Nonlinear optical neural network system and nonlinear optical computing method

Abstract

This specification provides a nonlinear optical neural network system and a nonlinear optical calculation method. The system includes: an optical feedback structure for forming a loop optical path to allow a light beam to propagate cyclically within the system; an input coupling component and an output coupling component, respectively used to couple an input light beam to the loop optical path and to couple a portion of the light beam from the loop optical path as an output light beam; a spatial light modulation device for modulating at least one of the amplitude and phase of the input light beam; and a diffraction control component for performing multi-layer diffraction operations and control on the input light beam. The spatial light modulation device and the diffraction control component are both disposed within the loop optical path, configured such that the light beam passes sequentially through the spatial light modulation device and the diffraction control component in each loop, and the light beam reaches the output coupling component after passing through the diffraction control component.

Description

Technical Field

[0001] One or more embodiments of this specification relate to the fields of optical computing and artificial intelligence, and in particular to a nonlinear optical neural network system and a nonlinear optical computing method. Background Technology

[0002] Optical neural networks, with their potential for high computational speed, ultra-high parallelism, low latency, and low power consumption, are considered an important supplement or alternative to traditional electronic computing technologies. However, the development of all-optical neural networks faces a core bottleneck: how to efficiently achieve controllable nonlinear mapping in the optical domain. Nonlinear mapping is crucial for neural networks to fit complex functions, identify high-level patterns, and learn abstract features.

[0003] Existing technologies for achieving optical nonlinearity suffer from the following drawbacks. One approach utilizes the nonlinear effects of nonlinear optical materials, but this method typically requires high-intensity pump lasers, leading to high system energy consumption, difficulty in miniaturization and integration, and its nonlinear response function is usually fixed with poor tunability. Another approach employs an optical-electrical-optical conversion method, first converting the optical signal into an electrical signal, then processing the nonlinear function using mature electronic circuits, and finally modulating the processed electrical signal back into an optical signal. While this approach can achieve flexible nonlinear functions, the introduced conversion process sacrifices the core advantages of optical computing in terms of speed and parallelism, and introduces additional energy loss and signal delay.

[0004] To leverage the advantages of optical computing, some technical solutions propose using optical feedback or recurrent structures to construct optical computing devices. However, in these existing designs, the input information is typically loaded and modulated onto the beam only once outside the recurrent optical path, and the beam entering the recurrent optical path only undergoes cyclic propagation and processing within the computational network. While this architecture can simulate the function of recurrent neural networks in processing time-series information, it cannot directly generate higher-order nonlinear terms related to the input signal through iterative processes because the beam no longer interacts with the input information during the recurrent process. Essentially, it remains limited by the expressive power of linear operations and fails to fundamentally solve the technical challenge of efficiently generating controllable nonlinear mappings across the entire optical domain. Therefore, a novel optical neural network architecture is urgently needed to achieve efficient, low-power, and controllable nonlinear computing across the entire optical domain. Summary of the Invention

[0005] In view of this, this specification aims to address the problems of high energy consumption, slow speed, poor controllability, or complex structure faced by existing optical neural networks in realizing all-optical nonlinear mapping. To achieve the above objective, one or more embodiments in this specification provide a nonlinear optical neural network system, including: An optical feedback structure is used to form a circulating optical path so that the light beam can circulate in the system. An input coupling component and an output coupling component are respectively used to couple an input beam to the circulating optical path and to couple a portion of the beam out from the circulating optical path; wherein, the input beam includes a coherent beam; A spatial light modulation device for controlling at least one of the amplitude and phase of an input light beam; A diffraction control component for performing multi-layer diffraction operations and control on its input beam; The spatial light modulation device and the diffraction control component are both disposed in the circulating optical path, and their configuration is such that the light beam passes through the spatial light modulation device and the diffraction control component sequentially in each cycle, and the light beam reaches the output coupling component after passing through the diffraction control component.

[0006] More preferably, the optical feedback structure includes an optical resonant cavity composed of at least three mirrors.

[0007] More preferably, the input coupling component and the output coupling component include a beam splitter.

[0008] More preferably, the spatial light modulation device includes a single-layer spatial light modulator, and the diffraction control component includes a multi-layer spatial light modulator arranged along the light propagation direction; the spatial light modulator includes a two-dimensionally distributed programmable control unit.

[0009] More preferably, the spatial light modulation device includes a reflective digital micromirror device.

[0010] More preferably, the diffraction modulation assembly includes a multilayer diffractive optical element (DOE) pre-fabricated with a phase distribution to statically modulate the input beam according to a pre-fabricated modulation mode.

