An optical neural network computing system based on nonlinear activation of doped gain medium

Abstract

The present application relates to the technical field of optical neural network computing, and provides an optical neural network computing system based on doped gain medium nonlinear activation, which comprises a signal input module, a core computing module and a signal output module.The signal input module is used for linearly mapping a preprocessed data matrix into optical power signals of different wavelengths, transmitting the signals to an optical fiber link after convergence, and avoiding interference of reflected light on the stability of a signal source.The core computing module comprises a doped optical fiber connected with a pump laser, and the doped optical fiber works in a gain near saturation region by injecting energy through the pump laser, and the gain saturation characteristics of the doped optical fiber are used to realize nonlinear activation conversion of optical signals.The signal output module is used for separating the multi-wavelength optical signals after nonlinear conversion, attenuating and regulating each channel optical signal based on training weights, and reading and outputting intensity after superimposing each optical signal, as a prediction result of a neural network.The present application realizes efficient computing with full optical domain, low delay, wide bandwidth and low power consumption.

Description

Technical Field

[0001] This invention relates to the technical field of optical neural network computing, and more particularly to an optical neural network computing system based on nonlinear activation of a doped gain medium. Background Technology

[0002] With the rapid iteration of artificial intelligence and machine learning technologies, the demand for computing resources for various complex tasks (such as large-scale data mining, high-dimensional image recognition, and deep learning model training) is increasing exponentially, with computing scale and energy consumption rising in tandem. Traditional electronic computing architectures, limited by the physical characteristics of electronic devices, are increasingly facing two core bottlenecks: on the one hand, the bandwidth limitation of electronic signal transmission makes it difficult to break through the ceiling of data processing speed, failing to meet the demands of ultra-high throughput computing; on the other hand, the heat dissipation problem generated by electronic components during operation is becoming increasingly prominent, not only increasing energy consumption and heat dissipation costs, but also potentially affecting device stability due to overheating, severely restricting the performance upgrade and scale expansion of computing systems.

[0003] Against this backdrop, optical computing, with its inherent technological advantages, is widely recognized by academia and industry as the core cornerstone for building next-generation high-performance information processing systems. Compared to electronic computing, optical computing possesses ultra-high signal transmission bandwidth, inherent large-scale parallel processing capabilities, and extremely low energy dissipation during photon transmission and interaction, fundamentally avoiding the bandwidth and heat dissipation dilemmas of electronic computing. Currently, research and industry have explored various optical neural network (ONN) architectures, covering different technical paths from free-space diffraction optics systems to highly integrated photonic chips. These architectures have demonstrated outstanding performance potential in ultra-fast inference and linear computing scenarios, laying the foundation for the practical application of all-optical computing.

[0004] However, in the process of advancing the practical application of all-optical computing, a key technical bottleneck has remained unresolved—the lack of an efficient, scalable, and purely optical nonlinear activation mechanism. In electronic neural networks, nonlinear activation functions (such as Sigmoid and ReLU) are core components for achieving complex feature abstraction and enhancing model expressive power. They can map linearly transformed features to a nonlinear space, thereby enabling the network to learn complex data patterns. However, achieving a nonlinear response equivalent to that of an electronic activation function in the optical domain is extremely difficult. Existing optical nonlinear schemes generally have significant limitations: some schemes rely on an optical-electrical-optical (OEO) conversion process, that is, first converting the optical signal into an electrical signal for nonlinear processing, and then converting it back into an optical signal for transmission. This conversion process not only introduces an electronic bottleneck, directly sacrificing the speed advantage and energy efficiency of optical computing, but also increases the system complexity. Another type of scheme uses the Kerreffect of materials or other third-order nonlinear effects to achieve optical nonlinearity, but these schemes often require extremely high peak optical power as a prerequisite, making it difficult to operate stably under continuous wave conditions, which greatly reduces their practicality. In addition, devices based on the resonance principle, such as microring resonators, can achieve a certain degree of nonlinear response, but they have inherent defects—extremely narrow operating bandwidth and high sensitivity to changes in ambient temperature. Even small temperature fluctuations can cause performance drift. At the same time, the structural characteristics of these devices also make it difficult to achieve large-scale integration and cascade expansion, which cannot meet the construction requirements of complex optical neural networks.

[0005] In summary, the fields of optical computing and optical neural networks urgently need to explore a completely new technological path: developing a nonlinear activation element that is essentially based on pure optical principles, has a working bandwidth that fully covers communication bands, operates with low power consumption, and is seamlessly compatible with existing mature fiber optic communication systems and integrated photonic platforms. A breakthrough in this core technology will effectively overcome the key bottlenecks in the physical realization of all-optical neural networks, propelling all-optical computing from the laboratory to industrial applications and providing core support for the development of next-generation high-performance computing systems. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide an optical neural network computing system based on nonlinear activation of doped gain media. Through a doped fiber nonlinear activation mechanism, a spectral parallel architecture, and a hybrid "digital training-optical inference" architecture, it achieves high-efficiency computing with low latency, wide bandwidth, and low power consumption across the entire optical domain. This improves computational density and feature extraction capabilities, avoids the challenges of all-optical training, and is compatible with existing fiber optic communication platforms with low industrialization costs. It demonstrates high accuracy and energy efficiency potential far exceeding that of traditional electronic processors in nonlinear classification and image recognition tasks.

[0007] The nonlinear activation principle used in this invention is not limited to fiber optic architectures. Based on the same physical logic, compact all-optical activation functionality can also be achieved by fabricating doped waveguides (such as erbium-doped waveguide amplifiers EDWA) on photonic integration platforms such as silicon-based, thin-film lithium niobate, and silicon nitride.

[0008] It should be understood that the optical nonlinear mechanism based on the gain response of doped gain media proposed in this invention is not limited to the Extreme Learning Machine (ELM) network model. This mechanism, as a reusable optical nonlinear operator, can be extended to various neural network models, including but not limited to feedforward fully connected networks (MLP), convolutional neural networks (CNN), recurrent neural networks (RNN / LSTM / GRU), Transformer networks, and reservoir / echo state networks, to implement the activation functions or nonlinear mapping processes in the aforementioned network models. Here, ELM is only used as a preferred embodiment to illustrate the hybrid implementation path of "digital training-optical inference".

[0009] Furthermore, by adjusting the pump power and / or input optical power, the doped gain medium can operate in the unsaturated region, near-saturated region, or saturated region, exhibiting near-linear gain response, weak nonlinear response, or gain-squeezed nonlinear response in different gain operating ranges, thereby providing diverse optical operator support for different neural network models or different network layers.

