An optical neural network nonlinear activation system, method and apparatus

Abstract

The application discloses an optical neural network nonlinear activation system, method and device, relates to the technical field of optical neural networks, and because a magneto-optical material is integrated in a phase adjustment waveguide arm of a nonlinear activation unit, a controller can control the magneto-optical material of the phase adjustment waveguide arm to generate a magneto-optical effect according to a current activation function type and weights of the optical neural network, so as to perform phase adjustment on an optical signal transmitted through the phase adjustment waveguide arm, and make the nonlinear activation unit adapt to the current activation function type and the weights, perform corresponding nonlinear calculation on a target input signal, and realize dynamic switching of multiple nonlinear activation functions, thereby improving the flexibility of the optical neural network.

Description

Technical Field

[0001] This application relates to the field of optical neural network technology, and in particular to an optical neural network nonlinear activation system, method and apparatus. Background Technology

[0002] With the exponential growth in computing power demands from artificial intelligence, optical neural networks (ONNs) have attracted attention due to their high parallelism, ultra-low transmission loss, and ultra-wide bandwidth characteristics in optical computing. As the core architecture of photonic computing, ONNs need to simultaneously perform linear operations (such as matrix multiplication) and nonlinear activation functions within the optical domain. Furthermore, their performance is highly dependent on the efficiency and reliability of the nonlinear activation units. Therefore, how the nonlinear activation units in ONNs achieve nonlinear computation has become a key research focus.

[0003] In related technologies, the nonlinear activation units of optical neural networks mainly include electro-optical modulators (EOM), electro-absorption modulators (EAM), thermo-optic modulators, semiconductor optical amplifiers, and microring resonators (MRRs). However, current nonlinear activation units can only perform single-type nonlinear calculations and cannot flexibly switch between multiple nonlinear activation functions, thus reducing the flexibility of optical neural networks. Summary of the Invention

[0004] This application provides an optical neural network nonlinear activation system, method, and apparatus to at least solve the problem that nonlinear activation units in related technologies can only perform a single type of nonlinear calculation, and cannot flexibly achieve dynamic switching of multiple nonlinear activation functions, thus reducing the flexibility of optical neural networks.

[0005] This application provides an optical neural network nonlinear activation system, including: multiple nonlinear activation units and a controller. The nonlinear activation units include a beam splitter and a beam combiner. The beam splitter and beam combiner are connected by multiple waveguide arms. The multiple waveguide arms include a phase-adjusting waveguide arm and a conventional waveguide arm. The phase-adjusting waveguide arm integrates a magneto-optical material.

[0006] The controller is used to control the magneto-optical material of the phase-adjusting waveguide arm to generate a magneto-optical effect according to the current activation function type and weight of the optical neural network.

[0007] The beam splitter is used to receive the target input signal of the optical neural network, split the target input signal into a corresponding number of optical signals, and then transmit them to the beam combiner through a phase-adjustable waveguide arm and a conventional waveguide arm, respectively. The magneto-optical material of the phase-adjustable waveguide arm adjusts the phase of the optical signal transmitted through the phase-adjustable waveguide arm based on the generated magneto-optical effect, so that the phase-adjustable waveguide arm transmits the phase-adjusted optical signal as the target optical signal to the beam combiner.

[0008] The beam combiner is used to perform signal interference on the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm, and then output the nonlinear activation result of the target input signal.

[0009] This application also provides a nonlinear activation method for optical neural networks, including:

[0010] Obtain the current activation function type and weights of the optical neural network;

[0011] Based on the current activation function type and weights of the optical neural network, the magneto-optical material of the phase-adjusting waveguide arm of the nonlinear activation unit is controlled to generate a magneto-optical effect. This causes the beam splitter of the nonlinear activation unit to split the received target input signal of the optical neural network into a corresponding number of optical signals, which are then transmitted to the beam combiner through the phase-adjusting waveguide arm and the conventional waveguide arm, respectively. The magneto-optical material of the phase-adjusting waveguide arm, based on the generated magneto-optical effect, performs phase adjustment on the optical signals transmitted through the phase-adjusting waveguide arm, so that the phase-adjusting waveguide arm transmits the phase-adjusted optical signals as the target optical signals to the beam combiner.

[0012] Based on the beam combiner, the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm are interfered and output to obtain and output the nonlinear activation result of the target input signal.

[0013] This application also provides an optical neural network nonlinear activation device, comprising:

[0014] The acquisition module is used to obtain the current activation function type and weights of the optical neural network;

[0015] The control module is used to control the magneto-optical material of the phase-adjusting waveguide arm of the nonlinear activation unit to generate a magneto-optical effect according to the current activation function type and weight of the optical neural network. This causes the beam splitter of the nonlinear activation unit to split the received target input signal of the optical neural network into a corresponding number of optical signals, which are then transmitted to the beam combiner through the phase-adjusting waveguide arm and the conventional waveguide arm, respectively. The magneto-optical material of the phase-adjusting waveguide arm adjusts the phase of the optical signals transmitted through the phase-adjusting waveguide arm based on the generated magneto-optical effect, so that the phase-adjusting waveguide arm transmits the phase-adjusted optical signals as the target optical signals to the beam combiner.

[0016] The output module is used to perform signal interference on the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm based on the beam combiner, and then output the nonlinear activation result of the target input signal.