[0011] More preferably, the system further includes a two-dimensional imaging detector for detecting the light field distribution of the output beam.

[0012] According to a second aspect of one or more embodiments of this specification, a method for implementing nonlinear optical computation is also provided, applied to a nonlinear optical neural network system. The system includes an optical feedback structure, an input coupling component, an output coupling component, a spatial light modulator, and a diffraction control component. The optical feedback structure is used to form a loop optical path to allow a light beam to propagate cyclically in the system. The spatial light modulator and the diffraction control component are both disposed in the loop optical path, and their configuration is such that the light beam passes through the spatial light modulator and the diffraction control component sequentially in each loop, and the light beam reaches the output coupling component after passing through the diffraction control component. The method includes the following steps: The input coupling component is used to couple at least a portion of the input beam into the loop optical path, the input beam comprising a coherent beam; The light beam is made to propagate cyclically in the circular optical path, and in each cycle, the following steps are performed sequentially: The spatial light modulation device is used to programmably spatially control at least one of the amplitude and phase of the input light beam; and Using the diffraction control component, multi-layer diffraction operations and control are performed on the spatially modulated beam; At least a portion of the beam that has undergone the multi-layer diffraction calculations and modulation is coupled out of the loop optical path.

[0013] More preferably, a two-dimensional imaging detector is used to detect the output beam coupled out of the cyclic optical path, wherein the output beam is a coherent superposition of beams of various orders coupled out after the beam has passed through the cyclic optical path a preset number of cycles.

[0014] More preferably, the spatial modulation step and the spatial transformation operation step are executed by a programmable spatial light modulator.

[0015] As can be seen from the above, in one or more embodiments of this specification, by placing both the spatial light modulation device and the diffraction control component inside the optical feedback structure, and allowing the light beam to pass through both sequentially in each cycle, the cyclic propagation of the light beam enables simple linear modulation and computation to equivalently generate a nonlinear transformation containing higher-order terms of the input information after multiple iterations. By adjusting the spatial light modulation device and the diffraction control component, the weights of different-order nonlinear terms can be controlled, thereby achieving controllable all-optical nonlinear mapping and enhancing the computational and expressive capabilities of the optical neural network. Furthermore, one or more embodiments provided in this specification achieve high hardware efficiency and a compact structure. That is, by cyclically reusing the same set of spatial light modulation devices and diffraction control components, the computational effect of a deep neural network is equivalently achieved without physically stacking a large number of optical components. This significantly improves hardware utilization efficiency, makes the system structure more compact, and reduces manufacturing costs and complexity. Meanwhile, one or more embodiments provided in this specification also achieve high energy efficiency and high speed, adopt a passive or low-power optical feedback structure, and realize nonlinear functions based on linear optical elements, avoiding the high-power pump light required by nonlinear crystals and the additional energy consumption and delay caused by photoelectric conversion; therefore, while realizing nonlinear calculation, the system inherits the inherent advantages of low power consumption and high speed of optical calculation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of a nonlinear optical neural network system based on a three-mirror resonator provided in an exemplary embodiment.

[0017] Figure 2 This is an illustrated diagram of the nonlinear generation principle provided in an exemplary embodiment.

[0018] Figure 3 This is a schematic diagram of the principle of a multilayer diffraction neural network provided in an exemplary embodiment.

[0019] Figure 4 This is a schematic diagram of the structure of a nonlinear optical neural network system based on a four-mirror resonator provided in an exemplary embodiment.

[0020] Figure 5 This is a schematic diagram illustrating the computational simulation results of a nonlinear diffraction neural network and a traditional linear diffraction neural network on different datasets, provided by an exemplary embodiment.

[0021] Figure 6 This is a flowchart of an exemplary embodiment of a nonlinear optical computation method. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0023] This specification provides an illustrative embodiment of a nonlinear optical neural network system based on optical feedback, the specific physical structure of which can be found in [reference needed]. Figure 1 . Figure 1 This is a schematic diagram of a nonlinear optical neural network system based on a three-mirror resonator provided in an embodiment of this specification. The system in this embodiment, through a novel and specific optical architecture, can efficiently and controllably generate nonlinear computational responses in the pure optical domain, thereby significantly improving the computational and expressive capabilities of the optical neural network.

[0024] like Figure 1 As shown, the system in this embodiment mainly includes an optical feedback structure, an input coupling component, an output coupling component, a spatial light modulation device, and a diffraction control component.