[0010] The above-mentioned objective of this invention is achieved through the following technical solutions: A computational system for optical neural networks based on nonlinear activation of doped gain media includes a signal input module, a core computation module, and a signal output module connected sequentially via an optical transmission link. The signal input module is used to linearly map the preprocessed data matrix into optical power signals of different wavelengths, which are then combined and transmitted to the optical fiber link, while avoiding interference from reflected light on the stability of the signal source. The core computing module includes a doped gain medium using any one of the following: doped fiber, rare-earth doped waveguide, or metal ion doped waveguide. The doped gain medium is connected to a pump laser, and energy is injected by the pump laser to make the doped gain medium work in a preset working range or a controlled working point. The gain working range includes at least one of the unsaturated region, near-saturated region, and saturated region of the doped gain medium. The nonlinear activation transformation of the optical signal is realized by using the gain response of the doped gain medium to the input optical power within the working range. The signal output module is used to separate the multi-wavelength optical signals after nonlinear transformation, adjust the attenuation of each channel optical signal based on the training weights, and then read the output intensity after superimposing the optical signals of each channel as the prediction result of the neural network.

[0011] Furthermore, the signal input module includes a control computer, a tunable laser array with a wavelength band matched to the doped gain medium, a wavelength division multiplexer, and a first single-stage optical isolator with a wavelength band matched to the laser. The tunable laser array is C-band and controlled by the control computer. The pre-processed data matrix, including grayscale image pixel values, is linearly mapped to optical power signals of different wavelengths through digital-to-analog conversion. The wavelength division multiplexer has a channel spacing of 100 GHz, which is used to combine different wavelength optical signals output by the tunable laser array into the same optical fiber link. The first single-stage optical isolator in the C-band is placed in the optical path before the core computing module to block reflected light from subsequent links and avoid interfering with the stability of the laser.

[0012] Furthermore, the signal input module and the core computing module work together to implement a spectral parallel computing architecture: Multiple optical signals of different wavelengths output from the tunable laser array are combined by the wavelength division multiplexer in the signal input module and simultaneously transmitted and processed in a single doped fiber of the core computing module. Due to the difference in emission / absorption cross sections and the shared inversion particle number, each wavelength signal in the doped gain medium forms a wavelength-dependent nonlinear gain response, which is accompanied by a controllable inter-channel coupling effect. This effect can be suppressed or utilized under appropriate channel spacing and power calibration conditions to enhance the feature mapping capability.

[0013] Furthermore, the core computing module is composed of a section of highly doped optical fiber; The high-concentration doped fiber, with a length of 15 cm, has an absorption coefficient of 110 dB / m at 1530 nm. It can achieve sufficient gain saturation effect within an extremely short fiber length, which reduces the time-of-flight delay of the optical signal and minimizes instability caused by dispersion and environmental disturbances.

[0014] Furthermore, in the core computing module, the doped fiber is connected to the pump laser via a 980 / 1550nm wavelength division multiplexing coupler. The pump laser is a 980nm pump laser that matches the wavelength of the doped fiber. By adjusting the pump power and / or the input optical signal power, the doped fiber can operate in at least one gain state in the unsaturated region, near-saturated region, or saturated region. When operating in the near-saturated region or saturated region, the fiber exhibits a strong nonlinear compression response to the input signal. A second single-stage optical isolator matching the pump source wavelength is also provided between the pump laser and the doped fiber.

[0015] Furthermore, the doped optical fiber of the core computing module can be selected from doped optical fiber, integrated optical planar waveguide, ridge waveguide or thin film waveguide; its doping ions include at least one of rare earth metal ions or transition metal ions, and there is no restriction on the type of doping. The rare earth metal ions include erbium, ytterbium, thulium, holmium, and neodymium ions, and the transition metal ions include bismuth ions.

[0016] Furthermore, the signal output module includes a two-stage isolator matching the wavelength band of the laser, a wavelength demultiplexer, a multi-channel variable optical attenuator array, an optical coupler, and a multi-channel optical power meter; The dual-stage isolator is disposed in the optical path before the wave demultiplexer; The wavelength demultiplexer is used to separate the multi-wavelength mixed optical signal after nonlinear transformation by doped fiber into independent channel signals. Then, the optical signal of each channel passes through a variable optical attenuator. The variable optical attenuator performs attenuation control based on training weights. The attenuation amount is determined by the trained output weights. In the inference stage, the weight matrix obtained by training is mapped to the optical attenuation value dB through mathematical transformation, thereby completing the multiplication part of the weighted summation operation in the optical domain. Finally, the processed optical signals are incoherently superimposed through an optical coupler, and the total light intensity after superposition is read by the multi-channel optical power meter as the classification prediction output of the neural network.

[0017] Furthermore, the core computing module and the signal output module work together to complete the nonlinear activation and weighting operations: When the doped gain medium in the core computing module operates in the unsaturated region, near-saturated region, or saturated region, it generates a near-linear gain response, a weak nonlinear response, or a gain-saturated nonlinear response to the input optical signal, forming an optical response relationship for linear mapping operators or nonlinear activation operators in the neural network model. During the inference phase, the control computer converts the trained output weight matrix into attenuation control commands and sends them to the multi-channel variable optical attenuator array of the signal output module. By adjusting the attenuation of each channel's variable optical attenuator, the multiplication part of the weighted summation operation is completed in the optical domain.

[0018] Furthermore, the system is used to implement an optical inference or digital training-optical inference hybrid computing architecture for a neural network model. The neural network model includes, but is not limited to, Extreme Learning Machine (ELM), Feedforward Fully Connected Neural Network (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), LSTM, GRU, Transformer network, Reservoir Computation / Echo State Network (RC / ESN), or combinations thereof. The optical gain response provided by the core computing module is used to implement the activation function or nonlinear operator in the neural network model, and the signal output module is used to implement weight modulation and weighted summation operations.

[0019] Furthermore, using the Extreme Learning Machine (ELM) algorithm as a preferred implementation method, a hybrid architecture of digital training and optical inference is constructed, which is linked with the three main modules: The signal input module fixes the input layer weights through linear mapping of optical power to achieve random projection; the core calculation module utilizes the gain saturation characteristics of doped optical fibers to complete the hidden layer nonlinear transformation; the output layer weights are calculated using the ridge regression algorithm in the digital domain and then adapted to the variable optical attenuator control requirements of the signal output module.

[0020] Furthermore, the specific workflow of the system includes: Data preprocessing: The computer controls the input vector x and the randomly generated input weight matrix W. in Multiply and add bias b to obtain an intermediate feature vector, then linearly scale the intermediate feature vector to the physically achievable milliwatt-level optical power range to generate the laser driving signal; Optical signal input: The control computer sends the drive signal to the tunable laser array of the signal input module, controls it to output an optical signal of corresponding power, and the optical signal is combined by a wavelength division multiplexer, isolated by a single-stage isolator, and then transmitted to the core computing module; Nonlinear activation: The optical signal is fed into the doped optical fiber of the core computing module in the pumping state. Due to the gain saturation characteristics of the optical fiber, the output optical power no longer increases linearly with the input, but exhibits a nonlinear curve similar to a gradual saturation, thus completing the nonlinear activation of the hidden layer. Training phase: The computer is used to collect the nonlinear response matrix H output by the core computing module, and the optimal output weight beta is calculated on the computer using the ridge regression algorithm; Inference phase: The control computer converts the output weight beta into attenuation control instructions for VOA and configures them into the hardware. The system then performs all-optical classification on the newly input unknown data and sends it to the multi-channel variable optical attenuator array of the signal output module to adjust the attenuation of each channel. The optical signal output by the core computing module is sequentially isolated by a double-stage isolator, separated by a wave demultiplexer, and then attenuated by the variable optical attenuator of the corresponding channel. After incoherent superposition by an optical coupler, the total optical intensity is finally read by a multi-channel optical power meter and used as the prediction output of the neural network. When the signal input module uses a tunable laser array with 5 wavelength channels and the core computing module is configured with highly doped optical fiber, the system can achieve an accuracy of over 90% in the all-optical classification task of the Fashion-MNIST dataset.