[0017] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for implementing the steps of any of the above-described optical neural network nonlinear activation methods when executing the computer program.

[0018] This application also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the above-described optical neural network nonlinear activation methods.

[0019] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described optical neural network nonlinear activation methods.

[0020] Through this application, since the phase-adjusting waveguide arm of the nonlinear activation unit integrates magneto-optical material, the controller can control the magneto-optical material of the phase-adjusting waveguide arm to generate a magneto-optical effect according to the current activation function type and weight of the optical neural network, so as to perform phase adjustment on the optical signal transmitted through the phase-adjusting waveguide arm. This allows the nonlinear activation unit to adapt to the current activation function type and weight, and perform corresponding nonlinear calculations on the target input signal, so as to realize the dynamic switching of multiple nonlinear activation functions and improve the flexibility of the optical neural network. Attached Figure Description

[0021] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the structure of the optical neural network nonlinear activation system provided in the embodiments of this application;

[0023] Figure 2 A schematic diagram of an exemplary optical neural network provided for embodiments of this application;

[0024] Figure 3 A schematic diagram of the structure of an exemplary nonlinear activation unit provided in an embodiment of this application;

[0025] Figure 4 A flowchart illustrating the nonlinear activation method for optical neural networks provided in this application embodiment;

[0026] Figure 5 This is a schematic diagram of the structure of the optical neural network nonlinear activation device provided in the embodiments of this application;

[0027] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0029] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, 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, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0030] With the exponential growth in computing power demands from artificial intelligence, traditional electronic chips, constrained by the slowdown of Moore's Law and the limitations of the von Neumann architecture, struggle to support the computational needs of complex models (such as large language models and real-time image processing). Photonic computing, with its high parallelism, ultra-low transmission loss, and ultra-wide bandwidth characteristics of optical signals, is considered one of the key technologies to overcome computing power bottlenecks. Optical Neural Networks (ONNs), as the core architecture of photonic computing, need to simultaneously perform linear operations (such as matrix multiplication) and nonlinear activation functions within the optical domain, and their performance is highly dependent on the efficiency and reliability of the nonlinear activation units. However, existing implementations of nonlinear activation units still face certain technical challenges in terms of speed, energy consumption, and integration, limiting the development and application of optical neural networks.

[0031] Existing optical neural network (OEO) nonlinear activation unit (NOU) technologies mainly include electro-optical modulators (EOMs) and electro-absorption modulators (EAMs), thermo-optical modulators, semiconductor optical amplifiers, microring resonators (MRRs), and solutions based on optical nonlinear effects and phase change materials. EOMs and EAMs achieve light intensity modulation by controlling the refractive index or absorption coefficient of the material through an electric field, offering advantages such as mature technology and fast response speed. However, they suffer from high energy consumption, speed limitations, and integration difficulties because OEO systems require a continuous driving voltage, and the photoelectric conversion process introduces additional power consumption. Thermo-optical modulators utilize temperature changes to induce changes in the refractive index of the material, achieving phase modulation. While structurally simple, they suffer from slow response speed and severe thermal crosstalk. Semiconductor optical amplifiers utilize gain saturation effects to achieve nonlinear modulation, offering the advantage of all-optical operation, but suffer from signal distortion and poor stability. MRRs achieve nonlinear modulation by changing the resonance conditions of the microring, offering a compact structure but being temperature-sensitive and having low process tolerance. Schemes based on optical nonlinear effects and phase change materials each have their own advantages and disadvantages. For example, the Kerr effect has a high threshold light intensity, the saturation absorption response speed is limited, the phase matching conditions for optical parametric processes are stringent, and the phase change materials have slow switching speeds and limited cycle lifetimes. In addition, most existing schemes can only achieve single nonlinear functions, making it difficult to flexibly implement dynamic switching of multiple activation functions (such as Sigmoid and ReLU), lacking dynamic tunability, and making it difficult to adapt to diverse AI tasks.

[0032] To address the aforementioned technical problems, this application provides an optical neural network nonlinear activation system, method, and apparatus. The system includes multiple nonlinear activation units and a controller. Each nonlinear activation unit includes a beam splitter and a beam combiner, connected by multiple waveguide arms. These waveguide arms include phase-adjustable waveguide arms and conventional waveguide arms, with magneto-optical material integrated into the phase-adjustable waveguide arms. The controller controls the magneto-optical material of the phase-adjustable waveguide arms to generate a magneto-optical effect based on the current activation function type and weights of the optical neural network. The beam splitter receives the target input signal of the optical neural network. The target input signal is split into a corresponding number of optical signals, which are then transmitted to a beam combiner via a phase-adjustable waveguide arm and a conventional waveguide arm, respectively. The magneto-optical material of the phase-adjustable waveguide arm, based on the generated magneto-optical effect, adjusts the phase of the optical signals transmitted through the phase-adjustable waveguide arm, so that the phase-adjustable waveguide arm transmits the phase-adjusted optical signals as the target optical signals to the beam combiner. The beam combiner is used to perform signal interference between the target optical signals transmitted through the phase-adjustable waveguide arm and the optical signals transmitted through the conventional waveguide arm, and then outputs the nonlinear activation result of the target input signal. In the system provided by the above scheme, because the phase-adjustable waveguide arm of the nonlinear activation unit integrates magneto-optical material, the controller can control the magneto-optical material of the phase-adjustable waveguide arm to generate a magneto-optical effect according to the current activation function type and weights of the optical neural network, thereby adjusting the phase of the optical signals transmitted through the phase-adjustable waveguide arm. This allows the nonlinear activation unit to adapt to the current activation function type and weights, performing corresponding nonlinear calculations on the target input signal, achieving dynamic switching of multiple nonlinear activation functions, and improving the flexibility of the optical neural network.

