System and method for realizing positive and negative weighting in optical domain

Abstract

The invention discloses a system and a method for realizing positive and negative weighting in an optical domain, and the system comprises a multi-wavelength light source array which provides multi-channel optical signals, and each wavelength corresponds to an input channel; the adjustable bias photoelectric modulator arrays are in one-to-one correspondence with the light sources, a quiescent working point is biased in a positive / negative slope linear region of a transmission curve through an independent bias circuit, positive and negative weight symbol control is realized, and an input electric signal is loaded; the multi-wavelength micro-ring weighting unit is used for weighting the modulated optical signal by regulating and controlling the transmissivity of the micro-ring; the photoelectric detection unit adopts a single-path detector to receive a composite optical signal and convert the composite optical signal into an electric signal; and the capacitor blocking and signal synthesis unit filters a direct current component, synthesizes an alternating current weighted signal and outputs a real number weighted sum. The system does not need to balance detectors and redundant devices, simplifies hardware architecture, reduces cost and size, reduces power consumption, improves weighting precision and reliability, supports flexible expansion of multi-wavelength channels, adapts to requirements of a large-scale optical neural network, and provides support for optical calculation integration and low-power-consumption landing.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical computing and optical neural network technology, and in particular to a system and method for implementing positive and negative weighting in the optical domain. BACKGROUND

[0002] With the continuous advancement of various high-complexity data processing tasks, a large number of artificial intelligence algorithms are widely used in signal processing, target recognition, cooperative control and other core scenarios to achieve efficient data analysis and decision output, which poses more stringent challenges to the computing power of computing hardware. However, most application scenarios have strict constraints on the speed, power consumption and heat dissipation of the computing system. When processing large-scale parallel data, the traditional electronic computing architecture needs to consume a large amount of electric energy to maintain efficient operation, and the heat dissipation problem of electronic devices will significantly increase the complexity of hardware design. At the same time, the delay and crosstalk problems existing in the process of electronic signal transmission are difficult to match the dual requirements of low latency, low power consumption and high reliability in high-demand scenarios. To break through this technical bottleneck, photonic neural networks have become a key solution. As an optical computing system, photonic neural networks use photons as information carriers and naturally have the core advantages of high parallelism, low power consumption and low latency, which can effectively adapt to the computing needs of multiple fields.

[0003] The core operation logic of neural networks is weighted summation and nonlinear activation, and real number weighting is the key to realizing complex feature extraction and model accuracy. In many classification tasks, the weights are usually required to be distributed in the entire real number domain, positive weights are used to strengthen effective signal features, and negative weights are used to suppress noise or interference features. For example, in the task of channel equalization, negative weights are needed to compensate for the phase shift or attenuation in signal transmission. Therefore, it is crucial for photonic neural networks to efficiently implement real number operations. However, most current photonic neural networks use the intensity of optical signals as the physical carrier, and the value is always non-negative, which cannot directly map the mathematical meaning of positive and negative weights in the electronic domain. This makes the expression and operation of negative numbers in photonic neural networks a core technical problem. If it cannot be solved, photonic neural networks will only be able to implement non-negative weight neural networks, and the model expression ability will be greatly limited, making it difficult to meet the needs of complex tasks.

[0004] Currently, there are three main solutions to the negative weight of photonic neural networks in the industry. The scheme using a balanced detector realizes negative numbers by differentiating two optical signals. The optical signals corresponding to positive and negative weights are input into the two ports of the balanced detector, and the output current of the detector is indirectly obtained by using the characteristic that the output current is proportional to the difference between the two optical intensities. This scheme has high requirements for the consistency of the two optical signals, and the volume and cost of the balanced detector are higher than those of the ordinary single-channel detector, which increases the complexity and power consumption of the system. The scheme of separate calculation of positive and negative numbers splits the weight into positive and negative components, respectively calculates the positive and negative weighted sums using two independent optical links, and finally obtains the total result by subtraction in the electronic domain. This scheme requires double optical modulators, optical delay lines and other devices, which significantly increases the hardware cost and volume. The matrix decomposition scheme decomposes the weight matrix containing negative weights into the product of multiple non-negative matrices through mathematical methods such as non-negative matrix decomposition, and then realizes equivalent negative weight through multi-stage optical calculation, which requires additional calculation overhead and increases the calculation complexity. The transformation mapping scheme adds a fixed offset to the weight to make it non-negative through mathematical transformations such as translation mapping. When calculating, the product of the offset weight and the input is first obtained, and then the product of the offset and the input is subtracted to offset the offset, realizing negative weight. This scheme also requires additional calculation overhead.

[0005] It should be noted that the information disclosed in the above background section is only for understanding the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The main purpose of the present application is to overcome the defects in the background art, provide a system and method for realizing positive and negative weighting in the optical domain, and realize a hardware architecture simplification, low cost, low power consumption and high precision optical neural network real number weighting operation scheme.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] A system for realizing positive and negative weighting in the optical domain, comprising:

[0009] A multi-wavelength light source array is used to provide multiple optical signals of different wavelengths, each wavelength corresponding to an input channel;

[0010] An adjustable bias photoelectric modulator array corresponds one-to-one to the multi-wavelength light source array, each modulator is configured with an independent bias circuit, which is used to bias the static operating point in the positive slope linear region or the negative slope linear region of the transmission curve, to load the input electrical signal and realize the symbol control of positive and negative weights;

[0011] A multi-wavelength micro-ring weighting unit is used to transmit the modulated multi-wavelength optical signal, and the weight size is controlled by the transmission rate of the micro-ring;

[0012] A photoelectric detection unit, using a single photoelectric detector, is used to receive the weighted multi-wavelength composite optical signal and convert it into an electrical signal;

[0013] A capacitance direct-current blocking and signal synthesis unit is used to filter out the direct-current component in the electrical signal and synthesize the alternating-current weighted signals of each channel, outputting a real number weighted sum.

