Optical neural network and optical neural chip based on dispersion nonlinearity

Abstract

This invention provides an optical neural network and optical neural chip based on dispersive nonlinearity, relating to the field of optical computing technology. The optical neural network includes: an optical encoding module comprising a tunable multi-wavelength light source for encoding data to be processed into light input signals of different wavelengths; a tunable dispersion module utilizing a tunable dispersion material based on structural or material dispersion to modulate the intensity of the light input signal according to its wavelength, thereby obtaining a light output signal; wherein the intensity of the light output signal and the wavelength of the light input signal have a nonlinear functional relationship; and a fully connected layer module comprising a photodetector for detecting the intensity of the light output signal to obtain a prediction result of the data to be processed based on the intensity of the light output signal. This invention provides an optical neural network with a simple structure, capable of implementing arbitrarily tunable nonlinear activation functions, and capable of realizing arbitrarily complex neural network functions.

Description

Technical Field

[0001] This invention relates to the field of optical computing technology, and in particular to an optical neural network based on dispersive nonlinearity and an optical neural chip. Background Technology

[0002] With the increasingly widespread and in-depth application of artificial intelligence, the demand for data processing, storage, and computing performance is also increasing dramatically. This is especially true in the field of generative artificial intelligence, which can generate complex information such as images, text, and music, placing higher demands on computing resources and energy efficiency. Traditionally, electronic chips based on complementary metal-oxide-semiconductor (CMOS) technology have followed Moore's Law, continuously reducing device size and improving performance. However, as the integrated circuit industry enters the post-Moore's Law era, traditional transistors are facing physical limitations, such as the quantum tunneling effect, while the in-memory / computation separation bottleneck of the von Neumann architecture is becoming increasingly prominent. Furthermore, while facing limits in terms of heat generation and power consumption, electronic chips also need to cope with the exponential growth in computing power demands brought about by artificial intelligence training and computation.

[0003] To address these issues, researchers have begun exploring optical neural networks as an alternative. Optical neural networks leverage the high bandwidth, low loss, low power consumption, high parallelism, and low latency of optical transmission, promising to significantly improve the computational performance and power efficiency of neural networks. However, existing optical neural network technologies are primarily based on multilayer perceptrons (MLPs). The high complexity of implementing fully connected layers and the difficulty in cascading layers within an in-plane optical neural network limit the construction of complex network structures, such as large-scale fully connected layers and multiple hidden layers. Furthermore, while MLP-based neural networks can achieve complex functions, their weights are contained within the nodes and the sheer number of parameters results in low interpretability, making them resemble a "black box."

[0004] Therefore, how to provide an optical neural network with a simple structure that can realize arbitrarily tunable nonlinear activation functions is an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides an optical neural network and an optical neural chip based on dispersive nonlinearity, which provides an optical neural network with a simple structure, capable of implementing arbitrarily tunable nonlinear activation functions, and capable of realizing arbitrarily complex neural network functions. This invention provides an optical neural network based on dispersive nonlinearity, comprising the following modules.

[0006] The optical encoding module includes a tunable multi-wavelength light source for encoding the data to be processed into optical input signals of different wavelengths; a tunable dispersion module for using a tunable dispersion material based on structural dispersion or material dispersion to modulate the intensity of the optical input signal according to the wavelength of the optical input signal to obtain an optical output signal; wherein the intensity of the optical output signal and the wavelength of the optical input signal have a nonlinear functional relationship; and a fully connected layer module includes a photodetector for detecting the intensity of the optical output signal to obtain a prediction result of the data to be processed based on the intensity of the optical output signal.

[0007] According to the present invention, an optical neural network based on dispersive nonlinearity is provided, wherein the tunable dispersion module includes: a dispersion layer, which is composed of a tunable dispersion material based on structural dispersion or material dispersion, for modulating the intensity of the optical input signal according to the wavelength of the optical input signal; and a nonlinear control layer, which is used to adjust the intensity of the optical input signal entering the nonlinear control layer using a preset method.

[0008] According to an optical neural network based on dispersive nonlinearity provided by the present invention, adjusting the intensity of the light input signal entering the nonlinear control layer using a preset method includes: changing the dispersive characteristics of the dispersive layer by acting on the physical field of the tunable dispersive module, so as to adjust the intensity of the light input signal entering the nonlinear control layer.

[0009] According to an optical neural network based on dispersive nonlinearity provided by the present invention, adjusting the intensity of the light input signal entering the nonlinear control layer using a preset method includes: adjusting the intensity of the light input signal entering the nonlinear control layer using an optical control device acting on the light input signal entering the nonlinear control layer.