[0021] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Break through the bottleneck of core technologies and achieve key breakthroughs in all-optical computing. Overcoming the challenges of purely optical nonlinear activation and circumventing electronic bottlenecks: Addressing the core pain point of existing optical neural networks lacking endogenous nonlinear activation elements that do not require optical-electrical-optical (OEO) conversion, this invention utilizes the gain saturation characteristics of near-saturation regions of doped optical fibers (erbium-doped, ytterbium-doped, etc.) to construct a physical nonlinear activation mechanism. Based on the energy level transition dynamics of rare-earth ions, a smooth nonlinear mapping between input and output optical power is naturally formed through the gain compression effect, which leads to the consumption of metastable particle numbers and a decrease in stimulated emission rate due to the enhancement of input optical power. This design achieves true all-optical nonlinear feature embedding, eliminating the need for photoelectric conversion throughout signal transmission and processing. It completely overcomes the electronic bottleneck limitations of traditional OEO conversion schemes, fully preserving the speed and energy efficiency advantages of optical computing while avoiding complex heterogeneous integration processes, thus clearing the core obstacles to the physical realization of all-optical neural networks.

[0022] Addressing the technical challenges of all-optical training and weight control: This invention innovatively integrates the Extreme Learning Machine (ELM) algorithm to construct a hybrid architecture of "digital training-optical inference," addressing the difficulties in physically implementing gradient backpropagation in all-optical neural networks and the high hardware complexity of active optical weight networks. Input layer weights are fixed through linear mapping of optical power to achieve random projection; the nonlinear transformation of the hidden layer is automatically completed by the physical saturation effect of the doped fiber; and the output layer weights are rapidly calculated and mapped to the physical attenuation of a variable optical attenuator (VOA) using a digital domain ridge regression algorithm. This architecture cleverly circumvents the technical difficulties of complex gradient backpropagation in the optical domain, transforming the highly complex dynamic weight control into simple passive optical attenuation control. This significantly reduces the hardware design complexity and implementation threshold of all-optical networks while ensuring high efficiency in network training and inference.

[0023] Overcoming the limitations of traditional optical fiber computing density and feature extraction dimensionality: Addressing the issues of large system size, low computing density, and the inability of a single physical response to provide rich feature extraction dimensions resulting from traditional single-channel optical fiber transmission, this invention employs a spectral parallel computing architecture. Through dense wavelength division multiplexing (DWDM) technology, multiple optical signals of different wavelengths are simultaneously transmitted and processed within a single doped optical fiber. Leveraging the different emission / absorption cross-sections of dopant ions at different wavelengths, each wavelength channel forms a unique nonlinear gain response, equivalent to constructing multiple physical neurons with unique activation characteristics. This design, without increasing the number of physical components, significantly improves the computing throughput of a single optical fiber and reduces hidden layer feature redundancy through "spectral diversity" and spectral orthogonality, significantly enhancing the network's ability to capture complex data patterns.

[0024] (2) Performance indicators have been comprehensively upgraded, with significant advantages in computing power and energy efficiency. Low latency, wide bandwidth, and high stability: All-optical signal transmission and processing eliminates photoelectric conversion, theoretically achieving nanosecond-level or even lower processing latency. This is far superior to traditional electronic computing architectures and optical computing solutions relying on OEO conversion, meeting the demands of ultrafast inference scenarios. The nonlinear response of doped optical fibers is a passive broadband response, fully covering the C-band and extending to the L-band, compatible with communication band requirements. It eliminates the need for narrowband devices such as microring resonators, avoiding the bandwidth limitations and environmental temperature sensitivity issues of narrowband solutions, significantly improving system stability.

[0025] Significantly improved computational density and throughput: The spectral parallel architecture carries parallel computing across multiple wavelength channels via a single optical fiber, eliminating the need for stacking numerous physical fibers. This significantly reduces system size and increases computational density per unit space, providing a compact implementation path for building large-scale optical neural networks. The parallel processing mode of multiple wavelength channels greatly improves data processing throughput, efficiently handling large-scale data computation needs. Compared to traditional single-channel fiber computing solutions, processing efficiency is improved by orders of magnitude.

[0026] The energy efficiency potential far exceeds that of traditional solutions, meeting the demand for low power consumption: Theoretical analysis shows that the computing efficiency of the core computing part of this invention (doped optical fiber) alone is expected to reach 194 TOPS / W, far exceeding the energy efficiency level of existing electronic AI accelerators, significantly alleviating the thermal power consumption bottleneck of traditional electronic computing architectures. Although the current system-level energy efficiency is limited by the peripheral laser source and photoelectric conversion interface, with the technological iteration towards integrated photonic platforms (such as thin-film lithium niobate waveguides and silicon-based light sources), the system's "wall-plug efficiency" is expected to achieve an order-of-magnitude improvement, further highlighting the low power consumption advantage and aligning with the development trend of green computing.

[0027] High computational accuracy and adaptability to complex tasks: Experimental data verifies that the system can achieve 100% accuracy in complex nonlinear classification tasks (such as the XOR problem) and more than 90% classification accuracy in high-dimensional image recognition tasks (such as the Fashion-MNIST dataset) using only 5 wavelength channels. Its performance is comparable to that of a digital computer and can meet the complex data processing needs in real-world application scenarios.

[0028] (3) It is highly practical and adaptable, and covers a wide range of scenarios. Universal fiber type adaptability, suitable for multi-band requirements: The technical solution of this invention has good universality for doped fiber types. It is not only applicable to doped fibers, but can also be extended to other rare earth doped fibers such as ytterbium-doped and thulium-doped fibers. It can be flexibly adapted to the computing requirements of different bands, providing diversified implementation paths for multi-band optical computing tasks.

[0029] Flexible task adaptability, covering a variety of computing scenarios: The system demonstrates excellent performance in tasks such as nonlinear classification and high-dimensional image recognition, and can be widely used in various computing scenarios related to artificial intelligence and machine learning, such as large-scale data mining, image recognition, and pattern classification. It has strong adaptability and broad application prospects.

[0030] (4) High feasibility for industrialization, low implementation cost and risk. Mature core components and low implementation threshold: The core components on which this system relies (tunable laser, wavelength division multiplexer, doped fiber, optical isolator, variable optical attenuator, etc.) are all commercially available products that are extremely mature and standardized in the optical fiber communication industry. There is no need to develop new semiconductor processes or special materials, which greatly reduces the R&D costs and cycle of technology implementation.

[0031] High system stability and low operation and maintenance costs: The system, built on mature commercial components, has extremely high stability and reliability, reducing the risk of failure and maintenance costs in subsequent operation and maintenance processes; at the same time, the solution avoids complex heterogeneous integration processes and difficult optical domain control designs, further reducing the uncertainty in the system operation process.