[0033] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] This application provides an optical neural network nonlinear activation system for calculating nonlinear activation functions in an optical neural network.

[0035] like Figure 1 The diagram shown is a structural schematic of an optical neural network nonlinear activation system provided in an embodiment of this application. The system includes: multiple nonlinear activation units and a controller. The nonlinear activation units include beam splitters and beam combiners. The beam splitters and beam combiners are connected by multiple waveguide arms. Among the multiple waveguide arms are phase-adjusting waveguide arms and conventional waveguide arms. The phase-adjusting waveguide arms integrate magneto-optical materials.

[0036] The controller controls the magneto-optical material of the phase-adjustable waveguide arm to generate a magneto-optical effect based on the current activation function type and weights of the optical neural network. The beam splitter receives the target input signal of the optical neural network, splits the target input signal into a corresponding number of optical signals, and transmits them to the beam combiner through the phase-adjustable waveguide arm and the conventional waveguide arm, respectively. The magneto-optical material of the phase-adjustable waveguide arm adjusts the phase of the optical signal transmitted through the phase-adjustable waveguide arm based on the generated magneto-optical effect, so that the phase-adjustable waveguide arm transmits the phase-adjusted optical signal as the target optical signal to the beam combiner. The beam combiner performs signal interference on the target optical signal transmitted through the phase-adjustable waveguide arm and the optical signal transmitted through the conventional waveguide arm, and outputs the result to obtain and output the nonlinear activation result of the target input signal.

[0037] Among them, the activation function type is a function that defines a nonlinear mapping relationship in an optical neural network (such as Sigmoid, ReLU, etc.). Different functions correspond to different signal processing logics and determine the characteristics of the nonlinear activation result. The weight is a quantitative parameter in the neural network that quantifies the importance of the input signal and affects the output intensity of the activation result.

[0038] It should be noted that optical neural networks, which use optical signals as information carriers and perform calculations within the optical domain, have advantages over traditional electronic neural networks, such as high parallelism, low transmission loss, and high bandwidth. Figure 2 As shown in the schematic diagram of the exemplary optical neural network provided in this application embodiment, the optical neural network includes multiple network layer structures. Each network layer consists of optical interference units and optical nonlinear units. Multiple optical interference units constitute an optical interference layer, and multiple optical nonlinear units constitute a nonlinear activation layer. The optical nonlinear unit is the same as the nonlinear activation unit. Both the optical interference unit and the optical nonlinear unit adopt a Mach-Zehnder interferometer (MZI). The target input signal received by any nonlinear activation unit is the linearly calculated input signal output by the optical interference unit.

[0039] Specifically, in one embodiment, the system provided in this application includes multiple nonlinear activation layers, each of which includes multiple nonlinear activation units. After obtaining a nonlinear activation result, any nonlinear activation unit uses the nonlinear activation result as a new input signal and inputs it to the next adjacent optical interference layer. This optical interference layer outputs a new target input signal to the beam splitter of the nonlinear activation unit in the next nonlinear activation layer, until all nonlinear activation units in the last nonlinear activation layer output nonlinear activation results. The nonlinear activation results output by all nonlinear activation units in the last nonlinear activation layer are used as the target nonlinear activation result of the optical neural network.

[0040] Specifically, optical neural networks utilize the optical interference effect of MZI structures to achieve weighted summation and activation functions. Each MZI unit simulates a neuron, and input data is mapped onto the MZI array using methods such as matrix decomposition and optical Fourier transform. The weights and biases of each MZI unit are adjusted by an external magnetic field and an internal phase modulator to achieve weighted summation and nonlinear activation of the data. Cascaded MZI structures can realize multi-layer neural networks and improve computational speed and efficiency by utilizing the parallel transmission characteristics of optical signals.

[0041] It's important to further explain that a beam splitter is an optical signal distribution element that breaks down the target input signal into multiple optical signals, ensuring that each signal can be processed by a different waveguide arm. A beam combiner is an optical element that complements the beam splitter, re-converging the multiple signals transmitted through different waveguide arms to prepare for subsequent signal interference. Waveguide arms are the channels for optical signal transmission. Conventional waveguide arms only handle optical signal transmission without altering core characteristics such as signal phase and intensity, serving as the reference channel for phase adjustment. Phase-adjusting waveguide arms are special waveguide arms integrating magneto-optical materials, capable of changing the phase of the optical signal through the magneto-optical effect, providing a channel for dynamic adjustment. Magneto-optical materials are materials whose optical properties (such as refractive index) change under the influence of a magnetic field, allowing for precise adjustment of the optical signal phase through magnetic field control.