[0010] According to an optical neural network based on dispersive nonlinearity provided by the present invention, adjusting the intensity of the light input signal entering the nonlinear control layer using a preset method includes: adjusting the photosensitivity of the photodetector of the fully connected layer module to adjust the intensity of the light output signal detected by the photodetector.

[0011] According to an optical neural network based on dispersive nonlinearity provided by the present invention, the light output signal is projected onto the target light receiving area of ​​the photodetector; the photodetector is used to detect the total intensity of the light output signal through the target light receiving area, and obtain the linear sum of the intensities of the light output signal at different wavelengths.

[0012] According to the present invention, an optical neural network based on dispersive nonlinearity is provided, wherein the tunable dispersion module includes multiple tunable dispersion units; the optical input signal is encoded by an optical encoding module into multiple optical input sub-signals, each optical input sub-signal entering a corresponding tunable dispersion unit; the tunable dispersion unit is used to modulate the intensity of the optical input sub-signal according to the wavelength of the optical input sub-signal entering it, to obtain a beam of optical output sub-signal; wherein the intensity of the optical output sub-signal corresponding to different tunable dispersion units has different nonlinear functional relationships with the wavelength of the optical input sub-signal; the photodetector includes multiple light receiving areas, the light receiving areas being used to detect the total intensity of the optical output sub-signal output by their corresponding tunable dispersion units.

[0013] The present invention also provides an optical neural chip for implementing the optical neural network based on dispersion nonlinearity as described above.

[0014] The present invention also provides a prediction method for an optical neural network, comprising the following steps.

[0015] The process involves: acquiring data to be processed; encoding the data into light input signals of different wavelengths using an optical encoding module; modulating the intensity of the light input signals according to their wavelengths using an adjustable dispersion module to obtain a light output signal; wherein the intensity of the light output signal and the wavelength of the light input signal have a nonlinear functional relationship; and detecting the intensity of the light output signal using a fully connected layer module to obtain a prediction result for the data to be processed based on the intensity of the light output signal.

[0016] According to the prediction method of the optical neural network provided by the present invention, the step of using an adjustable dispersion module to modulate the intensity of the optical input signal according to the wavelength of the optical input signal includes: using a dispersion layer to modulate the intensity of the optical input signal according to the wavelength of the optical input signal; and using a nonlinear control layer to adjust the intensity of the optical input signal entering the nonlinear control layer using a preset method.

[0017] This invention provides an optical neural network based on dispersive nonlinearity. Utilizing a tunable dispersion module, the intensity of the input light signal is modulated according to its wavelength to obtain the output light signal. This allows for an arbitrary, tunable nonlinear functional relationship between the intensity of the output light signal and the wavelength of the input light signal. By employing a fully connected layer module to detect the intensity of the output light signal, accurate predictions of the data to be processed can be obtained based on the output light signal intensity. Therefore, this invention provides an optical neural network with a simple structure, capable of implementing arbitrarily tunable nonlinear activation functions, and capable of realizing arbitrarily complex neural network functions (e.g., the KAN network architecture). Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the schematic diagrams of the optical neural network based on dispersive nonlinearity provided by the present invention.

[0020] Figure 2 This is the second schematic diagram of the structure of the optical neural network based on dispersive nonlinearity provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the adjustable dispersion module provided by the present invention.

[0022] Figure 4 This is a flowchart illustrating the prediction method of the optical neural network provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] To address the problems of high complexity, difficulty in implementing nonlinear activation functions, and low interpretability of existing optical neural networks, this invention proposes a novel optical neural network based on the Kolmogorov-Arnold Networks (KAN) architecture: Optics Kolmogorov-Arnold Networks (OKAN). For ease of description, this specification uses the English abbreviation OKAN to refer to this optical neural network, which can realize arbitrarily complex neural network functions and achieves the physical mapping of arbitrarily tunable nonlinear functions through spectral techniques.

[0025] The KAN neural network was inspired by the Kolmogorov-Arnold representation theorem, which proves that any multivariable continuous function... All of them can be expressed as a combination of a finite number of single-variable continuous functions: in, , It is a single-variable function.