[0032] Providing a path to large-scale deployment: The technical solution of this invention provides a practical technical path for building large-scale, low-power photonic computing clusters. Through the spectral parallel architecture and compatibility with existing optical fiber communication platforms, the system can be flexibly expanded and cascaded to meet different scale requirements from small and medium-sized computing tasks to large-scale cluster computing, and has great potential for industrialization and promotion. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall structure of the optical neural network computing system based on nonlinear activation of doped gain medium according to the present invention.

[0034] 1: Control computer; 2: Tunable laser array matched to doped fiber band; 3: Wavelength division multiplexer; 4: First single-stage optical isolator matched to laser band; 5: Doped fiber; 6: Second single-stage optical isolator matched to pump source band; 7: Pump laser matched to doped fiber band; 8: Double-stage isolator matched to laser band; 9: Wavelength demultiplexer; 10: Variable optical attenuator array; 11: Multi-channel optical power meter. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0037] It should be noted that the following embodiments use "doped fiber" as a preferred form of doped gain medium, ELM as a preferred algorithm framework, and near-saturation / saturation region as a preferred working range for illustration; however, the present invention is not limited thereto. The doped gain medium may also be rare earth doped waveguide or metal ion doped waveguide, the neural network model may also be MLP / CNN / RNN / Transformer, etc., and the gain working range may also include the unsaturated region, the near-saturated region, and the saturated region.

[0038] The problems solved by this invention specifically include the following, providing a complete overview of the pain points of existing technologies, innovative solutions, principles, and technical effects: (1) Main problems and corresponding solutions The main problem solved by this invention is that there has always been a lack of a highly efficient, wideband, low-power, and purely optical nonlinear activation element in optical neural network technology that does not require optical-electric-optical (OEO) conversion. This limits the ability of all-optical computing to abstract complex features.

[0039] The main innovation in solving this problem is to propose using the gain saturation characteristics of doped gain media (such as erbium-doped fiber, ytterbium-doped fiber, doped waveguide, etc.) operating in the near-saturation region as a nonlinear activation function for optical neurons.

[0040] The innovative principle behind solving the problem is based on the energy level transition dynamics of rare earth ions (such as erbium-doped ions). When the input optical signal power increases, the number of metastable particles in the optical fiber is greatly consumed, leading to a decrease in the stimulated emission rate and a reduction in the gain coefficient, forming a response where "gain is compressed as the input optical power increases" (i.e., gain saturation effect). This physical process naturally presents a smooth, simulated nonlinear mapping relationship between the input optical power and the output optical power, similar to the characteristics of activation functions commonly used in neural networks (such as Sigmoid).

[0041] The innovative features bring about the following technical effects: They achieve true all-optical nonlinear feature embedding, and the signal does not need to undergo photoelectric conversion during transmission. Theoretically, the delay can reach the sub-nanosecond level. At the same time, this nonlinearity is passive and broadband, which can cover the entire C-band and even extend to the L-band. It utilizes the medium properties of optical fiber communication and avoids complex heterogeneous integration processes.

[0042] (2) Secondary problem 1 and corresponding solutions Secondary problem 1 addressed by this invention: Traditional optical fiber devices are usually only used as single-channel transmission media. To build a large-scale optical neural network, it is often necessary to stack a large number of physical optical fibers, resulting in a large system size and low computational density. In addition, a single physical response is difficult to provide rich feature extraction dimensions.

[0043] Secondary innovation in solving this problem: The first innovation is to adopt a spectral parallel computing architecture and use dense wavelength division multiplexing (DWDM) technology to simultaneously transmit and process multiple optical signals of different wavelengths in a single doped fiber.

[0044] The principle behind the innovation is as follows: Due to the natural differences in the emission / absorption cross-sections of rare earth ions (such as erbium-doped ions) at different wavelengths, signals of different wavelengths in the doped gain medium form a wavelength-dependent nonlinear gain response due to the differences in emission / absorption cross-sections and the shared inversion particle number. This response may be accompanied by a controllable inter-channel coupling effect. This effect can be suppressed or utilized under appropriate channel spacing and power calibration conditions to enhance the feature mapping capability.

[0045] The technological effect brought about by the innovation is as follows: This physical mechanism not only improves the computing throughput of a single optical fiber, but also introduces spectral diversity. Each wavelength channel is equivalent to a physical neuron with unique activation characteristics. This spectral orthogonality reduces the redundancy of hidden layer features and significantly enhances the network's ability to capture complex data patterns without increasing the number of physical devices.

[0046] (3) Secondary problem 2 and corresponding solutions The second minor problem solved by this invention is that gradient backpropagation in all-optical neural networks is difficult to achieve physically, and the hardware complexity of dynamically controlling active optical weight networks is extremely high.

[0047] Secondary innovation in solving this problem 2: Combining physical layer design with Extreme Learning Machine (ELM) algorithm to build a hybrid architecture of digital training-optical inference.

[0048] The second principle of innovation in solving the problem: In this architecture, the input layer weights are fixed through linear mapping (random projection) of optical power; the nonlinear transformation of the hidden layer is automatically completed by the gain saturation effect of the doped fiber; and the output layer weights are calculated by the ridge regression algorithm in the digital domain and finally converted into the physical attenuation of the variable optical attenuator (VOA).

[0049] The second technological effect brought about by the innovation is that the design cleverly avoids the difficult problem of gradient backpropagation in all-optical networks, transforming the complex dynamic adjustment of weights into simple passive light attenuation control. Experimental data shows that the system can achieve 100% accuracy in complex tasks (such as the XOR problem) and more than 90% accuracy in high-dimensional classification tasks (such as Fashion-MNIST), while demonstrating energy efficiency potential far exceeding that of traditional electronic processors.

[0050] The technical solution of this invention has extremely high industrial feasibility and practical value. From a hardware perspective, the core components upon which this system relies—including tunable lasers, wavelength division multiplexers (WDM), erbium-doped fiber (EDF), optical isolators, and variable optical attenuators (VOAs)—are all extremely mature and standardized commercial off-the-shelf products in the optical fiber communication industry. This means that the implementation of this technology does not require the development of entirely new semiconductor processes or special materials, greatly reducing the implementation threshold and cost, and possessing extremely high system stability.

[0051] From a performance perspective, this scheme demonstrates excellent energy efficiency potential. Theoretical analysis shows that the computational efficiency considering only the fiber core could potentially reach 194 TOPS / W, far exceeding existing electronic AI accelerators. Although current system-level energy efficiency is limited by the peripheral laser source and photoelectric conversion interface, the "wall-plug efficiency" of this system is expected to improve by orders of magnitude with future migration to integrated photonic platforms (such as thin-film lithium niobate waveguides or silicon-based light sources). Furthermore, the universality of this scheme across fiber types means it is not limited to erbium-doped fibers but can be extended to other rare-earth-doped fibers such as ytterbium-doped and thulium-doped fibers to meet the computational needs of different wavelength bands, providing a practical technical path for building large-scale, low-power photonic computing clusters.