[0042] Specifically, the beam splitter first receives the target input signal from the optical neural network, then splits the signal into optical signals matching the number of waveguide arms, ensuring that each waveguide arm has a corresponding signal input. The controller generates control commands based on the activation function type and weights required by the current optical neural network, causing the magneto-optical material in the phase-adjustable waveguide arm to produce the corresponding magneto-optical effect. That is, the magneto-optical material in the phase-adjustable waveguide arm generates a magneto-optical effect under the controller's control, thereby changing its optical properties. When the split signal passes through the waveguide arm, its phase is adjusted to obtain the target optical signal, while the phase of the signal transmitted by the conventional waveguide arm remains unchanged, maintaining its original state. The beam combiner converges the two signals. Due to the phase difference between the two signals, optical interference occurs. The interfered signal is the nonlinear activation result of the target input signal, and is ultimately output to the next-level module of the optical neural network, i.e., to the optical interference unit of the next network layer.

[0043] This application embodiment dynamically adjusts the signal phase of the phase-adjusting waveguide arm according to the current activation function type and weight to adapt to the nonlinear characteristics of different activation functions, allowing the same nonlinear activation unit to meet multiple nonlinear activation requirements and improving the flexibility of the optical neural network.

[0044] Specifically, in one embodiment, an electromagnetic coil is integrated around the phase-adjusting waveguide arm; the magnetic field strength of the electromagnetic coil acts on the magneto-optical material of the phase-adjusting waveguide arm to cause the magneto-optical material to produce a magneto-optical effect.

[0045] Specifically, when current flows through the electromagnetic coil, a magnetic field is generated around it according to the principle of electromagnetic induction. The strength and other characteristics of the magnetic field can be controlled by adjusting parameters such as the current in the coil. The magnetic field generated by the electromagnetic coil integrated around the phase-adjustable waveguide arm acts on the magneto-optical material within the waveguide arm. Under the influence of the magnetic field, the optical properties of the magneto-optical material change, thus producing a magneto-optical effect. This magneto-optical effect allows the optical parameters, such as the phase, of the optical signal passing through the phase-adjustable waveguide arm to be adjusted, providing a foundation for subsequent optical neural networks to achieve functions such as nonlinear activation.

[0046] Among them, the magnetic field is generated by an electromagnetic coil, and the parameters such as the magnetic field strength can be flexibly adjusted by controlling the current of the coil, which provides a guarantee for the regulation of the magneto-optical effect.

[0047] Specifically, in one embodiment, the controller can control the magnetic field strength of the electromagnetic coil based on the current activation function type and weights of the optical neural network.

[0048] Specifically, the controller generates corresponding control logic based on the activation function type and weight, determines the target value of the magnetic field strength that the electromagnetic coil needs to achieve, and then controls the electromagnetic coil to generate a magnetic field strength that meets the target value. By applying this magnetic field strength to the magneto-optical material of the phase-adjusting waveguide arm, the phase adjustment amplitude of the optical signal transmitted through the phase-adjusting waveguide arm is matched with the nonlinear activation requirements of the optical neural network.

[0049] Specifically, in one embodiment, the controller can adjust the drive current of the electromagnetic coil according to the current activation function type and weight of the optical neural network to control the magnetic field strength of the electromagnetic coil.

[0050] It should be noted that the controller in this embodiment can be implemented using a field programmable gate array (FPGA), and the magnetic field strength of the electromagnetic coil changes with the driving current.

[0051] Specifically, the key to achieving dynamically adjustable nonlinear activation in this embodiment lies in flexibly switching activation function types and parameters through magnetic field programming using FPGA. At the hardware level, a miniature electromagnetic coil is integrated around the magneto-optical waveguide, and the magnetic field strength B is controlled by adjusting the driving current. At the software level, a dynamic programming interface is designed based on FPGA or microcontroller to control the electromagnetic coil current, thereby enabling real-time switching of the activation function type of the nonlinear unit. For example, the hidden layer of an optical neural network uses the ReLU activation function, while the output layer uses the Sigmoid activation function. The slope and threshold of the activation function can be dynamically adjusted through the magnetic field gradient. For example, increasing the rate of change of the magnetic field strength B can increase the steepness of the Sigmoid curve, while limiting the maximum value of the magnetic field strength B can set the saturation threshold of ReLU.

[0052] Specifically, in one embodiment, the controller can determine the driving current adjustment strategy of the electromagnetic coil according to the current activation function type of the optical neural network; and adjust the driving current of the electromagnetic coil according to the weights of the optical neural network according to the driving current adjustment strategy of the electromagnetic coil.

[0053] The drive current regulation strategy can be a current regulation rule preset by the controller that matches different activation function types, including at least the current regulation range.

[0054] Specifically, in one embodiment, the controller can determine the target phase difference increase range between the phase-adjustable waveguide arm and the conventional waveguide arm based on the current activation function type of the optical neural network; and determine the driving current adjustment strategy of the electromagnetic coil based on the target phase difference increase range between the phase-adjustable waveguide arm and the conventional waveguide arm, and the relationship function between the target phase difference increase range and the change in the optical signal refractive index of the phase-adjustable waveguide arm.

[0055] It should be noted that different activation function types correspond to different nonlinear output characteristics, which are determined by the phase difference between the two waveguide arm signals. For the ReLU function, a lower limit for the phase difference increase range needs to be set; when the phase difference is below this value, the output after interference is 0. For the Sigmoid function, a wide and continuous phase difference increase range needs to be set to ensure that the interference output exhibits a gradual curve between 0 and 1. Therefore, the controller can directly determine the target phase difference increase range based on the current activation function type. For example, the target phase difference increase range for the ReLU function includes an increase from π to 3π / 2, meaning the light intensity increases linearly with the magnetic field strength B. The target phase difference increase range for the ReLU function includes an increase from 3π / 2 to 2π, where the output light intensity transitions from linear growth to the saturation region, simulating the Sigmoid function.