[0026] Based on the above theorem, the KAN neural network uses a completely different approach to construct the neural network compared to MLP. In traditional MLP networks, linear weights are located at the "edges" of the neural network, and the data is weighted and then input into fixed activation functions in the "nodes" for nonlinear operations. The KAN neural network architecture is entirely different. The trainable activation functions are located at the "edges" of the neural network, and after the nonlinear operations of the activation functions, only linear summation is performed at the "nodes." Therefore, the KAN neural network does not have linear weights; instead, each univariate function uses spline functions as its basis functions.

[0027] Therefore, the KAN neural network decomposes complex high-dimensional functions into combinations of simple univariate functions. It models and fits complex problems by optimizing these univariate functions rather than the entire linear variable space, thus significantly reducing model complexity and the number of training parameters. Furthermore, the simplicity of the univariate functions greatly enhances the interpretability of the KAN neural network. In addition, due to the Kolmogorov-Arnold representation theorem, the KAN neural network can, in principle, fit functions of arbitrary complexity using only a two-layer neural network structure, significantly reducing network complexity and making it more suitable for optical integration.

[0028] The core of optical device design for realizing KAN neural networks lies in how to design arbitrarily tunable nonlinear functions. This invention proposes that the nonlinear function can be realized through a function corresponding to the optical transmission spectrum, which is formed by... Figure 1 The adjustable dispersion module shown generates this.

[0029] The following is combined Figures 1-3 The present invention describes an optical neural network based on dispersive nonlinearity.

[0030] Figure 1 This is one of the structural schematic diagrams of the optical neural network based on dispersive nonlinearity provided by the present invention.

[0031] like Figure 1 As shown, the optical neural network includes: an optical coding module, an adjustable dispersion module, and a fully connected layer module.

[0032] The optical encoding module includes a tunable multi-wavelength light source, which is used to encode the data to be processed into optical input signals of different wavelengths.

[0033] A tunable multi-wavelength light source is a light source whose output wavelength is continuously or discretely adjustable within a certain range, and can generate multiple light signals of different wavelengths simultaneously or time-divisionally. For example, a tunable multi-wavelength light source can be a multi-channel LED / LD light source, etc.

[0034] The data to be processed is data that needs to be processed by an optical neural network, such as handwritten digits or characters in handwriting recognition.

[0035] In the specific implementation process, an optical encoding module can be used to encode the data to be processed into optical input signals of different wavelengths in various ways, with different wavelengths representing different values ​​of independent variables.

[0036] For example, when performing handwriting recognition tasks, the optical encoding module uses a tunable multi-wavelength light source to encode each data point or data feature of the handwritten character into an optical input signal of different wavelengths.

[0037] Different colors can represent different wavelengths of light. For example, the wavelength of red light is approximately between 630 and 780 nanometers, while the wavelength of blue light is between 470 and 420 nanometers.

[0038] In some embodiments, the light encoding module can encode different intensity information in an image into light of different wavelengths. As an example only, when performing a handwriting recognition task, the brightness or on / off state of the light source can be adjusted to emit light of a wavelength corresponding to the color of each pixel on the image corresponding to the handwritten character.

[0039] The optical neural network provided by this invention encodes light input signals through the direct correspondence between light and pixels. Compared with the in-plane optical neural network implementation, this optical neural network has a smaller area and higher parallelism and integration.

[0040] The tunable dispersion module is used to modulate the intensity of the light input signal according to the wavelength of the light input signal using a tunable dispersion material based on structural dispersion or material dispersion, thereby obtaining a light output signal; wherein, there is a non-linear functional relationship between the intensity of the light output signal and the wavelength of the light input signal.

[0041] Dispersive materials exhibiting structural dispersion can include, but are not limited to, metasurfaces, metamaterials, photonic crystals, and surface plasmon polaritons (SPPs). Dispersive materials exhibiting material dispersion can include, but are not limited to, dyes, pigments, perovskites, ferroelectric materials, quantum dots, and nanowires, all of which possess specific absorption spectral lines.

[0042] When an optical input signal enters the tunable dispersion module, due to the dispersive characteristics of the tunable dispersion material based on structural or material dispersion, optical signals of different wavelengths will experience different propagation paths and phase delays, causing the intensity of the optical signals of different wavelengths to change accordingly. For example, the tunable dispersion material based on structural or material dispersion will reduce the intensity of optical signals within a specific wavelength range to a certain level; while for optical signals within other wavelength ranges, it may reduce their intensity to a lower level. Thus, the tunable dispersion module can be used to modulate the intensity of the optical input signal according to its wavelength.

[0043] During the training of optical neural networks, various distinct optical transmission spectra can be constructed by adjusting the dispersive properties of tunable dispersive materials based on structural or material dispersion. Furthermore, these optical transmission spectra can be used to realize an arbitrarily tunable nonlinear functional relationship between the intensity of the output light signal and the wavelength of the input light signal.