[0014] Further, the adjustable bias photoelectric modulator is a Mach-Zehnder modulator, whose working point is controlled by a bias voltage to be located in the positive slope linear region or the negative slope linear region of the transmission curve, so as to respectively realize the increase or decrease of the output light intensity with the increase of the input electrical signal, to respectively correspond to the positive weight or the negative weight symbol expression.

[0015] Further, the adjustable bias photoelectric modulator is an electro-absorption modulator, which is controlled by a bias voltage to be in a low absorption state or a high absorption state, respectively corresponding to the positive slope region and the negative slope region, to realize the increase or decrease of the output light intensity with the increase of the input electrical signal.

[0016] Further, the system further comprises an optical beam splitter arranged between the adjustable bias photoelectric modulator array and the multi-wavelength micro-ring weighting unit, for splitting and guiding the modulated multi-wavelength optical signal to each independent micro-ring resonator in the micro-ring weighting unit.

[0017] Further, the multi-wavelength micro-ring weighting unit adopts a silicon-based micro-ring array structure, each micro-ring has a designed radius according to the corresponding wavelength, and the transmittance of each wavelength optical signal is independently regulated by changing the micro-ring resonance state, to realize the attenuation weighting of each optical signal.

[0018] Further, the photoelectric detection unit further comprises a transimpedance amplifier for converting the current signal output by the photoelectric detector into a voltage signal, the voltage signal containing a direct-current bias component and an alternating-current modulation component caused by the static working point of the modulator.

[0019] Further, the direct-current blocking capacitance in the capacitance direct-current blocking and signal synthesis unit is connected in series at the output end of the photoelectric detector, for filtering out the direct-current bias component in the voltage signal, and only retaining the alternating-current weighted component related to the input feature and the positive and negative signs of the weight, and then superimposing the alternating-current components of each channel through a synthesis circuit to output a real number weighted sum;

[0020] The real number weighted sum signal output by the capacitance direct-current blocking and signal synthesis unit is used as input to a nonlinear activation unit to complete the complete operation of a single layer in the optical neural network.

[0021] A method for realizing positive and negative weighting in the optical domain using the system, comprising the following steps:

[0022] An array of multi-wavelength light sources is provided, outputting light signals of multiple wavelengths, each wavelength corresponding to an input feature channel;

[0023] The static operating points of each modulator are biased in the positive slope linear region or the negative slope linear region of the transmission curve by the array of adjustable bias photoelectric modulators, so as to correspond to the sign control of positive weight or negative weight, respectively.

[0024] The input electrical signal is loaded to the alternating current port of each modulator, and the modulated light signal is transmitted to the multi-wavelength micro-ring weighting unit for transmittance weighting, and the weight size is independently controlled by the micro-ring transmittance.

[0025] The weighted multi-wavelength composite light signal is received by a single-channel photoelectric detector and converted into an electrical signal containing a direct current component and an alternating current component.

[0026] The current signal is converted into a voltage signal by a transimpedance amplifier.

[0027] The direct current component in the voltage signal is removed by a capacitive direct current separation, and only the alternating current weighting component containing the sign information is extracted.

[0028] The alternating current weighting signals of all channels are synthesized, and the real number weighted sum of the optical neural network is output.

[0029] Further, the step of biasing the modulator in the negative slope linear region specifically includes: setting the operating point of the modulator in the linear region of the falling edge of the transmission curve by a bias voltage, so that the output light intensity decreases with the increase of the input electrical signal, the waveform is reversed, and the negative weight is expressed.

[0030] Further, the step of weighting the light signal by the micro-ring weighting unit specifically includes: dynamically adjusting the refractive index of the micro-ring by thermal tuning, electro-optic modulation or material phase change to continuously change the transmittance of the micro-ring, so as to continuously control the weight size.

[0031] The present application has the following beneficial effects:

[0032] The application provides a system and method for realizing positive and negative weighting in an optical domain, and the core is to construct an integrated system of "adjustable bias photoelectric modulator-multiple wavelength micro-ring-capacitance direct current separation detection", and the independent expression of positive and negative weights is realized by adjusting the static working point (Q point) of the photoelectric modulator, and the real number weighting operation of the optical neural network can be completed without a balanced detector. The bottleneck of the existing hardware redundancy, high complexity and dependence on additional operation of the negative weight realization scheme of the optical neural network (PNN) is broken. The application designs an optical neural network positive and negative weighting realization scheme based on the multiple wavelength micro-ring and modulator bias adjustment, solves the difficulty in expressing the negative weight caused by the non-negativity of the light intensity in the existing optical neural network, and solves the technical problems of high hardware complexity, high cost, high power consumption or low precision of the traditional scheme (such as a balanced detector, separate calculation of positive and negative numbers, etc.); the basic idea is to realize the independent adjustment of the weight size by using the normalized transmittance of the multiple wavelength micro-ring array, to bias the static working point of the photoelectric modulator in the positive slope linear region (corresponding to the positive weight) and the negative slope linear region (corresponding to the negative weight) respectively, to filter out the static light intensity component by combining the single photoelectric detection and the capacitance direct current separation technology, to directly extract the weighted alternating current signal containing positive and negative signs, and to finally realize the real number weighting operation of the optical neural network; the application does not need a balanced detector and redundant photoelectric devices, and the hardware architecture is simplified from the root, not only the hardware cost and volume are reduced, the power consumption is reduced, but also the weighting precision and system reliability are improved, and the multiple wavelength channel can be flexibly expanded, the efficient operation demand of the large-scale optical neural network in different scenes can be adapted, and the core technical support is provided for the integration and low-power landing of the optical computing system.

