator cavity Weight regulation step: regulating the interlayer coupling strength by coupling module , realize neural network weight mapping; Linear transmission steps: Optical signals are linearly transmitted in the neuron module via a ring resonant cavity array; Nonlinear output step: The probe waveguide output optical power is measured by the output module as a result of the nonlinear calculation. , as a result of the nonlinear calculation.