[0056] The relationship between the range of increase in the target phase difference and the change in the refractive index of the optical signal in the phase-adjusting waveguide arm is as follows:

[0057]

[0058] in, Indicates the range of increase in target phase difference. Represents pi (π). Indicates the wavelength of the target input signal. This represents the change in the refractive index of the optical signal in the phase-adjustable waveguide arm. This indicates the length of the magneto-optical waveguide.

[0059] Specifically, in one embodiment, the optical signal refractive index of the phase-adjusting waveguide arm changes according to the magnetic field strength.

[0060] The relationship between the change in the optical signal refractive index and the magnetic field strength in the phase-adjustable waveguide arm is as follows:

[0061]

[0062] in, This represents the change in the refractive index of the optical signal in the phase-adjustable waveguide arm. This represents a pre-defined constant, such as the Field constant. It represents the magnetic field strength.

[0063] Specifically, in one embodiment, the controller can also control the rate of change of the magnetic field strength of the electromagnetic coil with respect to the weights by adjusting the driving current of the electromagnetic coil based on the slope of the current activation function of the optical neural network.

[0064] Here, the slope of the current activation function represents the rate at which the output value of the activation function changes with the input value, and the rate at which the magnetic field strength changes with the weights represents the rate at which the magnetic field strength of the electromagnetic coil changes as the weights of the optical neural network increase or decrease.

[0065] Specifically, by binding the rate of change of magnetic field strength to the slope of the activation function, the change of phase difference of the optical signal is made to match the output requirements of the function, avoiding distortion of activation results caused by rate mismatch and improving the accuracy of nonlinear calculation.

[0066] Accordingly, in one embodiment, the controller can also limit the maximum value of the magnetic field strength of the electromagnetic coil by limiting the driving current of the electromagnetic coil, based on the threshold of the current activation function of the optical neural network.

[0067] The threshold of the current activation function is the critical input value that triggers a change in the output characteristics of the activation function.

[0068] Specifically, by limiting the maximum magnetic field strength, the activation result is prevented from exceeding the threshold constraint due to excessive phase difference.

[0069] Specifically, in one embodiment, the beam combiner is used to perform signal interference output on the target optical signal and the optical signal based on the phase difference between the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm, so as to obtain and output the nonlinear activation result of the target input signal.

[0070] Specifically, the variation of interference output intensity with phase difference can be directly mapped to the nonlinear characteristics of different activation functions. By adjusting the phase difference through the controller, the activation function type can be dynamically switched, thus improving the flexibility of the optical neural network.

[0071] Specifically, in one embodiment, the beam combiner determines the optical signal beam combining factor based on the phase difference between the target optical signal and the optical signal, and performs signal interference output on the target optical signal and the optical signal based on the optical signal beam combining factor, so as to obtain and output the nonlinear activation result of the target input signal.

[0072] Among them, such as Figure 3As shown in the schematic diagram of the exemplary nonlinear activation unit provided in this application embodiment, the beam splitter is responsible for splitting the two input optical signals E1 and E2 into multiple signals, which are then sent to different waveguide paths (including phase-adjusting waveguide arms with integrated MO phase shifters and conventional waveguide arms). The MO phase shifter is a magneto-optical material in the phase-adjusting waveguide arm, used to change the phase of the optical signal under the action of an external magnetic field, thereby achieving phase modulation of the target optical signal. The beam combiner is used to combine and interfere with the optical signals transmitted through different paths, ultimately outputting E3 and E4. The refractive index change of the MO material (magneto-optical material) depends only on the external magnetic field, resisting temperature drift and process fluctuations, which can greatly improve output stability. Through optimized design of the Faraday rotation coefficient of the MO material, only a μT-level magnetic field is needed to trigger phase modulation, and the input light intensity is as low as the μW level. The phase difference is precisely controlled by the magnetic field strength. It can dynamically switch between various activation functions such as Sigmoid and ReLU, and the slope and threshold can be programmed online. Through a magneto-optical field driving mechanism, it achieves fast response and low-energy modulation, effectively resolving the contradiction between speed and energy efficiency in related technologies. Compared with electro-optic modulators, the magneto-optical modulator can achieve a sub-nanosecond response speed and reduce energy consumption by more than 90%, meeting the real-time information processing requirements of optical neural networks. Simultaneously, through environmentally robust design, the refractive index change of the MO material depends only on the external magnetic field, resisting temperature drift and process fluctuations, ensuring output stability, and further improving the performance and reliability of the ONN.

[0073] The beam combiner determines the nonlinear activation result of the target input signal based on the following formula:

[0074]

[0075] in, This represents the nonlinear activation result of the target input signal. Indicates the target input signal. This represents the phase difference between the target optical signal and the optical signal. This represents the optical signal combining factor.

[0076] Specifically, by adjusting the magnetic field B, precise control can be achieved. The numerical range of, makes It exhibits nonlinear characteristics such as Sigmoid or ReLU.