[0044] In practical implementation, arbitrarily tunable nonlinear function relationships can be realized through various methods using tunable dispersion modules. For specific embodiments, please refer to [link to specific implementation examples]. Figure 3 The relevant content will not be repeated here.

[0045] The nonlinear functional relationship is a functional relationship determined based on the target task to be performed by the optical neural network.

[0046] As an example, to perform a handwriting recognition task, an optical neural network needs to learn and determine a nonlinear functional relationship. This functional relationship, through a complex mapping, can convert the wavelength information of the light input signal into the intensity information of the light output signal, thereby enabling accurate differentiation of different handwritten characters based on the intensity information of the light output signal.

[0047] In practical implementation, the tunable dispersion material in the tunable dispersion module, based on structural or material dispersion, can be divided into multiple parts. Each part is arranged in an array at regular intervals on a two-dimensional plane. The light input signal is encoded in the wavelength dimension while maintaining the same light intensity before entering the array. Therefore, a tunable dispersion module can be regarded as an optical KAN neural network layer, i.e., an OKAN layer, where one array can be regarded as a neuron node. For example, a 5x5 array of tunable dispersion material based on structural or material dispersion constitutes an OKAN layer with 25 neurons capable of parallel processing.

[0048] The optical neural network proposed in this invention has good scalability. By increasing the array size of the tunable dispersion module in the plane, it is equivalent to increasing the dimension of the nonlinear activation function in an OKAN layer.

[0049] In some embodiments, such as Figure 2 As shown, an optical neural network can include multiple tunable dispersion modules. The array configuration within each tunable dispersion module is different. The light output signal of each tunable dispersion module, after intensity-to-wavelength mapping, serves as the light input signal for the next tunable dispersion module. This is equivalent to stacking OKAN layers with different parameters and different numbers of input and output nodes, thereby enabling the implementation of more complex network functions.

[0050] The fully connected layer module includes a photodetector for detecting the intensity of the light output signal, so as to obtain the prediction result of the data to be processed based on the intensity of the light output signal.

[0051] In some embodiments, the light output signal is projected onto the target light receiving area of ​​the photodetector; the photodetector is used to detect the total intensity of the light output signal through the target light receiving area, and to obtain the linear sum of the intensities of the light output signals at different wavelengths.

[0052] In some embodiments, the tunable dispersion module includes multiple tunable dispersion units; the optical input signal is encoded by the optical encoding module into multiple optical input sub-signals, each optical input sub-signal entering a corresponding tunable dispersion unit; the tunable dispersion unit is used to modulate the intensity of the optical input sub-signal according to the wavelength of the optical input sub-signal entering it, to obtain a beam of optical output sub-signal; wherein, the intensity of the optical output sub-signal corresponding to different tunable dispersion units and the wavelength of the optical input sub-signal have different nonlinear functional relationships; the photodetector includes multiple light receiving areas, which are used to detect the total intensity of the optical output sub-signal output by their corresponding tunable dispersion units.

[0053] In this embodiment, each light receiving area can realize an output node of a KAN neural network, and the intensity of each light output sub-signal can be used as the value of the corresponding output node.

[0054] The optical neural network OKAN provided by this invention can be widely applied to various scenarios, such as image recognition and multi-class prediction. Furthermore, in scenarios requiring greater interpretability, such as solving differential equations, fitting data, and graph neural networks, the optical neural network proposed in this invention can achieve simpler, more direct results and better interpretability.

[0055] This invention overcomes the limitation of existing technologies in realizing arbitrary tunable optical nonlinear functions through a tunable dispersion module, achieving the OKAN optical neural network based on KAN. This network can easily realize tunable nonlinear functions, thus possessing great potential for online training. After training is complete and network parameters are fixed, this optical neural network only requires light input to operate, achieving zero-power inference. According to the Kolmogorov-Arnold representation theorem, the optical neural network provided by this invention can achieve equivalent functionality to more complex multilayer perceptron (MLP) networks with fewer training parameters, and can construct MLP networks of arbitrary complexity. Furthermore, through in-depth analysis of univariate nonlinear functions, the optical neural network provided by this invention offers better interpretability than traditional MLP networks.

[0056] Figure 3 This is a schematic diagram of the adjustable dispersion module provided by the present invention. Figure 3 As shown, the adjustable dispersion module includes a dispersion layer and a nonlinear control layer.