[0033] Compared with the prior art, the significant effect of the present application is first embodied in the direct technical level brought by the core features: the negative Q point bias feature does not need to rely on the "two-path difference" structure of the balanced detector, but directly realizes waveform inversion by regulating the working point of the modulator, thus avoiding the stringent requirement of the balanced detector on the "consistency of two-path optical devices (modulator, amplifier, detector)", reducing the difficulty of device selection and the complexity of calibration, and reducing the amount of optical devices by 50% (without the need for a backup link); the design of multi-wavelength micro-ring integration enables multi-wavelength signals to be transmitted in parallel in the same optical waveguide, which not only compresses the system size compared with the traditional "single-wavelength multi-link" architecture, but also enables the number of wavelength channels to be flexibly expanded - only the number of micro-rings needs to be increased, without the need for additional expansion space, laying a foundation for the integration of large-scale optical neural networks; in addition, the capacitance direct current isolation feature can filter out the direct current component by only connecting a capacitor with a cost of less than 0.1 yuan, effectively replacing the operational amplifier, comparator and other devices required for "electronic domain subtraction operation" in the traditional scheme, not only significantly reducing the hardware cost, but also avoiding the noise introduced by electronic domain operation, so that the signal-to-noise ratio of the weighted signal is improved; in addition, the use of a single-path detector can receive composite optical signals of all wavelengths with only one photodetector, which greatly reduces the number of detectors (n-channel input only needs one detector, while the traditional scheme needs 2n detectors), and avoids the weighted error caused by the response difference of the multi-path detector, which is conducive to the improvement of calculation accuracy.

[0034] In addition to the direct effect of the core features, the present application also exhibits significant advantages at the overall system level: in terms of hardware cost and size optimization, the system eliminates redundant devices such as balanced detectors, backup optical links and electronic subtraction circuits, not only reducing hardware costs, but also compressing system size, which can adapt to the constraints of "lightweight and miniaturization" of on-board payloads; in terms of power consumption and delay, since there is no need to run real-time calibration algorithms for balanced detectors or perform electronic subtraction operations, the system power consumption and weighted operation delay can be effectively reduced; in terms of reliability and scalability, the reduction of the number of devices and link nodes reduces the system failure rate, and only the addition of micro-rings and modulator arrays is needed to support different sizes of optical neural network models, further improving the adaptability and practical value of the system.

[0035] The present application has important technical value and engineering significance for solving the problem of negative number weighting in photonic neural networks, enabling photonic neural networks to have complete real number operation capability and breaking free from the restriction of non-negative weights.

[0036] Other beneficial effects of the embodiments of the present application will be further described below. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1A multi-wavelength micro-ring positive and negative weighting system architecture diagram of an embodiment of the present application.

[0038] Figure 2 An MZM transmission curve diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0039] The embodiments of the present application will be described in detail below. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present application and its applications.

[0040] It should be noted that when an element is referred to as being “fixed” or “set” on another element, it can be directly on the other element or indirectly on the other element. When an element is referred to as being “connected” to another element, it can be directly connected to the other element or indirectly connected to the other element. In addition, the connection can be for fixing or for coupling or communicating.

[0041] It should be understood that the terms “length”, “width”, “upper”, “lower”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inner”, “outer” and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the embodiments of the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0042] In addition, the terms “first”, “second” are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features referred to. Therefore, the features defined with “first”, “second” can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the meaning of “a plurality of” is two or more, unless otherwise specifically limited.

[0043] The present application aims to solve the problem of existing optical neural networks that are difficult to express negative weights due to the non-negativity of light intensity, and the problem of traditional negative weight schemes (such as balanced detectors, separate calculation of positive and negative numbers) that have high hardware complexity, high cost, high power consumption or low precision, and proposes to construct an integrated system of “adjustable bias optoelectronic modulator-multi-wavelength micro-ring-capacitive direct current separation detection”; by adjusting the static operating point of the modulator (positive / negative slope linear region corresponding to positive / negative weight), the sign control is realized, the micro-ring array controls the weight size, combined with single-channel detection and capacitive direct current separation to extract the weighted signal, without redundant devices, which can reduce the hardware cost and volume, reduce power consumption, improve precision, and also can flexibly expand the wavelength channel to adapt to the demand of large-scale optical neural networks.

[0044] Reference Figure 1The embodiment of the present application provides a system for realizing positive and negative weighting in an optical domain, comprising a multi-wavelength light source array, an adjustable bias photoelectric modulator array, a multi-wavelength micro-ring weighting unit, a photoelectric detection unit and a capacitance direct-current isolation and signal synthesis unit. The multi-wavelength light source array is used for providing a plurality of light signals of different wavelengths, each wavelength corresponding to an input channel; the adjustable bias photoelectric modulator array corresponds to the multi-wavelength light source array one by one, each modulator is configured with an independent bias circuit, which is used for biasing a static operating point in a positive slope linear region or a negative slope linear region of a transmission curve, so as to load an input electrical signal and realize symbol control of positive and negative weights; the multi-wavelength micro-ring weighting unit is used for transmitting the modulated multi-wavelength light signals and weighting the transmittance, and the weight size is controlled by the transmittance of the micro-ring; the photoelectric detection unit adopts a single photoelectric detector, which is used for receiving the weighted multi-wavelength composite light signals and converting them into electrical signals; and the capacitance direct-current isolation and signal synthesis unit is used for filtering direct-current components in the electrical signals and synthesizing alternating-current weighted signals of the channels, and outputting a real number weighted sum.