[0025] Specifically, in this embodiment, the optical feedback structure is embodied as an optical resonant cavity composed of three highly reflective mirrors. This resonant cavity is formed by mirrors 102, 105, and 107, which together enclose a stable triangular closed optical path. This geometric configuration, such as an equilateral or isosceles triangle, is carefully designed to ensure that the light beam can stably circulate multiple times within the cavity without diverging or deviating from the optical path, providing a solid physical basis for subsequent iterative calculations. It is understood that mirrors 102, 105, and 107 are preferably dielectric film mirrors with high reflectivity (e.g., greater than 99.5%, or even greater than 99.9%). These mirrors are typically composed of multiple layers of dielectric thin films with different refractive indices (such as alternating layers of titanium dioxide / silicon dioxide), which minimizes energy loss of the light beam during each reflection. This directly relates to the quality factor (Q value) of the resonant cavity; a high Q value means a longer average lifetime of photons within the cavity, allowing the light beam to circulate effectively a sufficient number of times, thereby enabling the generation and accumulation of higher-order nonlinear terms.

[0026] In this embodiment, the input coupling component can be a beam splitter 101 with a specific beam splitting ratio, which is partially transmitted / partially reflected. Its function is to couple an external input beam into the loop optical path. For example, a coherent laser beam (e.g., from a single-longitudinal-mode laser) that has been collimated and expanded to match the intracavity mode illuminates the beam splitter 101 at a specific incident angle. The beam splitter 101 transmits a portion of the beam into the resonant cavity formed by the mirrors 102, 105, and 107 as injected energy, while the other portion of the beam is reflected away from the main optical path. By carefully selecting the beam splitting ratio of the beam splitter 101, impedance matching with the resonant cavity can be achieved, thereby optimizing the optical energy injected into the resonant cavity and improving the overall optical efficiency of the system.

[0027] Accordingly, the output coupling component, in this embodiment, can also be a beam splitter 106 with a specific beam splitting ratio. Its function is to couple a portion of the beam from the loop optical path as the output of the calculation result of the entire system. After the beam completes one loop propagation within the resonant cavity, it reaches the beam splitter 106. The beam splitter 106 also has a specific beam splitting ratio. It transmits a small portion of the beam to form the output component of this loop, while reflecting most of the beam back into the resonant cavity, which then propagates through mirrors 107 and 102 to begin the next loop propagation. It is important to note that the beam splitting ratio of the output coupling component is a key adjustable parameter. It determines how much energy is extracted as output in each loop and how much energy is retained for subsequent loops, which directly affects the relative weight distribution of different orders of nonlinear terms in the final output optical field. By adjusting this beam splitting ratio, the intensity and form of the system's nonlinear response can be flexibly controlled.

[0028] It is worth noting that the input coupling component and output coupling component described in this specification are not limited to beam splitters. Based on the properties of the beam and actual needs, those skilled in the art can choose various optical beam splitting coupling devices based on different principles, such as wavelength division multiplexers, polarization beam splitters, and diffraction beam splitters.

[0029] Furthermore, the spatial light modulation device described in this embodiment can be a modulator that can control the light field in at least two-dimensional space, including various types such as spatial light modulators (SLMs), electro-optic modulators, acousto-optic modulators, and semiconductor light modulators, used to modulate the amplitude (light intensity), phase, and other physical properties of the light beam. Figure 1 As shown, in this embodiment, the spatial light modulation device can be a single-layer transmissive spatial light modulator (SLM), such as a liquid crystal spatial light modulator (LCoS-SLM). A spatial light modulator is an optical device that can dynamically and programmably control the spatial distribution of a light field. It typically consists of a two-dimensional pixel array, where each pixel unit can independently modulate the amplitude or phase of the light wave passing through it. In this embodiment, the single-layer spatial light modulator 103 spatially modulates the light beam propagating in the loop optical path based on the information to be processed provided by an external computer or signal source (e.g., a digital image, a data vector, a timing signal, etc.). For example, when performing a handwritten digit recognition task, a 28x28 pixel MNIST handwritten digit image can be loaded as input information onto the spatial light modulator (such as the single-layer spatial light modulator 103). Each pixel unit modulates the amplitude or phase of the light beam according to the grayscale value of the corresponding image pixel. Thus, the input information to be processed is encoded in real time onto the spatial complex amplitude distribution of the light field.