[0052] The following is an illustration through specific examples: First Embodiment like Figure 1 As shown, this embodiment provides an optical neural network computing system based on nonlinear activation of doped gain medium, including a signal input module, a core computing module, and a signal output module connected in sequence via optical fibers; This embodiment uses doped fiber + ELM as an example and is not a limitation; similarly, it can be replaced with doped waveguide and adapted to MLP / CNN / Transformer; The signal input module is used to linearly map the preprocessed data matrix into optical power signals of different wavelengths, which are then combined and transmitted to the optical fiber link, while avoiding interference from reflected light on the stability of the signal source. The core computing module includes a section of doped fiber 5, which is connected to a pump laser 7. Energy is injected by the pump laser 7 to make the doped fiber 5 work in the gain saturation region. The nonlinear activation transformation of the optical signal is realized by utilizing the gain saturation characteristics of the doped fiber 5. The signal output module is used to separate the multi-wavelength optical signals after nonlinear transformation, adjust the attenuation of each channel optical signal based on the training weights, and then read the output intensity after superimposing the optical signals of each channel as the prediction result of the neural network.

[0053] The specific structures of the signal input module, core computing module, and signal output module are described below: The signal input module includes a control computer 1, a tunable laser array 2 with a wavelength band matched to the doped fiber 5, a wavelength division multiplexer 3, and a first single-stage optical isolator 4 with a wavelength band matched to the laser. The tunable laser array 2 is a C-band laser controlled by the control computer 1, which linearly maps the pre-processed data matrix, including grayscale image pixel values, into optical power signals of different wavelengths through digital-to-analog conversion. The wavelength division multiplexer 3 (DWDM MUX) has a channel spacing of 100 GHz and is used to combine different wavelength optical signals output by the tunable laser array 2 into the same optical fiber link. The first single-stage optical isolator 4 in the C-band is placed in the optical path before the core computing module to block reflected light from subsequent links and avoid interfering with the stability of the laser.

[0054] The signal input module and the core computing module work together to implement a parallel spectral computing architecture. Multiple optical signals of different wavelengths output by the tunable laser array 2 are combined by the wavelength division multiplexer 3 in the signal input module and simultaneously transmitted and processed in the single doped fiber 5 of the core computing module. Each wavelength signal utilizes the natural difference in the emission / absorption cross-section of the doped ions at different wavelengths in the doped fiber 5 to form a unique nonlinear gain response.

[0055] The core design of this signal input module is to achieve efficient conversion and stable transmission of "data-optical signal" and to provide fundamental support for spectral parallel computing. The working logic of each component and its collaborative mechanism is as follows: The control computer 1 serves as the core control unit for signal input. Its pre-processed data (including grayscale image pixel values, etc.) needs to be converted into signals that can be processed in the optical domain. The C-band tunable laser array 2 is the key device for realizing this conversion. Through digital-to-analog conversion technology, the discrete digital data matrix is ​​linearly mapped into optical power signals corresponding to different wavelengths, so that the strength of the data can be intuitively reflected by the level of optical power. Moreover, the selection of the C-band is in line with the mainstream application band of optical fiber communication, and has the advantages of low loss and wide bandwidth.

[0056] The wavelength division multiplexer 3 adopts a 100GHz channel spacing design. This parameter ensures high-density integration and low crosstalk transmission of multi-wavelength signals. It can efficiently merge multiple optical signals of different wavelengths output by the tunable laser array 2 into the same optical fiber link, avoiding the cumbersome design of multiple optical fibers required for multiple channels, and greatly improving the system integration and transmission efficiency.

[0057] The C-band first single-stage optical isolator 4, located in front of the core computing module, is specifically designed to block reflected light generated by subsequent links (such as the doped fiber 5 of the core computing module and the reflective interface), prevent reflected light from flowing back into the tunable laser array 2, avoid stability problems such as laser operating point offset and output power fluctuation, and ensure the continuity and accuracy of optical signal transmission.

[0058] The collaborative design of the signal input module and the core computing module is the key to realizing parallel spectral computing: the multi-wavelength optical signals after being combined by the wavelength division multiplexer 3 are synchronously transmitted and processed in a single doped fiber 5 without the need to increase the number of physical fibers, which significantly improves the computing density; at the same time, by taking advantage of the natural differences in the emission / absorption cross-sections of doped ions (such as erbium-doped ions) at different wavelengths in the doped fiber 5, each wavelength signal will form a unique nonlinear gain response, which is equivalent to building multiple parallel "physical neurons", providing rich dimensional support for feature extraction and nonlinear processing of subsequent complex data, and realizing the parallel computing capability of the spectral dimension from the hardware level.

[0059] The core computing module consists of a section of highly doped optical fiber 5; the highly doped optical fiber 5 is 15 cm long and has an absorption coefficient of 110 dB / m at 1530 nm; the high doping concentration design can achieve sufficient gain saturation effect within an extremely short fiber length, which reduces the time-of-flight delay of the optical signal and minimizes instability caused by dispersion and environmental disturbances.

[0060] In the core computing module, the doped fiber 5 is connected to the pump laser 7 via a 980 / 1550nm wavelength division multiplexing coupler. The pump laser 7 is a 980nm pump laser that matches the wavelength of the doped fiber 5. The power injected into the doped fiber 5 is adjusted to the equilibrium point of near-saturation. This power value is the equilibrium point power that makes the doped fiber 5 work stably in the gain saturation region, so that the fiber works in the gain saturation region. At this time, the fiber exhibits a strong nonlinear compression response to the input signal. A second single-stage optical isolator 6, matching the waveband of the pump source, is also provided between the pump laser 7 and the doped fiber 5.

[0061] The doped fiber 5 of the core computing module can be selected from at least one of doped fiber 5, ytterbium-doped fiber or thulium-doped fiber, to adapt to the optical signal processing requirements of different bands.

[0062] The core design of the core computing module revolves around building a highly efficient, stable, and adaptable all-optical nonlinear activation unit. The selection and connection logic of each component are all based on "achieving accurate gain saturation response," as explained below: The high-concentration doped fiber 5, as the core component of the core computing module, is an erbium-doped medium with an absorption coefficient of 110 dB / m at 1530 nm and a length of only 15 cm. Its core advantage lies in the high efficiency brought by the high doping concentration. It allows the optical signal to fully trigger the gain saturation effect during transmission without the need for excessively long fiber length. This significantly shortens the flight time of the optical signal, reduces transmission delay, reduces dispersion accumulation caused by long-distance fiber, and weakens the interference of environmental disturbances on the optical signal, ensuring the stability and accuracy of the nonlinear response.

[0063] To ensure stable operation of the doped fiber 5 in the gain saturation region, the module uses a 980 / 1550nm wavelength division multiplexing coupler to precisely connect the doped fiber 5 to the 980nm pump laser 7. The 980nm pump wavelength perfectly matches the absorption band of the doped fiber (such as erbium-doped fiber), efficiently providing excitation energy for rare-earth ions within the fiber. Precisely adjusting the pump power to a near-saturation equilibrium point is a key optimized parameter. This power ensures that the number of metastable particles within the fiber reaches saturation, resulting in a strong nonlinear compression response (i.e., gain decreases with increasing input optical power), while avoiding signal distortion due to excessive power or failure to trigger saturation due to insufficient power, thus ensuring consistent nonlinear activation.