[0045] In some embodiments, the adjustable bias photoelectric modulator is a Mach-Zehnder modulator, which controls the operating point in the positive slope linear region or the negative slope linear region of the transmission curve through a bias voltage, so as to respectively realize that the output light intensity increases or decreases with the increase of the input electrical signal, so as to correspond to the symbol expression of the positive weight or the negative weight.

[0046] In some embodiments, the adjustable bias photoelectric modulator is an electro-absorption modulator, which controls the low absorption state or the high absorption state through a bias voltage, respectively corresponding to the positive slope region and the negative slope region, so as to realize that the output light intensity increases or decreases with the increase of the input electrical signal.

[0047] In some embodiments, the system further comprises an optical beam splitter arranged between the adjustable bias photoelectric modulator array and the multi-wavelength micro-ring weighting unit, which is used for splitting and guiding the modulated multi-wavelength light signals to each independent micro-ring resonator in the micro-ring weighting unit.

[0048] In some embodiments, the multi-wavelength micro-ring weighting unit adopts a silicon-based micro-ring array structure, each micro-ring has a designed radius according to the corresponding wavelength, and the transmittance of each wavelength light signal is independently controlled by changing the micro-ring resonance state, so as to realize the attenuation weighting of each light signal.

[0049] In some embodiments, the photoelectric detection unit further comprises a transimpedance amplifier, which is used for converting a current signal output by the photoelectric detector into a voltage signal, and the voltage signal comprises a direct-current bias component and an alternating-current modulation component caused by the static operating point of the modulator.

[0050] In some embodiments, the direct-current blocking capacitor in series in the capacitive direct-current blocking and signal synthesizing unit is connected to the output end of the photodetector, for filtering out the direct-current bias component in the voltage signal and retaining only the alternating-current weighted component related to the input feature and the positive and negative signs of the weights, and then superimposing the alternating-current components of each channel through the synthesizing circuit to output the real number weighted sum.

[0051] The real number weighted sum signal output by the capacitive direct-current blocking and signal synthesizing unit is input to a nonlinear activation unit to complete the complete operation of a single layer in the optical neural network.

[0052] Referring to Figure 1 The embodiments of the present application also provide a method for realizing positive and negative weighting in the optical domain using the system, comprising the following steps:

[0053] A multi-wavelength light source array is provided to output light signals of multiple wavelengths, each wavelength corresponding to an input feature channel;

[0054] The static operating points of the modulators are biased in the positive slope linear region or the negative slope linear region of the transmission curve through the adjustable bias photoelectric modulator array, so as to correspond to the sign control of the positive weights or the negative weights, respectively.

[0055] The input electrical signals are loaded to the alternating-current ports of the modulators, and the modulated light signals are transmitted to the multi-wavelength micro-ring weighting unit for transmittance weighting, and the weight size is independently controlled by the micro-ring transmittance.

[0056] The weighted multi-wavelength composite light signal is received by a single-channel photodetector and converted into an electrical signal containing a direct-current component and an alternating-current component.

[0057] The current signal is converted into a voltage signal by a transimpedance amplifier.

[0058] The direct-current component in the voltage signal is filtered out by the capacitive direct-current blocking, and the alternating-current weighted component containing only the sign information is extracted.

[0059] The alternating-current weighted signals of all channels are synthesized to output the real number weighted sum of the optical neural network.

[0060] In some embodiments, the step of biasing the modulator in the negative slope linear region specifically comprises: setting the operating point of the modulator in the linear region of the falling edge of the transmission curve by a bias voltage, so that the output light intensity decreases with the increase of the input electrical signal, realizing waveform inversion, thereby expressing the negative weight.

[0061] In some embodiments, the step of weighting the light signal by the micro-ring weighting unit specifically comprises: dynamically adjusting the refractive index of the micro-ring by heat tuning, electro-optic modulation or material phase change to continuously change the transmittance thereof, thereby realizing continuous control of the weight size.

[0062] The system and method for realizing positive and negative weighting in the optical domain are provided, taking an integrated system of "adjustable bias photoelectric modulator-multiple wavelength micro-ring-capacitance direct current separation detection" as the core, efficiently breaking through the negative weight expression difficulty of optical neural network and the bottleneck of traditional scheme, directly realizing symbol control by regulating the static working point of photoelectric modulator (positive / negative slope linear region corresponding to positive / negative weight), without the "two-way differential" structure of balanced detector, reducing the amount of optical devices by 50% and reducing the selection and calibration complexity; relying on a multiple wavelength micro-ring array, allowing multiple wavelength signals to be transmitted in parallel in the same optical waveguide, compressing the system volume and expanding the channel by only adding micro-rings, laying the foundation for large-scale integration; combining single photoelectric detection (n channels only need 1, and the traditional one needs 2n) with the cost of less than 0.1 yuan of capacitance direct current separation, avoiding the error of multiple detectors, replacing the electronic domain subtraction device to reduce costs, and reducing electronic noise to improve the signal-to-noise ratio. The overall system also adapts to the lightweight demand of on-board load, reduces power consumption and weighting delay, and increases micro-rings / modulators to support different sizes of optical neural networks, improve reliability and scalability, and provide core support for optical computing integration and low-power landing.