[0057] The dispersion layer, composed of tunable dispersion materials based on structural dispersion or material dispersion, is used to modulate the intensity of the light input signal according to the wavelength of the light input signal.

[0058] In practice, the intensity of light input signals of different wavelengths output by the optical encoding module is the same. After passing through the dispersion layer, under the action of tunable dispersion materials based on structural dispersion or material dispersion, the intensity of light input signals of different wavelengths is modulated to different values.

[0059] The nonlinear control layer is used to adjust the intensity of the light input signal entering the nonlinear control layer using a preset method and a nonlinear functional relationship, thereby establishing a nonlinear functional relationship between the wavelength of the light input signal output from the optical encoding module and the intensity of the light output signal output from the tunable dispersion module.

[0060] In some embodiments, the dispersion characteristics of the dispersion layer (e.g., piezoelectric metasurface, phase change material metasurface, etc.) can be changed by the physical field (e.g., electromagnetic field, temperature field, etc.) acting on the tunable dispersion module, so as to adjust the intensity of the light input signal entering the nonlinear control layer, thereby realizing an arbitrary tunable nonlinear function relationship.

[0061] As an example only, the tunable dispersion module includes an electro-optic material as a dispersion layer. A pair of electrodes are positioned in the nonlinear control layer, through which a controllable voltage is applied to the electro-optic material, thereby generating an electric field. The electro-optic effect causes the refractive index of the electro-optic material to change with the strength of the electric field, thus achieving adjustment of the intensity of the light input signal.

[0062] In some embodiments, the intensity of the light input signal entering the nonlinear modulation layer can be adjusted using an optical modulation device that acts on the light input signal entering the nonlinear modulation layer.

[0063] Optical control devices may include, but are not limited to: electrical amplifiers (e.g., adjusting the analog gain inside a CCD), DMDs (digital micromirrors), liquid crystals (spatial light modulators), memristors, etc.

[0064] In some embodiments, the intensity of the light output signal detected by the photodetector can be adjusted by regulating (e.g., based on bias regulation) the photosensitivity of the photodetector in the fully connected layer module, which is equivalent to adjusting the intensity of the light input signal entering the nonlinear modulation layer.

[0065] In the embodiments provided by the present invention, arbitrary tunable nonlinear activation functions can be efficiently realized through the collaborative work of the dispersive layer and the nonlinear control layer.

[0066] Figure 4 This is a flowchart illustrating the prediction method of the optical neural network provided by the present invention, as shown below. Figure 4 As shown, the method includes the following: Step 401: Obtain the data to be processed.

[0067] As an example only, for handwriting recognition tasks, handwritten numbers, letters, or symbols can be collected through handwriting input devices (such as touch screens or handwriting tablets) as data to be processed.

[0068] Step 402: Use the optical encoding module to encode the data to be processed into optical input signals of different wavelengths.

[0069] For a detailed description of this step, please refer to [link / reference]. Figure 1 The relevant content will not be repeated here.

[0070] Step 403: Using an adjustable dispersion module, the intensity of the light input signal is modulated according to the wavelength of the light input signal to obtain the light output signal; wherein, there is a non-linear functional relationship between the intensity of the light output signal and the wavelength of the light input signal.

[0071] For a detailed description of this step, please refer to [link / reference]. Figure 1 and Figure 3 The relevant content will not be repeated here.

[0072] Step 404: Use the fully connected layer module to detect the intensity of the light output signal, so as to obtain the prediction result of the data to be processed based on the intensity of the light output signal.

[0073] For a detailed description of this step, please refer to [link / reference]. Figure 1 The relevant content will not be repeated here.

[0074] In practice, the photodetector converts the intensity of the received light output signal into an electrical signal, and the processing device can obtain the prediction result of the data to be processed based on the electrical signal, such as the recognition result of handwritten characters.

[0075] In some embodiments, a dispersive layer can be used to modulate the intensity of the optical input signal according to the wavelength of the optical input signal.

[0076] A nonlinear modulation layer is used to adjust the intensity of the optical input signal entering the nonlinear modulation layer using a preset method.

[0077] On the other hand, the present invention also provides an optical neural chip for implementing the dispersion-based nonlinear optical neural network as described above. It includes: an optical encoding module comprising a tunable multi-wavelength light source for encoding data to be processed into optical input signals of different wavelengths; a tunable dispersion module for modulating the intensity of the optical input signal according to the wavelength of the optical input signal using a tunable dispersion material based on structural dispersion or material dispersion to obtain an optical output signal; wherein the intensity of the optical output signal and the wavelength of the optical input signal have a nonlinear functional relationship; and a fully connected layer module comprising a photodetector for detecting the intensity of the optical output signal to obtain a prediction result of the data to be processed based on the intensity of the optical output signal.