[0064] A second single-stage optical isolator 6, positioned between the pump laser 7 and the doped fiber 5, is specifically adapted to the waveband of the pump source. Utilizing the Faraday effect of a magneto-optical crystal, it achieves unidirectional transmission of the optical signal, effectively blocking reflected light from the doped fiber 5 or subsequent links from flowing back into the pump laser 7. This design avoids problems such as mode hopping, output power fluctuations, or frequency shifts in the pump laser caused by reflected light, ensuring stable operation of the pump source and providing continuous and reliable energy support for the gain saturation state of the doped fiber 5.

[0065] In addition, the core computing module supports the selection of various types of optical fibers, such as doped fiber, ytterbium-doped fiber, or thulium-doped fiber. Different doped fibers correspond to different working bands (such as erbium-doped fiber for C / L band, ytterbium-doped fiber for near-infrared band, etc.). Users can flexibly choose according to the actual optical signal processing band requirements, which greatly improves the module's scene adaptability and allows the system to cover more all-optical computing tasks in different bands.

[0066] The signal output module includes a dual-stage isolator 8 that matches the wavelength of the laser, a wavelength demultiplexer 9 (DWDMDEMUX), a multi-channel variable optical attenuator array 10 (VOA), an optical coupler, and a multi-channel optical power meter 11. The dual-stage isolator 8 is disposed in the optical path before the wave demultiplexer 9; The wavelength demultiplexer 9 is used to separate the multi-wavelength mixed optical signal after nonlinear transformation by the doped fiber 5 into independent channel signals. Then, the optical signal of each channel passes through a variable optical attenuator. The variable optical attenuator performs attenuation control based on training weights. The attenuation is determined by the trained output weights. In the inference stage, the weight matrix obtained by training is mapped to the optical attenuation value dB through mathematical transformation, thereby completing the multiplication part of the weighted summation operation in the optical domain. Finally, the processed optical signals are incoherently superimposed through an optical coupler, and the total optical intensity after superposition is read by the multi-channel optical power meter 11 as the prediction output of the neural network.

[0067] The core computing module and the signal output module work together to complete the nonlinear activation and weighting operation: In the core computing module, the doped fiber 5 in the gain saturation region generates a gain saturation response to the input optical signal, forming a nonlinear mapping relationship similar to asymptotic saturation to complete the hidden layer activation. During the inference phase, the control computer 1 converts the trained output weight matrix into attenuation control commands and sends them to the multi-channel variable optical attenuator array 10 of the signal output module. By adjusting the attenuation of each channel's variable optical attenuator, the multiplication part of the weighted summation operation is completed in the optical domain.

[0068] A hybrid architecture for digital training and optical inference is constructed by combining the Extreme Learning Machine (ELM) algorithm with three main modules: The signal input module fixes the input layer weights through linear mapping of optical power to achieve random projection; the core calculation module utilizes the gain saturation characteristics of the doped fiber 5 to complete the hidden layer nonlinear transformation; the output layer weights are calculated using the ridge regression algorithm in the digital domain and then adapted to the variable optical attenuator control requirements of the signal output module.

[0069] The core design of the signal output module is to achieve accurate conversion from "nonlinear optical signal to weighted output". Through component collaboration and algorithm linkage, it completes the final key step of all-optical inference. The functions and collaborative logic of each component are explained as follows: The dual-stage isolator 8, using a wavelength matched to the laser, is positioned before the wavelength demultiplexer 9. Its core function is to completely block back-reflected light. Compared to a single-stage isolator, the dual-stage design provides up to 60dB of isolation, more thoroughly shielding reflected light from subsequent optical paths (such as the wavelength demultiplexer 9 and optical coupler interfaces). This prevents reflected light from flowing back to the doped fiber 5 of the core computing module, thus preventing interference with the stability of its gain saturation state and providing a clean optical signal foundation for subsequent signal separation and weighting.

[0070] Wavelength demultiplexer 9 is a key component for achieving multi-channel signal separation. It can accurately receive multi-wavelength mixed optical signals after nonlinear transformation by doped fiber 5 and split them back into independent channels according to wavelength. This design complements the wavelength division multiplexer 3 in the signal input module, ensuring the integrity of parallel spectral computing and creating conditions for independent weighted control of each channel, allowing the "characteristic signals" corresponding to different wavelengths to be processed individually.

[0071] The multi-channel variable optical attenuator array 10 is the core component for weighted computation in the optical domain. Each channel corresponds to an independent variable optical attenuator, and its attenuation is precisely controlled by the trained output weights. During the inference phase, the output weight matrix calculated in the digital domain using the ridge regression algorithm is transformed into a specific optical attenuation value (unit: dB) through mathematical transformation. The variable optical attenuator adjusts its own attenuation level to directly complete the multiplication part of the weighted summation operation in the optical domain—without photoelectric conversion. This improves computation speed and reduces energy loss, perfectly meeting the needs of all-optical inference.

[0072] The optical coupler and the multi-channel optical power meter 11 constitute the signal aggregation and reading unit at the output end: the optical coupler performs incoherent superposition of the independently weighted optical signals to realize the addition part of the weighted summation operation, and finally forms the total optical signal that reflects the overall reasoning result; the multi-channel optical power meter 11 accurately reads the total optical intensity, converts the physical quantity of the optical domain into a quantifiable digital result, and uses it as the final prediction output of the neural network to complete the closed loop of "optical signal-data result".

[0073] The collaboration between the core computing module and the signal output module is key to all-optical nonlinear activation and weighting: the doped fiber 5 of the core computing module generates a nonlinear response similar to asymptotic saturation in the gain saturation region, completing the extraction of hidden layer features; in the inference stage, the control computer 1 converts the trained weight matrix into attenuation control commands and sends them to the variable optical attenuator array 10. By adjusting the attenuation of each channel, weighting is completed in the optical domain, achieving a seamless connection between "nonlinear activation and optical domain weighting".

[0074] The integrated design with the Extreme Learning Machine (ELM) algorithm further optimizes the logic chain of all-optical inference: the signal input module fixes the input layer weights through linear mapping of optical power to achieve random projection; the core computing module completes the nonlinear transformation of the hidden layer; and the output layer weights are efficiently calculated through the ridge regression algorithm in the digital domain and then precisely adapted to the control range of the variable optical attenuator, forming a hybrid architecture of "digital training to determine weights and optical inference to perform calculations". This avoids the technical difficulties of backpropagation of optical gradients and maximizes the speed and energy efficiency advantages of all-optical computing.