[0077] Specifically, the light intensity of the input signal can be considered as the input value of the neural network, while the phase difference of the MZI units corresponds to the connection weights. By precisely controlling the phase difference of each MZI unit, complex weighted summation operations can be achieved, which is the core of the forward propagation process of the neural network. Cascaded MZI structures can construct multi-layer neural networks. The output light signal of each layer can serve as the input signal of the next layer. By cascading multiple MZI units, a multi-layer network structure can be realized. In each layer, the output light signal of the MZI unit is processed by a nonlinear activation function before being passed to the next layer. This hierarchical structure enables optical neural networks to handle more complex tasks.

[0078] To enhance the flexibility and adaptability of optical neural networks, dynamically adjustable nonlinear activations are crucial. This means that the network can adjust the parameters of the nonlinear activation function, such as slope, threshold, and shape, through software control without altering the hardware structure. This dynamic adjustment capability is essential for optimizing network performance, adapting to different learning tasks, and responding to environmental changes.

[0079] The optical neural network nonlinear activation system provided in this application includes: multiple nonlinear activation units and a controller. Each nonlinear activation unit includes a beam splitter and a beam combiner, connected by multiple waveguide arms. These waveguide arms include a phase-adjustable waveguide arm and a conventional waveguide arm, with magneto-optical material integrated into the phase-adjustable waveguide arm. The controller controls the magneto-optical material of the phase-adjustable waveguide arm to generate a magneto-optical effect based on the current activation function type and weights of the optical neural network. The beam splitter receives the target input signal of the optical neural network, splits it into a corresponding number of optical signals, and transmits them to the beam combiner via the phase-adjustable waveguide arm and the conventional waveguide arm, respectively. The magneto-optical material of the phase-adjustable waveguide arm, based on the generated magneto-optical effect, performs phase adjustment on the optical signals transmitted through the phase-adjustable waveguide arm, so that the phase-adjustable waveguide arm transmits the phase-adjusted optical signals as the target optical signal to the beam combiner. The beam combiner performs signal interference on the target optical signal transmitted through the phase-adjustable waveguide arm and the optical signal transmitted through the conventional waveguide arm, and then outputs the nonlinear activation result of the target input signal. The system provided by the above solution integrates magneto-optical material in the phase-adjusting waveguide arm of the nonlinear activation unit. The controller can control the magneto-optical material of the phase-adjusting waveguide arm to generate a magneto-optical effect based on the current activation function type and weights of the optical neural network. This allows for phase adjustment of the optical signal transmitted through the phase-adjusting waveguide arm, enabling the nonlinear activation unit to adapt to the current activation function type and weights and perform corresponding nonlinear calculations on the target input signal. This achieves dynamic switching between multiple nonlinear activation functions, improving the flexibility of the optical neural network. Furthermore, various activation functions, such as Sigmoid and ReLU, can be flexibly implemented by adjusting the external magnetic field. Compared with related technologies, the system provided in this application overcomes the limitation of single nonlinear function and can meet the differentiated activation function requirements of different neural network layers. In addition, this application also supports online programming of the slope and threshold of the activation function, providing more flexible configuration options for the optical neural network and improving its generalization ability and adaptability to different tasks.

[0080] This application also provides a nonlinear activation method for optical neural networks, applied to the nonlinear activation system for optical neural networks provided in the above embodiments. The executing entity of this application embodiment is an electronic device, such as a server, desktop computer, laptop computer, tablet computer, or other electronic devices that can be used to control the nonlinear activation units in an optical neural network.

[0081] like Figure 4 The diagram shown is a flowchart illustrating the nonlinear activation method for optical neural networks provided in this application embodiment. The method includes:

[0082] Step 401: Obtain the current activation function type and weights of the optical neural network;

[0083] Step 402: Based on the current activation function type and weights of the optical neural network, control the magneto-optical material of the phase-adjusting waveguide arm of the nonlinear activation unit to generate a magneto-optical effect, so that the beam splitter of the nonlinear activation unit splits the received target input signal of the optical neural network into a corresponding number of optical signals, which are then transmitted to the beam combiner through the phase-adjusting waveguide arm and the conventional waveguide arm, respectively; wherein, based on the generated magneto-optical effect, the magneto-optical material of the phase-adjusting waveguide arm performs phase adjustment on the optical signals transmitted through the phase-adjusting waveguide arm, so that the phase-adjusting waveguide arm transmits the phase-adjusted optical signals as the target optical signals to the beam combiner;

[0084] Step 403: Based on the beam combiner, the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm are subjected to signal interference and output to obtain and output the nonlinear activation result of the target input signal.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0086] For a description of the features in the embodiment corresponding to the optical neural network nonlinear activation method, please refer to the relevant description of the embodiment corresponding to the optical neural network nonlinear activation system, which will not be repeated here.

[0087] The embodiments of this application also provide an optical neural network nonlinear activation device for performing the optical neural network nonlinear activation method provided in the above embodiments.

[0088] like Figure 5 The diagram shown is a structural schematic of an optical neural network nonlinear activation device provided in an embodiment of this application. The optical neural network nonlinear activation device 50 includes: an acquisition module 501, a control module 502, and an output module 503.