[0030] The aforementioned diffraction modulation component is positioned downstream of the spatial light modulator and is used to perform a preset spatial transformation operation on the light beam modulated by the spatial light modulator. This operation functionally simulates the mathematical transformations of each network layer in an artificial neural network. As an optional implementation, the diffraction modulation component may include one or more two-dimensional light modulators arranged along the light propagation direction, such as... Figure 1 The multilayer spatial light modulator 104 shown forms a multilayer diffraction neural network (DNN). As an example, Figure 3This demonstrates the principle of a multilayer diffraction neural network. After passing through diffraction layer l-1, the propagation of light waves follows the Huygens-Fresnel principle, and each point on the layer can be considered a new spherical wavelet source. These wavelets propagate a distance in free space (physically corresponding to Fresnel or Fraunhofer diffraction) and then coherently superimpose on the plane of diffraction layer l, forming a new light field distribution. By offline training and optimization of the phase or amplitude of each diffraction layer (i.e., each spatial light modulator), arbitrary linear spatial transformations can be achieved. Therefore, the function of the diffraction control component is to apply an optimized transformation operator representing the network weights to the incident light field.

[0031] It is understood that a core technical feature of this embodiment lies in the arrangement of the spatial light modulation device and the diffraction control component. For example... Figure 1 As shown, the single-layer spatial light modulator 103 and the multi-layer spatial light modulator 104 are both arranged in series in the loop optical path formed by the mirrors 102, 105, and 107, located on the propagation path of the light beam. This "intracavity series" configuration is the key to achieving the technical effect of this specification. It ensures that the light beam passes through the two-dimensional light modulator and the diffraction control component sequentially in each loop propagation process, thereby realizing the iterative multiplication of the input information and the network weights.

[0032] It is worth noting that, as an optional implementation, the two-dimensional light modulator and diffraction control components can also be implemented using a reflective spatial light modulator, such as a digital micromirror device (DMD). A DMD is a semiconductor device composed of millions of tiny mirrors, each of which can be independently and rapidly deflected between two or more angles. When used in the nonlinear optical neural network system described in this specification, spatial modulation of the light field amplitude can be achieved by programming the micromirror array (e.g., recording light reflected into the optical path as "1" and light reflected onto an aperture or absorber as "0"). When using reflective devices, as described in this specification... Figure 1 or Figure 5 The transmissive optical path layout shown needs to be adjusted accordingly. For example, the light beam needs to be incident on the surface of the digital micromirror device at a certain angle, and then the reflected light is collected to enter the next optical path link. This usually requires a folded optical path design.

[0033] As an alternative implementation, particularly for a pre-trained, task-specific system, the multilayer spatial light modulator, as a diffraction modulation component, can be replaced by a set of fixed, custom-manufactured multilayer diffractive optical elements (DOEs). A diffractive optical element is a device that statically modulates the phase of a light wave through surface microstructures. Once the weights of a diffractive neural network are configured as described above... Figure 1 or Figure 5Once the training or computer simulation training of the nonlinear optical neural network system shown is determined, these optimized phase distributions can be solidified onto a multilayer transparent substrate (such as quartz glass or polymer) using micro-nano fabrication techniques such as photolithography and electron beam direct writing, forming a set of multilayer diffractive optical elements. Figure 1 or Figure 5 Compared to programmable spatial light modulators such as the multilayer spatial light modulator shown, this fixed diffractive optical element has significant advantages: lower cost (especially in mass production), near-zero power consumption (passive device), more stable optical performance, and response speed limited only by the speed of light. In this case, the system still retains its powerful nonlinear computing capabilities, but its network weights become a non-programmable "read-only" mode. The spatial light modulator can still be a programmable spatial light modulator or a digital micromirror device to maintain the system's ability to load different input data, forming a hybrid system with programmable input and fixed computation.

[0034] Therefore, this embodiment demonstrates that the "spatial light modulation device" and "diffraction control component" defined in this specification should be understood as functional limitations. Their specific physical implementation is not limited to a transmissive liquid crystal spatial light modulator, but can also be a reflective digital micromirror device, a solidified diffractive optical element, or an acousto-optic / electro-optic modulator array, etc. As long as the device can achieve spatial modulation of the light field or perform preset spatial transformation calculations, it falls within the protection scope of this specification.

[0035] The following will combine Figure 2 The schematic diagram of the optical propagation process in each cycle of the optical cavity illustrates the working process of the system in this embodiment in more detail. A collimated coherent beam with a uniform initial complex amplitude is incident from the outside onto the input coupling component, such as the beam splitter 101. A portion of the light energy is coupled into the resonant cavity, reflected by the mirror 102, and reaches the single-layer spatial light modulator 103. First, some definitions are made: the incident plane wave beam is defined as... The reflections of beam splitter 101 and beam splitter 106 are respectively set as and The spatial light field information input of the single-layer spatial light modulator 103 is The spatial light field modulation of the multilayer spatial light modulator 104 is as follows: ,in This represents the number of layers in a multi-layer spatial light modulator. Light field propagation and diffraction in free space are used... It means that, among them It represents the distance the light field travels in free space.