[0078] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An optical neural network based on dispersion nonlinearity, characterized in that, include: The optical encoding module includes a tunable multi-wavelength light source for encoding the data to be processed into optical input signals of different wavelengths. A tunable dispersion module is used to modulate the intensity of the light input signal according to the wavelength of the light input signal using a tunable dispersion material based on structural dispersion or material dispersion, thereby obtaining a light output signal; wherein the intensity of the light output signal and the wavelength of the light input signal have a non-linear functional relationship. The fully connected layer module includes a photodetector for detecting the intensity of the light output signal, so as to obtain a prediction result of the data to be processed based on the intensity of the light output signal.

2. The optical neural network based on dispersion nonlinearity according to claim 1, characterized in that, The adjustable dispersion module includes: A dispersive layer, composed of a tunable dispersive material based on structural dispersion or material dispersion, is used to modulate the intensity of the optical input signal according to the wavelength of the optical input signal; A nonlinear modulation layer is used to adjust the intensity of the optical input signal entering the nonlinear modulation layer using a preset method.

3. The optical neural network based on dispersive nonlinearity according to claim 2, characterized in that, The method of adjusting the intensity of the optical input signal entering the nonlinear modulation layer using a preset method includes: By applying a physical field to the tunable dispersion module, the dispersion characteristics of the dispersion layer are changed, thereby adjusting the intensity of the light input signal entering the nonlinear modulation layer.

4. The optical neural network based on dispersive nonlinearity according to claim 2, characterized in that, The method of adjusting the intensity of the optical input signal entering the nonlinear modulation layer using a preset method includes: The intensity of the light input signal entering the nonlinear modulation layer is adjusted using an optical modulation device that acts on the light input signal entering the nonlinear modulation layer.

5. The optical neural network based on dispersion nonlinearity according to claim 2, wherein, The method of adjusting the intensity of the optical input signal entering the nonlinear modulation layer using a preset method includes: The intensity of the light output signal detected by the photodetector can be adjusted by regulating the photosensitivity of the photodetector in the fully connected layer module.

6. The optical neural network based on dispersion nonlinearity according to any one of claims 1-5, wherein, The optical output signal is projected onto the target light receiving area of ​​the photodetector; The photodetector is used to detect the total intensity of the light output signal through the target light receiving area, and to obtain a linear sum of the intensities of the light output signal at different wavelengths.

7. The optical neural network based on dispersion nonlinearity according to any one of claims 1-5, wherein, The adjustable dispersion module includes multiple adjustable dispersion units; The optical input signal is encoded into multiple optical input sub-signals by the optical encoding module, and each optical input sub-signal enters a corresponding tunable dispersion unit. The tunable dispersion unit is used to modulate the intensity of the light input sub-signal according to the wavelength of the light input sub-signal entering it, to obtain a light output sub-signal; wherein, the intensity of the light output sub-signal and the wavelength of the light input sub-signal corresponding to different tunable dispersion units have different nonlinear functional relationships. The photodetector includes multiple light receiving areas, which are used to detect the total intensity of the light output sub-signals output by their corresponding tunable dispersion units.

8. An optical neurochip, characterized by, Used to implement the optical neural network based on dispersive nonlinearity as described in any one of claims 1 to 7.

9. A prediction method for an optical neural network, applied to an optical neural network based on dispersive nonlinearity as described in any one of claims 1 to 7, the method comprising: Obtain the data to be processed; The data to be processed is encoded into optical input signals of different wavelengths using an optical encoding module. An adjustable dispersion module is used to modulate the intensity of the light input signal according to the wavelength of the light input signal to obtain a light output signal; wherein, the intensity of the light output signal and the wavelength of the light input signal have a non-linear functional relationship. The intensity of the optical output signal is detected using a fully connected layer module, so as to obtain the prediction result of the data to be processed based on the intensity of the optical output signal.

10. The method of claim 9, wherein, The method of using a tunable dispersion module to modulate the intensity of the optical input signal according to the wavelength of the optical input signal includes: The intensity of the optical input signal is modulated according to the wavelength of the optical input signal using a dispersive layer. A nonlinear modulation layer is used to adjust the intensity of the optical input signal entering the nonlinear modulation layer using a preset method.

Images