[0063] The specific embodiments and experimental verification of the present application are further described below.

[0064] The core of the present application is to construct an integrated system of "adjustable bias photoelectric modulator-multiple wavelength micro-ring-capacitance direct current separation detection", to realize independent expression of positive and negative weights by regulating the static working point (Q point) of the photoelectric modulator, and to complete the real number weighting operation of the optical neural network without balanced detector. The following explains from three aspects of system architecture, core principle and signal processing flow:

[0065] Overall architecture of the system:

[0066] As shown in Figure 1 (multiple wavelength micro-ring positive and negative weighting system architecture diagram), the system mainly includes five modules:

[0067] Multiple wavelength light source array: providing n continuous lasers with different center wavelengths (λ1, λ2,..., λn), each wavelength corresponding to an input feature channel of the optical neural network, and the wavelength interval meeting the channel isolation requirement of the micro-ring filter;

[0068] Bias photoelectric modulator array: one-to-one corresponding to the multiple wavelength light source array, each modulator (such as Mach-Zehnder modulator MZM, electro-absorption modulator EAM) is configured with an independent direct current bias circuit, which can bias the static working point in the positive slope linear region (positive Q point) or the negative slope linear region (negative Q point) of the transmission curve, used to load the input electrical signal (corresponding to the input feature X i of the neural network) and map positive and negative weights;

[0069] Multi-wavelength micro-ring coupling unit: adopt silicon-based micro-ring array structure, couple N modulated wavelength signals to the same optical transmission link (such as silicon-based optical waveguide), control the transmissivity w of the micro-ring i (0≤w i ≤1), realize the attenuation weighting (w i is the normalized transmissivity, only represents the weight size) of the light intensity of each wavelength, and the micro-ring radius is designed according to the wavelength;

[0070] Photoelectric detection unit: adopt a single photoelectric detector (such as PIN photodiode, avalanche photodiode APD), receive the multi-wavelength composite light signal output by the modulator and then weighted by the micro-ring, convert the light intensity signal into a current signal, and then convert it into a voltage signal through a transimpedance amplifier (TIA);

[0071] Capacitive direct current isolation and signal synthesis unit: connect a direct current isolation capacitor C in series at the output end of the photoelectric detector, filter out the direct current component in the signal, and then superimpose the weighted results of all channels through a signal synthesis circuit, and finally output the real weighted sum of the optical neural network.

[0072] Core principle: based on the negative weight realization of the modulator bias and the capacitive direct current isolation.

[0073] The key to breaking through the negative weight expression problem in the application is to utilize the transmission curve characteristics of the photoelectric modulator and the direct current isolation characteristics of the capacitor, which is specifically divided into two branches of “positive weight mapping” and “negative weight mapping”, and the principles are as follows:

[0074] (1) Transmission curve characteristics of photoelectric modulator:

[0075] Taking the commonly used Mach-Zehnder modulator (MZM) as an example, the transmission curve is as shown in Figure 2 (MZM transmission curve diagram), wherein the horizontal axis represents the input voltage V, and the vertical axis represents the output optical power P(V). The relationship between the output light intensity I EOM (t) and the input electrical signal V(t) (including direct current bias V bias and alternating current input V sig (t)) satisfies the following formula:

[0076]

[0077] Among them, I in is the input light intensity of the modulator (from a multi-wavelength light source); T EOM is the inherent transmissivity of the modulator; V bias is the direct current bias voltage (used to set the static working point Q); V sig (t) is the alternating current input voltage (corresponding to the neural network input feature X i , V sig (t)∝X i ); Vπ Let φ be the modulator half-wave voltage (an inherent parameter of the device, typically 0.5–5V). For simplicity, assume the initial phase φ is 0. Figure 2 It can be seen that the MZM transmission curve has two linear regions:

[0078] ● Positive slope linear region (positive Q point): when V bias =-V π At / 2, the operating point is located on the linear segment of the rising edge of the curve, and the output light intensity I is at this time. EOM (t) increases as V(t) increases, and the input voltage can be expressed as V bias +V sig (t), then the output light intensity of the modulator can be expressed as:

[0079]

[0080] ● Negative slope linear region (negative Q point): when V bias =V π At / 2, the operating point is located on the linear segment of the descending edge of the curve, and the output light intensity I is at this time. EOM The input voltage (t) decreases as V(t) increases, achieving waveform reversal. At this time, the input voltage can be expressed as V bias +V π +V sig (t), then the output light intensity of the modulator is:

[0081]

[0082] (2) Capacitor DC blocking and negative weight extraction:

[0083] The microring resonator weights the output light intensity of the modulator, and the absolute value of the weight, |w|, is mapped to the transmittance of the microring. The photodetector then converts the light intensity signal output by the microring into a current signal I. det (t), satisfying I det (t)=R·|w|·I EOM (t)(R is the detector responsivity, in A / W), which is then converted into a voltage signal V by a transimpedance amplifier. det (t), satisfying V det (t)=G·I det (t)=G·R·|w|·I EOM (t)(G is the transimpedance amplifier gain). Therefore, the voltages at the TIA output of the signal biased at the positive and negative Q points are respectively:

[0084]

[0085] V det (t) After connecting the DC blocking capacitor C, the capacitor will filter out the DC component U in the signal.bias (only AC component is reserved), obtaining the weighted voltage signal ±U(t). Finally, the signal synthesis unit superimposes U(t) of all channels, obtaining the weighted sum of the optical neural network Perfectly realize real number weighting operation.