[0036] like Figure 2 As shown, in the first loop ( In this process, the beam first passes through an input single-layer spatial light modulator 103, where its complex amplitude is multiplied by a spatial modulation function representing the input information. (For example, a two-dimensional matrix). Subsequently, the beam carrying the input information propagates to the multilayer spatial light modulator 104 and is subjected to a spatial transformation operator. The transformed beam is When the beam is reflected by mirror 105 and reaches beam splitter 106, a small portion of the beam is transmitted out, forming the output component of the first cycle. , represented as: (1) As a plane wave with no spatial information distribution, it propagates without diffraction, and there is no need to consider its diffraction before reaching the single-layer spatial light modulator 103. The distance between the single-layer spatial light modulator 103 and the multi-layer spatial light modulator 104. The distance between the multilayer spatial light modulator 104 and the output detector is denoted as .

[0037] set up (2) (3) Formula (1) can be presented as: (4) Meanwhile, at beam splitter 106, most of the beam is reflected back to the resonant cavity, and then returns to the front of the single-layer spatial light modulator 103 via mirrors 7 and 2, ready to begin the second cycle. Importantly, when the beam begins its second cycle, its complex amplitude already carries the results of the first cycle's calculations. Considering output coupling, this beam, after being transmitted through beam splitters 106 and 101, passes again through the single-layer spatial light modulator 103, where its complex amplitude is multiplied again by the input modulation function x. Then, the beam passes through the multi-layer spatial light modulator 104 again, where the transformation operator Tn is applied. At beam splitter 106, a portion of this beam is transmitted, forming the output component of the second cycle. See formula (5).

[0038] (5) in, To ensure that the light field reaches the distance of the single-layer spatial light modulator 103 again after passing through the multi-layer spatial light modulator 104 and undergoing a cycle, the following settings are made: (6) Formula (5) becomes: (7) This component contains information related to X. 2 The relevant term is the quadratic nonlinear term. Cycle 3, which refers to the output of the light field circulating three times within the optical cavity, can be expressed as:

[0039] (8) After simplification, it becomes: (9) And so on, loop m That is, the light field circulates within the optical cavity. m The output of the loop can be expressed as: (10) Figure 2 This process is vividly illustrated. It can be seen that as the light field continuously circulates within the optical resonant cavity, it receives spatial light field information input with each cycle. and spatial light field modulation , producing higher order Item and This term generates higher-order nonlinear effects. The final optical field output can be expressed as: (11) Ultimately, the total output light field received on the detector plane outside the system (e.g., a two-dimensional imaging detector, such as a charge-coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) image sensor) is... It is a coherent superposition of the output components generated by all cycles m. Since the optical path difference of each cycle is an integer multiple of the wavelength (satisfying the resonance condition), this superposition is coherent, i.e., a vector sum of the electric fields. The final output optical field contains the output of each cycle within the optical resonant cavity, including outputs from linear first order to nonlinear higher orders (second order to infinite order), with the coefficient ratios between each order determined by the reflectivity of beam splitters 101 and 106. and The decision is made because each order has an output, and there are losses in the cavity devices; the output energy of higher orders is lower than that of lower orders. The total output optical field... It includes first-order, second-order, third-order, and even higher-order terms of X, thus physically implementing a strongly nonlinear mapping function directly and in parallel.

[0040] In this embodiment, the total output light field can be controlled by adjusting the beam splitting ratio (i.e., transmittance) of the beam splitter 106. The convergence rate of each term in the series expansion is used to adjust the intensity and form of the nonlinear response. For example, higher transmittance results in a shorter average lifetime of photons within the cavity, with lower-order terms (e.g., m=1, 2) dominating the total output and leading to weaker nonlinearity; while lower transmittance allows the beam to cycle more times, significantly increasing the contribution of higher-order terms (e.g., m>3) and resulting in stronger nonlinearity. This ability to directly control the form of the nonlinear function at the physical level is a significant advantage of the scheme described in this specification.