[0075] The specific workflow of the system includes: Data preprocessing: The computer 1 controls the input vector x and the randomly generated input weight matrix W. in Multiply and add a bias b to obtain an intermediate feature vector. Then, linearly scale the intermediate feature vector to a physically achievable milliwatt-level optical power range (e.g., 0.1mW to 1.0mW) to generate a laser drive signal. Optical signal input: The control computer 1 sends the drive signal to the tunable laser array 2 of the signal input module, controls it to output an optical signal with corresponding power, and the optical signal is combined by the wavelength division multiplexer 3, isolated by the single-stage isolator, and then transmitted to the core computing module. Nonlinear activation: The optical signal is sent into the doped optical fiber 5 of the core computing module in the pumping state. Due to the gain saturation characteristics of the optical fiber, the output optical power no longer increases linearly with the input, but exhibits a nonlinear curve similar to a gradual saturation, thus completing the nonlinear activation of the hidden layer. Training phase: The computer 1 is used to collect the nonlinear response matrix H output by the core computing module, and the optimal output weight beta is calculated on the computer using the ridge regression algorithm; Inference phase: The control computer 1 converts the output weight beta into attenuation control instructions for VOA and configures them into the hardware. The system then performs all-optical classification on the newly input unknown data and sends it to the multi-channel variable optical attenuator array 10 of the signal output module to adjust the attenuation of each channel. The optical signal output by the core computing module is isolated by the dual-stage isolator 8 and separated by the wave demultiplexer 9. After attenuation control by the variable optical attenuator of the corresponding channel, it is incoherently superimposed by the optical coupler. Finally, the total optical intensity is read by the multi-channel optical power meter 11 and used as the prediction output of the neural network. When the signal input module uses a tunable laser array with 5 wavelength channels 2 and the core computing module is configured with a high-concentration doped fiber 5, the system can achieve an accuracy of over 90% in the all-optical classification task of the Fashion-MNIST dataset.

[0076] The system's workflow revolves around a hybrid architecture of "digital training - optical inference," proceeding step-by-step according to the logic of "data conversion - optical domain processing - result output." The design logic and technical implications of each step are explained below: The data preprocessing stage is a crucial bridge connecting digital data and optical domain processing. The control computer 1 first multiplies the input vector x with a randomly generated input weight matrix Win and adds a bias b. This step perfectly aligns with the core idea of ​​the Extreme Learning Machine (ELM) algorithm: "randomly initializing input layer weights and biases." Matrix operations naturally achieve random projection of the input data, quickly obtaining intermediate feature vectors without iterative optimization. This feature vector is then linearly scaled to the milliwatt-level optical power range because the output power of the tunable laser array 2 has a physically feasible range. The milliwatt-level range ensures that the optical signal strength meets the requirements of subsequent transmission and processing while avoiding device losses due to excessive power or signal noise interference caused by insufficient power. The resulting laser drive signal achieves a precise mapping between "digital data" and "optical power signal."

[0077] The optical signal input stage is responsible for converting the digital drive signal into a parallel optical signal stream. The drive signal sent by the control computer 1 directly controls the tunable laser array 2, causing each wavelength channel to output an optical signal with the corresponding power, allowing different dimensions of feature data to correspond to different wavelength optical signal carriers. The wavelength division multiplexer 3 combines multiple optical signals of different wavelengths into the same optical fiber link, realizing parallel spectral transmission and significantly improving data transmission efficiency; while the single-stage isolator effectively blocks reflected light from subsequent links, preventing it from flowing back to the laser and affecting output stability, ensuring that the optical signal input to the core computing module is pure and accurate.

[0078] The nonlinear activation stage is the core of all-optical computing, enabling nonlinear extraction of hidden layer features. The highly doped fiber 5, under pumping conditions, operates stably in the gain saturation region under the influence of the pump power in the near-saturation region. At this point, the optical signal, as it propagates through the fiber, exhibits a gain compression effect due to the depletion of metastable rare-earth ion particles. This physical characteristic creates a nonlinear mapping relationship between the output and input optical power, similar to asymptotic saturation. This eliminates the need for optical-to-electrical-to-optical conversion, directly activating the hidden layers of the neural network in the optical domain. This approach retains the high-speed advantage of optical computing while avoiding electronic bottlenecks.

[0079] The core of the training phase is to efficiently solve for the optimal weights of the output layer, which aligns with the fast training characteristics of the ELM algorithm. Computer 1 controls the acquisition of the nonlinear response matrix H output by the core computing module. This matrix is ​​essentially a high-dimensional feature matrix of the input data after random projection and nonlinear activation. The optimal output weight beta is calculated using the ridge regression algorithm in the digital domain. Compared to the iterative optimization of gradient descent in traditional neural networks, ridge regression can be solved directly through matrix operations, resulting in extremely fast training speed and effectively avoiding overfitting, providing accurate weight support for subsequent optical domain inference.

[0080] The inference phase completes the closed loop of "weighted optical signal - output result," realizing the final stage of all-optical inference. The control computer 1 converts the trained output weight matrix into attenuation control commands and sends them to the multi-channel variable optical attenuator array 10. This step maps the digital domain weight values ​​to specific optical attenuation values ​​(dB) through mathematical transformation, enabling the variable optical attenuators to precisely execute the weighted control of each channel, directly completing the multiplication part of the weighted summation operation in the optical domain. The multi-wavelength mixed optical signal output from the core computing module is first completely shielded from reflected light by a dual-stage isolator 8, then split back into independent channels by a wavelength demultiplexer 9. Each channel's optical signal is weighted by its corresponding variable optical attenuator, and then incoherently superimposed (i.e., the addition part of the weighted summation) through an optical coupler. Finally, the total optical intensity is read by a multi-channel optical power meter 11, converting the physical quantity in the optical domain into a quantifiable digital result, which serves as the prediction output of the neural network.

[0081] Experimental results show that when the system is configured with a tunable laser array with 5 wavelength channels and a highly doped fiber, it achieves an accuracy of over 90% in the all-optical classification task of the Fashion-MNIST dataset. This fully demonstrates the feasibility and efficiency of the workflow, leveraging the rapid training advantage of the ELM algorithm while maximizing the high-speed and parallel characteristics of all-optical computing.

[0082] A computer-readable storage medium stores computer code that, when executed, performs the methods described above. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0083] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

[0084] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A computational system for optical neural networks based on nonlinear activation of doped gain media, characterized in that, It includes a signal input module, a core computing module, and a signal output module, which are connected sequentially via an optical transmission link. The signal input module is used to linearly map the preprocessed data matrix into optical power signals of different wavelengths, which are then combined and transmitted to the optical fiber link, while avoiding interference from reflected light on the stability of the signal source. The core computing module includes a doped gain medium using any one of the following: doped fiber, rare-earth doped waveguide, or metal ion doped waveguide. The doped gain medium is connected to a pump laser, and energy is injected by the pump laser to make the doped gain medium work in a preset working range or a controlled working point. The gain working range includes at least one of the unsaturated region, near-saturated region, and saturated region of the doped gain medium. The nonlinear activation transformation of the optical signal is realized by using the gain response of the doped gain medium to the input optical power within the working range. The signal output module is used to separate the multi-wavelength optical signals after nonlinear transformation, adjust the attenuation of each channel optical signal based on the training weights, and then read the output intensity after superimposing the optical signals of each channel as the prediction result of the neural network.

2. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 1, characterized in that, The signal input module includes a control computer, a tunable laser array with a wavelength band matched to the doped gain medium, a wavelength division multiplexer, and a first single-stage optical isolator with a wavelength band matched to the laser. The tunable laser array is C-band and controlled by the control computer. The pre-processed data matrix, including grayscale image pixel values, is linearly mapped to optical power signals of different wavelengths through digital-to-analog conversion. The wavelength division multiplexer has a channel spacing of 100 GHz, which is used to combine different wavelength optical signals output by the tunable laser array into the same optical fiber link. The first single-stage optical isolator in the C-band is placed in the optical path before the core computing module to block reflected light from subsequent links and avoid interfering with the stability of the laser.

3. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 2, characterized in that, The signal input module and the core computing module work together to implement a parallel spectral computing architecture. Multiple optical signals of different wavelengths output from the tunable laser array are combined by the wavelength division multiplexer in the signal input module and simultaneously transmitted and processed in a single doped fiber of the core computing module. Due to the difference in emission / absorption cross sections and the shared inversion particle number, each wavelength signal in the doped gain medium forms a wavelength-dependent nonlinear gain response, which is accompanied by a controllable inter-channel coupling effect. This effect can be suppressed or utilized under appropriate channel spacing and power calibration conditions to enhance the feature mapping capability.

4. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 1, characterized in that, The core computing module consists of a section of highly doped optical fiber. The high-doped fiber is 15 cm long and has an absorption coefficient of 110 dB / m at 1530 nm. The high doping concentration design enables sufficient gain saturation effect within a very short fiber length, which reduces the time-of-flight delay of the optical signal and minimizes instability caused by dispersion and environmental disturbances.

5. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 1, characterized in that, In the core computing module, the doped fiber is connected to the pump laser via a 980 / 1550nm wavelength division multiplexing coupler. The pump laser is a 980nm pump laser that matches the wavelength of the doped fiber. By adjusting the pump power and / or the input optical signal power, the doped fiber can be made to operate in at least one gain operating state in the unsaturated region, near-saturated region, or saturated region, thereby obtaining corresponding linear, weakly nonlinear, or gain-squeezed nonlinear responses in different operating ranges. A second single-stage optical isolator matching the pump source wavelength is also provided between the pump laser and the doped fiber.

6. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 1, characterized in that, The doped fiber of the core computing module can be selected from doped fiber, integrated optical planar waveguide, ridge waveguide or thin film waveguide; its doping ions include at least one of rare earth metal ions or transition metal ions, and there is no restriction on the type of doping. The rare earth metal ions include erbium, ytterbium, thulium, holmium, and neodymium ions, and the transition metal ions include bismuth ions.

7. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 1, characterized in that, The signal output module includes a two-stage isolator with a wavelength matched to the laser, a wavelength demultiplexer, a multi-channel variable optical attenuator array, an optical coupler, and a multi-channel optical power meter. The dual-stage isolator is disposed in the optical path before the wave demultiplexer; The wavelength demultiplexer is used to separate the multi-wavelength mixed optical signal after nonlinear transformation by doped fiber into independent channel signals. Then, the optical signal of each channel passes through a variable optical attenuator. The variable optical attenuator performs attenuation control based on training weights. The attenuation amount is determined by the trained output weights. In the inference stage, the weight matrix obtained by training is mapped to the optical attenuation value dB through mathematical transformation, thereby completing the multiplication part of the weighted summation operation in the optical domain. Finally, the processed optical signals are incoherently superimposed through an optical coupler, and the total light intensity after superposition is read by the multi-channel optical power meter as the classification prediction output of the neural network.

8. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 1, characterized in that, The core computing module and the signal output module work together to complete the nonlinear activation and weighting operation: When the doped gain medium in the core computing module operates in the unsaturated region, near-saturated region, or saturated region, it generates a linear, weakly nonlinear, or gain-saturated nonlinear response to the input optical signal, forming an optical response relationship for linear mapping operators or nonlinear activation operators in the neural network model. During the inference phase, the control computer converts the trained output weight matrix into attenuation control commands and sends them to the multi-channel variable optical attenuator array of the signal output module. By adjusting the attenuation of each channel's variable optical attenuator, the multiplication part of the weighted summation operation is completed in the optical domain.

9. The optical neural network computing system based on nonlinear activation of a doped gain medium according to claim 1, characterized in that, The system is used to implement optical inference or digital training-optical inference hybrid computing architecture for neural network models. The neural network models include, but are not limited to, Extreme Learning Machine (ELM), Feedforward Fully Connected Neural Network (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), LSTM, GRU, Transformer network, Reservoir Computation / Echo State Network (RC / ESN), or combinations thereof. The optical gain response provided by the core computing module is used to implement the activation function or nonlinear operator in the neural network model, and the signal output module is used to implement weight modulation and weighted summation operations.

10. The optical neural network computing system based on nonlinear activation of a doped gain medium according to claim 1 or 9, characterized in that, Using the Extreme Learning Machine (ELM) algorithm as a preferred implementation method, a hybrid architecture of digital training and optical inference is constructed, which is linked with the three main modules: The signal input module fixes the input layer weights through linear mapping of optical power to achieve random projection; the core calculation module utilizes the gain saturation characteristics of doped optical fibers to complete the hidden layer nonlinear transformation; the output layer weights are calculated using the ridge regression algorithm in the digital domain and then adapted to the variable optical attenuator control requirements of the signal output module.

11. The optical neural network computing system based on nonlinear activation of doped gain medium according to claim 1, characterized in that, The specific workflow of the system includes: Data preprocessing: The computer controls the input vector x and the randomly generated input weight matrix W. in Multiply and add bias b to obtain an intermediate feature vector, then linearly scale the intermediate feature vector to the physically achievable milliwatt-level optical power range to generate the laser driving signal; Optical signal input: The control computer sends the drive signal to the tunable laser array of the signal input module, controls it to output an optical signal with corresponding power, and the optical signal is combined by a wavelength division multiplexer, isolated by a single-stage isolator, and then transmitted to the core computing module; Nonlinear activation: The optical signal is fed into the doped optical fiber of the core computing module in the pumping state. Due to the gain saturation characteristics of the optical fiber, the output optical power no longer increases linearly with the input, but exhibits a nonlinear curve similar to a gradual saturation, thus completing the nonlinear activation of the hidden layer. Training phase: The computer is used to collect the nonlinear response matrix H output by the core computing module, and the optimal output weight beta is calculated on the computer using the ridge regression algorithm; Inference phase: The control computer converts the output weight beta into attenuation control instructions for VOA and configures them into the hardware. The system then performs all-optical classification on the newly input unknown data and sends it to the multi-channel variable optical attenuator array of the signal output module to adjust the attenuation of each channel. The optical signal output by the core computing module is sequentially isolated by a double-stage isolator, separated by a wave demultiplexer, and then attenuated by the variable optical attenuator of the corresponding channel. After incoherent superposition by an optical coupler, the total optical intensity is finally read by a multi-channel optical power meter and used as the prediction output of the neural network. When the signal input module uses a tunable laser array with 5 wavelength channels and the core computing module is configured with highly doped optical fiber, the system can achieve an accuracy of over 90% in the all-optical classification task of the Fashion-MNIST dataset.

Images