[0086] Signal processing complete flow:

[0087] Take 3-channel input (x1, x2, x3, corresponding weights w1>0, w2<0, w3>0) of a certain layer of the optical neural network as an example, the complete processing flow is as follows:

[0088] 1. Multi-wavelength light source outputs λ1 (corresponding to x1), λ2 (corresponding to x2), and λ3 (corresponding to x3) three lasers;

[0089] 2. Adjust the modulator bias circuit: bias the modulators corresponding to λ1 and λ3 at the positive Q point (V bias =-V π / 2), and bias the modulator corresponding to λ2 at the negative Q point (V bias =V π / 2);

[0090] 3. Input electrical signals x1, x2, and x3 are loaded into the AC ports of the corresponding modulators, and the modulator output light signals are coupled to the same optical waveguide through the micro-ring weighting unit, obtaining the light signal containing weight information

[0091] 4. The composite light signal is received by the photodetector and converted into a voltage signal containing a direct current component

[0092] 5. The direct current component in V det (t) is filtered out by the direct current capacitor C, obtaining the pure AC weighted signal U total (t)=U1(t)-U2(t)+U3(t);

[0093] 6. U total (t) is input to the subsequent nonlinear activation unit to complete one layer of operation of the optical neural network.

[0094] Based on the core principle, the present application can be implemented in the following various embodiments, the main differences are reflected in the device selection, architecture topology and bias control method, all of which can achieve the goal of "no need for balanced photodetectors to realize positive and negative weighting".

[0095] Based on different types of optoelectronic modulators.

[0096] Embodiment 1, Mach-Zehnder modulator (MZM) type:

[0097] As described above, the MZM is used as the core modulator, and the positive and negative weights are realized by adjusting V_bias to -V_π / 2 (positive Q point) and V_π / 2 (negative Q point), which is suitable for scenarios with high requirements for modulation linearity (such as high-precision signal processing). The half-wave voltage V_π of the MZM can be selected as 0.5-2V (low-voltage driving, reducing power consumption).

[0098] Embodiment 2, EAM type:

[0099] The transmission curve (relationship between light absorption coefficient and bias voltage) of the EAM also has a linear region. When the bias voltage is V_low (low absorption, positive slope region), the output light intensity increases with the increase of the input electrical signal (positive weight). When the bias voltage is V_high (high absorption, negative slope region), the output light intensity decreases with the increase of the input electrical signal (negative weight). The advantage of this mode is that the EAM has a smaller size (which can be integrated on a micro ring) and a faster response speed (up to 100GHz), and is suitable for high-speed signal processing scenarios.

[0100] Embodiment 3, micro ring modulator type:

[0101] The modulation electrode is integrated on the micro ring, and the effective refractive index of the micro ring is changed by adjusting the bias voltage, thereby changing the transmittance (transmission curve). When the bias voltage is V_res (negative slope region), the increase of the input electrical signal leads to the decrease of the transmittance (negative weight). When the bias voltage is V_off (positive slope region), the increase of the input electrical signal leads to the increase of the transmittance (positive weight). This mode can realize the integration of “micro ring-modulator”, and has the smallest size (the area of a single unit is less than 100μm 2 ), and is suitable for super large scale integration scenarios.

[0102] Embodiments based on different micro ring architecture topologies.

[0103] Embodiment 4, bus type micro ring architecture:

[0104] The topology of “single bus optical waveguide + multiple groups of micro rings” is adopted, each group of micro rings corresponds to a wavelength channel, the micro rings are connected with the bus waveguide through the coupling area, and the multi-wavelength signals are transmitted in parallel in the bus, and finally received by the single-channel detector at the end of the bus. This architecture has a simple structure and high coupling efficiency, and is suitable for scenarios with a small number of channels (≤64 channels).

[0105] Embodiment 5, mesh type micro ring architecture:

[0106] The topology of "multiple horizontal buses + multiple vertical buses + cross micro-ring" is adopted, each cross micro-ring corresponds to a wavelength input channel, the two-dimensional mapping of "multiple wavelength and multiple input" can be realized, the number of channels can be expanded to more than 1024, and it is suitable for large-scale optical neural networks (such as deep CNN), but the crosstalk at the cross of the micro-ring needs to be optimized.

[0107] Based on the implementation of different bias control modes.

[0108] Embodiment 6, electric bias mode:

[0109] The direct current bias voltage V_bias is output through a digital potentiometer or a DAC (digital-to-analog converter), the working point can be adjusted in real time (such as dynamically calibrating the Q point according to the interstellar environment change), the control precision is high (the voltage resolution is less than 1 mV), and it is suitable for scenes with large environmental fluctuations.

[0110] Embodiment 7, optical bias mode:

[0111] A low-power bias laser (with a wavelength different from that of the signal) is superimposed at the light input end of the modulator, the equivalent working point of the modulator is indirectly adjusted by changing the light intensity of the bias laser (utilizing the photo-thermal effect or the photo-induced refractive index change), without the need of electrical signal wiring, and it is suitable for all-optical integrated systems (avoiding "electro-optical" interference).

[0112] Embodiments based on different wavelength numbers and synthesis modes.