[0041] Furthermore, Figure 3 This embodiment demonstrates the basic principle of spatial light field manipulation through a multilayer spatial light modulator 104, forming a multilayer diffraction neural network. According to the Huygens-Fresnel principle, each point on the diffraction layer can be considered a secondary wave source, its amplitude and phase determined by the input wave and the complex-valued transmission or reflection coefficient of that point. Therefore, a point on the diffraction layer (for the spatial light modulator, this means each pixel) can represent a neuron in the optical network, and each neuron connects to other neurons in the next layer through light diffraction, such as... Figure 3 As shown. Free-space diffraction between diffraction layers can be described using Rayleigh-Sommerfeld theory. The subwave diffraction of each neuron on the diffraction layer can be expressed as: (12) in, These are the two spatial dimensions of light wave diffraction. The spatial dimension of the beam propagation; l Indicates the first l Layer diffraction layer (i.e., each optical modulator layer included in a multilayer spatial light modulator). i Indicates that it is located at the th l layer The i One neuron, λ It is the wavelength of the incident wave. , j This represents the imaginary part of a complex number. Therefore, by the first... l The first layer of modulation i one neuron The light field at that point can be represented as: (13) in, It is from the ( l The incident wave from layer -1), M contains the (-1)th layer l -1) All neurons on layer 1. , which represents the complex-valued transmission or reflection coefficient at the neuron. For a pure phase-type multilayer diffraction neural network architecture, It is a constant; it can be adjusted by changing the properties of the pixels in each layer of the multilayer spatial light modulator. This is a learnable network parameter. During network training, this parameter can be iteratively adjusted using gradient descent or backpropagation algorithms.

[0042] To verify the beneficial effects of this embodiment, simulation experiments are also provided in this specification. Please refer to... Figure 4 This figure illustrates the computational simulation results of conventional linear diffraction neural networks with different pixel counts and nonlinear diffraction neural networks based on optical resonators on the MNIST-Handwritten digits and MNIST-Fashion datasets. Existing linear optical diffraction neural networks, due to their inherent linearity, have limited expressive power and struggle to solve complex nonlinearly separable problems. This specification significantly enhances the computational and expressive capabilities of optical neural networks by introducing the nonlinear computational power generated by a recurrent feedback structure.

[0043] In this computational simulation, the wavelength of the light source was 633 nm, the number of diffraction layers was 3, and the horizontal axis represented the total number of pixels per layer. The training set consisted of 50,000 pixels, and the test set consisted of 10,000 pixels. The results show that both networks reached their optimal performance when the number of pixels reached 78,400 (280×280). For both the MNIST-Handwritten digits and MNIST-Fashion datasets, the nonlinear diffraction neural network based on an optical resonator provided in this specification significantly improved performance compared to the traditional linear diffraction neural network at the same network size (same number of diffraction layers and same number of pixels). For the widely used MNIST-Handwritten digits dataset, the recognition accuracy of the nonlinear diffraction neural network (93.4%) was 10.9% higher than that of the linear neural network (82.5%). For the more challenging MNIST-Fashion dataset, the recognition accuracy of the nonlinear diffraction neural network (79.1%) was 5.3% higher than that of the linear neural network (73.8%). These quantitative and significant performance improvements strongly demonstrate the superiority of the solution described in this specification.

[0044] Furthermore, the system in this embodiment may also include a two-dimensional imaging detector for detecting the light field distribution of the output beam. This detector is placed after the transmission optical path of the output coupling component, with its photosensitive surface located on the system's output focal plane or image plane, to capture the final output light spot pattern carrying the calculation results. By analyzing the light intensity of different regions on the detector, classification results or other calculation results can be obtained. For example, in a classification task, 10 separate target regions can be preset on the detector plane, each corresponding to one of 10 different categories. The region receiving the largest integrated light intensity is determined to belong to that category.

[0045] As another embodiment, the technical solution proposed in this specification is not limited to a specific optical feedback structure geometry. Please refer to... Figure 5 This figure is a schematic diagram of another nonlinear optical neural network system based on a four-mirror resonator provided in an embodiment of this specification. This embodiment aims to illustrate the structural flexibility of the solution in this specification.

[0046] Unlike the triangular resonant cavity in Embodiment 1, the optical feedback structure in this embodiment consists of four highly reflective mirrors: mirror 502, mirror 505, mirror 507, and mirror 508. These four mirrors together form a stable, nearly rectangular closed loop optical path, sometimes referred to as a "bowtie" shaped or square cavity.

[0047] In this embodiment, the functions, types, and configurations of the input coupling component, output coupling component, spatial light modulation device, and diffraction control component in the system are similar to the corresponding components in the aforementioned embodiments. Specifically, the input coupling component, namely the beam splitter 501, is used to couple the input beam into the circulating optical path, and the output coupling component, namely the beam splitter 506, is used to couple a portion of the beam from the circulating optical path as the output.

[0048] Similarly, spatial light modulation devices (such as...) Figure 5 The single-layer optical modulator 503 shown) and diffraction control components (such as...) Figure 5 The multilayer optical modulators 50 shown are all arranged in series inside the loop optical path formed by the four-mirror resonant cavity. After entering through the beam splitter 501, the beam is guided by the reflector 502 to the single-layer optical modulator 503, where input information is loaded. Subsequently, the beam passes through the multilayer optical modulator 504 to perform network operations. Afterward, the beam reaches the beam splitter 506 through the reflector 505. Part of the light is transmitted out as output, while the other part is reflected and reflected in sequence by the reflectors 507 and 508, finally returning to a position after the beam splitter 501 and before the reflector 502, thus completing one complete cycle and starting the next cycle.