[0089] The system comprises the following modules: an acquisition module for acquiring the current activation function type and weights of the optical neural network; a control module for controlling the magneto-optical material of the phase-adjusting waveguide arm of the nonlinear activation unit to generate a magneto-optical effect based on the current activation function type and weights of the optical neural network, so that the beam splitter of the nonlinear activation unit splits the received target input signal of the optical neural network into a corresponding number of optical signals, which are then transmitted to the beam combiner through the phase-adjusting waveguide arm and the conventional waveguide arm, respectively; wherein, the magneto-optical material of the phase-adjusting waveguide arm performs phase adjustment on the optical signals transmitted through the phase-adjusting waveguide arm based on the generated magneto-optical effect, so that the phase-adjusting waveguide arm transmits the phase-adjusted optical signals as the target optical signals to the beam combiner; and an output module for performing signal interference on the target optical signals transmitted through the phase-adjusting waveguide arm and the optical signals transmitted through the conventional waveguide arm based on the beam combiner, and then outputting the nonlinear activation result of the target input signal.

[0090] For a description of the features in the embodiment corresponding to the optical neural network nonlinear activation device, please refer to the relevant description of the embodiment corresponding to the optical neural network nonlinear activation method, which will not be repeated here.

[0091] Embodiments of this application also provide an electronic device, such as... Figure 6 The diagram shown is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, including a processor 10 and a memory 20. The memory 20 stores a computer program, and the processor 10 is configured to run the computer program to execute the steps in any of the above-described embodiments of the optical neural network nonlinear activation method.

[0092] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above embodiments of the optical neural network nonlinear activation method when running.

[0093] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0094] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the optical neural network nonlinear activation method described above.

[0095] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the optical neural network nonlinear activation method.

[0096] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] The foregoing has provided a detailed description of an optical neural network nonlinear activation system, method, and apparatus provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only intended to aid in understanding the method and core ideas of this application. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A nonlinear activation system for an optical neural network, characterized in that, include: Multiple nonlinear activation units and a controller are provided. The nonlinear activation units include beam splitters and beam combiners. The beam splitters and beam combiners are connected by multiple waveguide arms. The multiple waveguide arms include phase-adjustable waveguide arms and conventional waveguide arms. The phase-adjustable waveguide arms integrate magneto-optical materials. The controller is used to control the magneto-optical material of the phase-adjusting waveguide arm to generate a magneto-optical effect according to the current activation function type and weight of the optical neural network. The beam splitter is used to receive the target input signal of the optical neural network, split the target input signal into a corresponding number of optical signals, and then transmit them to the beam combiner through a phase-adjustable waveguide arm and a conventional waveguide arm, respectively. The magneto-optical material of the phase-adjustable waveguide arm adjusts the phase of the optical signal transmitted through the phase-adjustable waveguide arm based on the generated magneto-optical effect, so that the phase-adjustable waveguide arm transmits the phase-adjusted optical signal as the target optical signal to the beam combiner. The beam combiner is used to perform signal interference on the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm, and then output the nonlinear activation result of the target input signal. The phase-adjustable waveguide arm is surrounded by an electromagnetic coil. The magnetic field strength of the electromagnetic coil acts on the magneto-optical material of the phase-adjusting waveguide arm, so that the magneto-optical material produces a magneto-optical effect; The controller is used for: Based on the current activation function type of the optical neural network, determine the target phase difference increase range between the phase-adjusting waveguide arm and the conventional waveguide arm; Based on the target phase difference increase range between the phase-adjustable waveguide arm and the conventional waveguide arm, and the relationship function between the target phase difference increase range and the change in the optical signal refractive index of the phase-adjustable waveguide arm, the driving current adjustment strategy of the electromagnetic coil is determined. According to the driving current adjustment strategy of the electromagnetic coil, the driving current of the electromagnetic coil is adjusted according to the weights of the optical neural network. The drive current regulation strategy is a current regulation rule preset by the controller that matches different activation function types, and at least includes the current regulation range.

2. The optical neural network nonlinear activation system according to claim 1, characterized in that, The relationship between the target phase difference increase range and the change in the optical signal refractive index of the phase-adjusting waveguide arm is as follows: in, This indicates the range of increase in the target phase difference. Represents pi (π). This represents the wavelength of the target input signal. This represents the change in the refractive index of the optical signal in the phase-adjusting waveguide arm. This indicates the length of the magneto-optical waveguide.

3. The optical neural network nonlinear activation system according to claim 1, characterized in that, The controller is also used for: Based on the slope of the current activation function of the optical neural network, the rate of change of the magnetic field strength of the electromagnetic coil with respect to the weights is controlled by adjusting the driving current of the electromagnetic coil.

4. The optical neural network nonlinear activation system according to claim 1, characterized in that, The controller is also used for: Based on the threshold of the current activation function of the optical neural network, the maximum value of the magnetic field strength of the electromagnetic coil is limited by limiting the driving current of the electromagnetic coil.

5. The optical neural network nonlinear activation system according to claim 1, characterized in that, The optical signal refractive index of the phase-adjusting waveguide arm changes according to the magnetic field strength.

6. The optical neural network nonlinear activation system according to claim 5, characterized in that, The relationship between the change in the optical signal refractive index of the phase-adjusting waveguide arm and the magnetic field strength is as follows: in, This represents the change in the refractive index of the optical signal in the phase-adjusting waveguide arm. This represents a preset constant. This indicates the strength of the magnetic field.

7. The optical neural network nonlinear activation system according to claim 1, characterized in that, The bundle combiner is used for: Based on the phase difference between the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm, signal interference is performed on the target optical signal and the conventional waveguide arm to obtain and output the nonlinear activation result of the target input signal.