[0113] Embodiment 8, few-wavelength (4-16) parallel mode:

[0114] A DFB laser array is used as a multi-wavelength light source, the wavelength interval is 1 nm, the bus-type micro-ring coupling is adopted, and a PIN diode (low cost) is used as a detector, and it is suitable for small and medium-sized optical neural networks, such as shallow CNN.

[0115] Embodiment 9, multi-wavelength (64-1024) wavelength division multiplexing mode:

[0116] An ultra-continuous spectrum light source + wavelength division multiplexer (WDM) is used to generate a multi-wavelength signal, the wavelength interval is 0.1 nm, the cross-grid type micro-ring architecture coupling is adopted, and an APD (high sensitivity, suitable for weak light signals) is used as a detector, and it is suitable for large-scale optical neural networks, such as the Transformer model.

[0117] The innovation core of the application is to break through the "hardware redundancy, high complexity and dependence on additional operation" bottleneck of the existing optical neural network (PNN) negative weight implementation scheme, and the key difference points revolve around "weight size and sign decoupling control" and "signal extraction without redundant devices", and the specific embodiments are as follows:

[0118] Point 1: The bias of the static operating point (Q point) of the electro-optical modulator realizes the physical control of the weight symbol.

[0119] Limitations of the prior art: The existing negative weight scheme (such as balanced detector difference, matrix decomposition, and transformation mapping) does not directly utilize the transmission curve characteristics of the electro-optical modulator itself to realize symbol control. The balanced detector needs to rely on the difference between two optical signals to indirectly obtain the negative weight, the matrix decomposition needs offline mathematical operation, and the transformation mapping needs electronic domain offset compensation. All of them need additional links or operation steps, which increases complexity and delay.

[0120] Innovation of the present application: Based on the slope characteristics of the linear region of the transmission curve of the electro-optical modulator (such as MZM and EAM), the weight symbol is directly controlled by the bias of the static operating point (Q point): the modulator is biased in the "positive slope linear region" (such as V_bias=-Vπ / 2), and the output light intensity increases with the increase of the input signal, corresponding to the positive weight (+wi); the modulator is biased in the "negative slope linear region" (such as V_bias=+Vπ / 2), and the output light intensity decreases with the increase of the input signal, corresponding to the negative weight (-wi). The symbol control completely relies on the physical characteristics of the modulator itself, without the need for additional difference links or mathematical transformation.

[0121] Technical advantages of the present application: The dependence on balanced detectors and electronic domain subtraction operations is completely eliminated, the symbol control has no additional delay and error, and the hardware complexity is greatly reduced.

[0122] Point 2: Single electro-optical detection + capacitance direct current (DC) separation realizes direct extraction of positive and negative weighted signals.

[0123] Limitations of the prior art: The traditional scheme needs multiple detectors (such as 2n detectors for separate calculation of positive and negative numbers) or complex electronic circuits (such as operational amplifiers for subtraction) to extract weighted signals, which not only increases the cost and size of devices, but also introduces electronic domain noise (such as dark current and operational amplifier offset voltage), reducing the calculation accuracy.

[0124] Innovation of the present application: A single electro-optical detector is used to receive a multi-wavelength composite optical signal, and the light intensity signal is converted into a voltage signal containing a "static light intensity component + alternating weighted component"; then a series of DC separation capacitors are used to filter out the fixed static light intensity component (irrelevant to the weight symbol), and directly retain the alternating weighted component with positive and negative symbols, without additional electronic subtraction operation.

[0125] Technical advantages of the present application: The use of a single detector reduces the amount of detector devices by more than 50%, the DC separation capacitor has no additional noise, and the signal-to-noise ratio and extraction efficiency of the weighted signal are significantly improved.

[0126] Based on the core logic of the present application, only the combination scheme of specific device selection, architecture topology, or regulation mode adjustment can be divided into the following categories:

[0127] 1. Combination scheme based on "different types of optoelectronic modulators".

[0128] Core logic: Keep "micro-ring control weight size + Q-point symbol control + single-path detection + fixed capacitance direct current isolation" unchanged, only replace the type of optoelectronic modulator, specific combinations are as follows:

[0129] Combination 1: Multi-wavelength bus-type micro-ring array + Mach-Zehnder modulator (MZM) + single-path PIN detector + fixed capacitance direct current isolation (adapted to high-precision scenarios, MZM has high linearity and Q-point bias accuracy is easy to control).

[0130] Combination 2: Multi-wavelength grid-type micro-ring array + electro-absorption modulator (EAM) + single-path APD detector + adjustable capacitance direct current isolation (adapted to high-speed scenarios, EAM has a response speed of up to 100GHz, APD is suitable for interstellar weak light environment, and adjustable capacitance is suitable for different wavelength signals).

[0131] Combination 3: Multi-wavelength micro-ring-modulator integrated array (micro-ring integrated modulation electrode) + single-path PIN detector + fixed capacitance direct current isolation (adapted to ultra-large-scale integrated scenarios, micro-ring and modulator are integrated, volume <100μm 2 / cell).

[0132] 2. Combination scheme based on "different micro-ring architecture topologies".

[0133] Core logic: Keep "micro-ring control weight size + Q-point symbol control + single-path detection + fixed capacitance direct current isolation" unchanged, only adjust the architecture of the micro-ring array, specific combinations are as follows:

[0134] Combination 4: Single bus + multi-branch micro-ring array (≤64 channels) + MZM (Q-point electrical adjustment) + single-path PIN detector + fixed capacitance (adapted to small and medium-sized optical neural networks, such as shallow CNN, simple architecture).