[0049] Therefore, although the geometry of the resonant cavity and the number of mirrors have changed, the core technical feature of this specification, namely, "the spatial light modulation device and the diffraction control component are both disposed in the cyclic optical path, and their configuration is such that the beam passes through the spatial light modulation device and the diffraction control component sequentially in each cycle," is completely retained in this embodiment. Therefore, the physical principle of generating the nonlinear response is the same as in the previous embodiment: through the cyclic iteration of the beam, a series expansion containing higher-order terms of the input information is formed by coherent superposition at the output end. This embodiment shows that as long as a stable closed optical path that allows the beam to propagate cyclically can be formed, whether using a polygonal resonant cavity composed of three, four, or more mirrors, or other forms of optical feedback structures (such as fiber optic ring cavities), the technical solution of this specification can be implemented. The specific geometry can be optimized based on considerations such as system compactness, stability, and ease of alignment.

[0050] Another embodiment of this specification provides a method for implementing nonlinear optical calculations, which can be executed using the system provided in any of the above embodiments. Please refer to... Figure 6 The flowchart shown illustrates the method. The method includes the following steps: Step S602: Provide an external coherent beam and couple a portion of its energy into the loop optical path provided in the aforementioned embodiments.

[0051] Step S604: During each cycle of the light beam's propagation, firstly, based on the input information, the light beam is spatially modulated using a spatial light modulation device located in the cyclic optical path to encode the input information onto the complex amplitude of the optical field.

[0052] Step S606: In the same loop, using a diffraction control component disposed in the loop optical path and located downstream of the spatial light modulator, diffraction and modulation calculations for simulating neural network layers are performed on the spatially modulated beam.

[0053] Step S608: After the network operation step S606 is completed, when the beam reaches an output coupling point, a portion of its energy is fed back to the starting point of the loop optical path to perform the next loop iteration including steps S604 and S606; at the same time, at the output coupling point, a portion of the beam is coupled out from the loop optical path to form the output beam component of this loop.

[0054] During beam cycling, the output beams can be coherently superimposed in each cycle to generate a nonlinear response. The beam components coupled from different cycles, originating from the same coherent source and with stable optical path differences, undergo coherent superposition (i.e., electric field vector sum) when propagating to the detector plane, forming a stable total output light field distribution that includes nonlinear terms of each order of the input information (such as first-order, second-order, and third-order terms). This directly generates a nonlinear function response with respect to the input information at the physical level.

[0055] Optionally, the method further includes a detection and interpretation step S610: that is, using a two-dimensional imaging detector to detect the light field intensity distribution of the final output beam formed in step S608 after a preset number of cycles, and obtaining the final calculation results, such as classification labels, regression values, etc., by analyzing (e.g., integrating) the light intensity values ​​of a preset area on the detector.

[0056] In summary, this specification, through its innovative structure in which both the spatial light modulation device and the diffraction control component are located inside the optical feedback structure, cleverly utilizes the cyclic iterative physical process of the light beam to achieve efficient and controllable nonlinear computation in a purely optical system. This overcomes the limitations of linear computation in existing optical neural networks and provides a novel and highly feasible technical solution for the development of high-performance, low-power, and high-speed all-optical neural networks and photonic processors.

[0057] The above description is merely a preferred embodiment of this specification and is not intended to limit this specification. Those skilled in the art can make various changes, modifications, substitutions, and variations to these embodiments without departing from the spirit and principles of this specification. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of protection of this specification.

[0058] What those skilled in the art will understand is: In this specification, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitation, the presence of additional identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded.

[0059] In this specification, “a,” “an,” and “the” do not specifically refer to the singular, but may also include the plural.

[0060] In this specification, ordinal numbers such as "first," "second," etc., do not necessarily indicate order; they are often used to distinguish between objects. For example, "first server" and "second server" usually refer to two servers. To differentiate between these two servers, they are described as "first server" and "second server." Of course, sometimes these two servers may be the same server.

[0061] In this specification, unless explicitly stated otherwise, "receiving and sending data" does not necessarily mean direct receiving and sending; it can also mean indirect receiving and sending. For example, A receiving data sent by B can be understood as A directly receiving the data sent by B, or it can be understood as A indirectly receiving the data sent by B through other entities such as C. Similarly, B sending data to A can be understood as B sending the data directly to A, or it can be understood as B indirectly sending the data to A through other entities such as C. Here, C can be one entity, or it can be two or more entities.