8. The optical neural network nonlinear activation system according to claim 7, characterized in that, The bundle combiner is specifically used for: The optical signal combining factor is determined based on the phase difference between the target optical signal and the optical signal. Based on the optical signal combining factor, signal interference is performed on the target optical signal and the optical signal to obtain and output the nonlinear activation result of the target input signal.

9. The optical neural network nonlinear activation system according to claim 8, characterized in that, The bundle combiner is used for: The nonlinear activation result of the target input signal is determined based on the following formula: in, This represents the nonlinear activation result of the target input signal. This represents the target input signal. This indicates the phase difference between the target optical signal and the optical signal. This represents the optical signal combining factor.

10. The optical neural network nonlinear activation system according to claim 1, characterized in that, include: Multiple nonlinear activation layers, each of which includes multiple nonlinear activation units; After obtaining the nonlinear activation result, any of the nonlinear activation units uses the nonlinear activation result as a new input signal to input to the next adjacent optical interference layer, so as to output a new target input signal through the optical interference layer to the beam splitter of the nonlinear activation unit in the next nonlinear activation layer, until all nonlinear activation units in the last nonlinear activation layer output the nonlinear activation result; The nonlinear activation results output by all nonlinear activation units in the last nonlinear activation layer are taken as the target nonlinear activation result of the optical neural network.

11. A nonlinear activation method for an optical neural network, characterized in that, include: Obtain the current activation function type and weights of the optical neural network; Based on the current activation function type and weights of the optical neural network, the magneto-optical material of the phase-adjusting waveguide arm of the nonlinear activation unit is controlled to generate a magneto-optical effect. This causes the beam splitter of the nonlinear activation unit to split the received target input signal of the optical neural network into a corresponding number of optical signals, which are then transmitted to the beam combiner through the phase-adjusting waveguide arm and the conventional waveguide arm, respectively. The magneto-optical material of the phase-adjusting waveguide arm, based on the generated magneto-optical effect, performs phase adjustment on the optical signals transmitted through the phase-adjusting waveguide arm, so that the phase-adjusting waveguide arm transmits the phase-adjusted optical signals as the target optical signals to the beam combiner. Based on the beam combiner, the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm are interfered and output to obtain and output the nonlinear activation result of the target input signal. The phase-adjusting waveguide arm is surrounded by an electromagnetic coil; the magnetic field strength of the electromagnetic coil acts on the magneto-optical material of the phase-adjusting waveguide arm to cause the magneto-optical material to produce a magneto-optical effect. The method further includes: Based on the controller's determination of the target phase difference increase range between the phase-adjusting waveguide arm and the conventional waveguide arm according to the current activation function type of the optical neural network; Based on the target phase difference increase range between the phase-adjustable waveguide arm and the conventional waveguide arm, and the relationship function between the target phase difference increase range and the change in the optical signal refractive index of the phase-adjustable waveguide arm, the driving current adjustment strategy of the electromagnetic coil is determined. According to the driving current adjustment strategy of the electromagnetic coil, the driving current of the electromagnetic coil is adjusted according to the weights of the optical neural network. The drive current regulation strategy is a current regulation rule preset by the controller that matches different activation function types, and at least includes the current regulation range.

12. A nonlinear activation device for an optical neural network, characterized in that, include: The acquisition module is used to obtain the current activation function type and weights of the optical neural network; The control module is used to control the magneto-optical material of the phase-adjusting waveguide arm of the nonlinear activation unit to generate a magneto-optical effect according to the current activation function type and weight of the optical neural network. This causes the beam splitter of the nonlinear activation unit to split the received target input signal of the optical neural network into a corresponding number of optical signals, which are then transmitted to the beam combiner through the phase-adjusting waveguide arm and the conventional waveguide arm, respectively. The magneto-optical material of the phase-adjusting waveguide arm adjusts the phase of the optical signals transmitted through the phase-adjusting waveguide arm based on the generated magneto-optical effect, so that the phase-adjusting waveguide arm transmits the phase-adjusted optical signals as the target optical signals to the beam combiner. The output module is used to perform signal interference on the target optical signal transmitted through the phase-adjusting waveguide arm and the optical signal transmitted through the conventional waveguide arm based on the beam combiner, and then output the nonlinear activation result of the target input signal. The phase-adjusting waveguide arm is surrounded by an electromagnetic coil; the magnetic field strength of the electromagnetic coil acts on the magneto-optical material of the phase-adjusting waveguide arm to cause the magneto-optical material to produce a magneto-optical effect. The control module is also used for: Based on the controller's determination of the target phase difference increase range between the phase-adjusting waveguide arm and the conventional waveguide arm according to the current activation function type of the optical neural network; Based on the target phase difference increase range between the phase-adjustable waveguide arm and the conventional waveguide arm, and the relationship function between the target phase difference increase range and the change in the optical signal refractive index of the phase-adjustable waveguide arm, the driving current adjustment strategy of the electromagnetic coil is determined. According to the driving current adjustment strategy of the electromagnetic coil, the driving current of the electromagnetic coil is adjusted according to the weights of the optical neural network. The drive current regulation strategy is a current regulation rule preset by the controller that matches different activation function types, and at least includes the current regulation range.

13. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the nonlinear activation method for an optical neural network as described in claim 11 when executing the computer program.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the nonlinear activation method for optical neural networks as described in claim 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the nonlinear activation method for optical neural networks as described in claim 11.

Images