[0062] In this specification, unless explicitly stated otherwise, the relationships between structures can be direct or indirect. For example, when describing "A is connected to B," unless it is explicitly stated that A and B are directly connected, it should be understood that A can be directly connected to B or indirectly connected to B. Similarly, when describing "A is on top of B," unless it is explicitly stated that A is directly above B (AB is adjacent and A is above B), it should be understood that A can be directly above B or indirectly above B (AB is separated by other elements, and A is above B). And so on.

[0063] This specification uses specific terms to describe embodiments thereof. For example, "one embodiment" and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that "one embodiment" or "an alternative embodiment" mentioned twice or more in different places in this specification do not necessarily refer to the same embodiment. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples, without contradiction.

[0064] Although one or more embodiments of this specification provide method steps as described in the embodiments or flowcharts, it is understood that the order of steps listed in the embodiments or flowcharts is only one of many possible execution orders and does not represent the only execution order. Therefore, when the claims involve method steps, any changes or adjustments to the order of such steps, or the parallelism between steps, are also within the scope of protection of the claims.

Claims

1. A nonlinear optical neural network system, characterized in that, include: An optical feedback structure is used to form a closed loop optical path so that the light beam propagates cyclically in the closed loop optical path; An input coupling component and an output coupling component are respectively used to couple an input beam to the circulating optical path and to couple a portion of the beam out from the circulating optical path; wherein, the input beam includes a coherent beam; A spatial light modulation device for controlling at least one of the amplitude and phase of an input light beam; A diffraction control component for performing multi-layer diffraction operations and control on its input beam; The spatial light modulator and the diffraction control component are both disposed in the closed loop optical path. The configuration is such that the light beam passes through the spatial light modulator and the diffraction control component sequentially in each loop, and the light beam reaches the output coupling component after passing through the diffraction control component. The spatial light modulation device includes a single-layer spatial light modulator (SLM), and the diffraction control component includes a multi-layer spatial light modulator (SLM) arranged along the light propagation direction; the spatial light modulator (SLM) includes a plurality of two-dimensionally distributed, programmable controllable pixel units.

2. The system according to claim 1, characterized in that, The optical feedback structure includes an optical resonant cavity composed of at least three mirrors.

3. The system according to claim 1, characterized in that, The input coupling component and the output coupling component include a beam splitter.

4. The system according to claim 1, 2, or 3, characterized in that, The spatial light modulation device includes a reflective digital micromirror device.

5. The system according to claim 1, 2, or 3, characterized in that, The diffraction modulation assembly includes a multilayer diffractive optical element (DOE) with a pre-processed phase distribution to statically modulate the input beam according to a pre-processed modulation mode.

6. The system according to claim 1, characterized in that, The system also includes a two-dimensional imaging detector for detecting the output light field distribution.

7. A method for realizing nonlinear optical calculations, characterized in that, An application is made in a nonlinear optical neural network system, the system comprising an optical feedback structure, an input coupling component, an output coupling component, a spatial light modulator, and a diffraction control component; the optical feedback structure is used to form a closed loop optical path to allow a light beam to propagate cyclically within the closed loop optical path; the spatial light modulator and the diffraction control component are both disposed within the closed loop optical path, configured such that the light beam passes sequentially through the spatial light modulator and the diffraction control component in each loop, and the light beam reaches the output coupling component after passing through the diffraction control component; the spatial light modulator includes a single-layer spatial light modulator (SLM), and the diffraction control component includes multiple layers of spatial light modulators (SLMs) arranged along the light propagation direction; the spatial light modulator (SLM) includes multiple two-dimensionally distributed, programmable controllable pixel units; The method includes the following steps: The input coupling component is used to couple at least a portion of the input beam into the loop optical path, the input beam comprising a coherent beam; The light beam is made to propagate cyclically in the circular optical path, and in each cycle, the following steps are performed sequentially: The spatial light modulation device is used to programmably spatially control at least one of the amplitude and phase of the input light beam; and Using the diffraction control component, multi-layer diffraction operations and control are performed on the spatially modulated beam; At least a portion of the beam that has undergone the multi-layer diffraction calculations and modulation is coupled out of the loop optical path.

8. The method according to claim 7, characterized in that, It also includes: using a two-dimensional imaging detector to detect the output beam coupled out of the cyclic optical path, wherein the output beam is a coherent superposition of beams of various orders coupled out after the beam has passed through the cyclic optical path a preset number of cycles.

9. The method according to claim 7 or 8, characterized in that, The spatial modulation step and the spatial transformation operation step are executed by a programmable spatial light modulator.

Images