[0135] Combination 5: Multi-lateral bus + multi-vertical bus cross micro-ring array (≥1024 channels) + EAM (Q-point optical adjustment) + single-path APD detector + adjustable capacitance (adapted to large-scale optical neural networks, cross micro-ring supports multi-wavelength-multiple-input two-dimensional mapping).

[0136] 3. Combination scheme based on "different Q-point bias control methods".

[0137] Core logic: Keep "micro-ring control weight size + Q-point symbol control + single-path detection + fixed capacitance direct current isolation" unchanged, only adjust the bias control method of Q-point, specific combinations are as follows:

[0138] Combination 6: Bus-type microring array + MZM (Q-point electrical adjustment: DAC dynamically adjusts V_bias, precision <1mV) + single PIN detector + fixed capacitor (adapt to environmental fluctuation scenarios, such as inter-satellite temperature changes, DAC real-time calibration of Q-point, ensure stable linear region)

[0139] Combination 7: Grid-type microring array + EAM (Q-point optical adjustment: auxiliary laser controls photothermal effect, adjusts equivalent V_bias) + single APD detector + adjustable capacitor (adapt to all-optical integration scenarios, no electrical signal wiring interference, Q-point regulation without electromagnetic crosstalk).

[0140] 4. Combination scheme based on "different application scenarios"

[0141] Core logic: keep the core difference point unchanged, adjust the device parameters for different scenes, and the specific combinations are as follows:

[0142] Combination 8: Few-wavelength (4-16 channels) DFB laser source + bus-type microring array + MZM (low Vπ=1V) + single PIN detector + fixed capacitor (adapt to shallow target recognition, low Vπ reduces power consumption, and few wavelengths simplify control).

[0143] Combination 9: Multi-wavelength (64-1024 channels) super-continuous spectrum light source + cross microring array + EAM (high response speed) + single APD detector + adjustable capacitor (adapt to large-scale cooperative scheduling, APD high sensitivity adapts to weak light).

[0144] The core of all combination schemes of the present application is: through the synergistic effect of "multi-wavelength microring control weight size + modulator Q-point symbol control + single detection + capacitor direct current isolation", without balanced detectors and redundant optoelectronic devices, the optical neural network positive and negative weighting is realized.

[0145] In summary, the present application proposes a system and method for realizing positive and negative weighting in the optical domain. The present application designs an optical neural network positive and negative weighting implementation scheme based on multi-wavelength micro-ring and modulator bias control, solves the difficulty of negative weight expression caused by the non-negativity of light intensity in existing optical neural networks, and the technical problems of high hardware complexity, high cost, high power consumption or low precision of traditional schemes (such as balanced detector, positive and negative number separate calculation, etc.); the basic idea is to use the normalized transmittance of the multi-wavelength micro-ring array to realize independent control of the weight size, and at the same time, the static operating point of the optoelectronic modulator is biased in the positive slope linear region (corresponding to the positive weight) and the negative slope linear region (corresponding to the negative weight), combined with single-path optoelectronic detection and capacitance direct-current separation technology to filter out the static light intensity component, directly extract the weighted alternating current signal containing positive and negative signs, and finally realize the real number weighting operation of the optical neural network; the present application does not need balanced detector and redundant optoelectronic devices, which simplifies the hardware architecture from the root, not only reduces the hardware cost and volume, reduces the power consumption, but also improves the weighting precision and system reliability, while supporting flexible expansion of multi-wavelength channels, which can adapt to the efficient operation demand of large-scale optical neural networks in different scenarios, and provides core technical support for the integration and low-power landing of optical computing systems.

[0146] Compared with the prior art, the significant effects of the present application are embodied in the following aspects:

[0147] 1. Direct technical effects brought by core features:

[0148] The effect of the "negative Q point bias" feature: without relying on the "two-path differential" structure of the balanced detector, the waveform inversion is directly realized by controlling the modulator operating point, which avoids the strict requirement of the balanced detector on the "consistency of two-path optical devices (modulator, amplifier, detector)", reduces the difficulty of device selection and calibration complexity, and reduces the amount of optical devices by 50% (without the need for a backup link).

[0149] The effect of the "multi-wavelength micro-ring integration" feature: multi-wavelength signals are transmitted in parallel in the same optical waveguide, which compresses the system volume compared to the traditional "single-wavelength multi-link" architecture, and the number of wavelength channels can be flexibly expanded (only the number of micro-rings needs to be increased, without the need for additional expansion space), laying the foundation for the integration of large-scale optical neural networks.

[0150] The effect of the "capacitance direct-current separation" feature: only a series capacitance (cost <0.1 yuan) is needed to filter out the direct-current component, replacing the operational amplifier, comparator and other devices required for "electronic domain subtraction operation" in traditional schemes, which not only reduces the hardware cost, but also avoids the noise introduced by electronic domain operation, improving the signal-to-noise ratio of the weighted signal.

[0151] The effect of the "single-channel detector" feature: only one photodetector is needed to receive the composite light signal of all wavelengths, compared with the "multi-channel detector + electronic synthesis" structure of the positive and negative number separate calculation scheme, reducing the number of detectors (n-channel input only needs one detector, and the traditional scheme needs 2n detectors), while avoiding the weighting error caused by the response difference of the multi-channel detector, which is conducive to the improvement of calculation accuracy.

[0152] 2. Technical effects of the system as a whole:

[0153] Hardware cost and volume optimization: The system hardware cost is reduced, the volume is compressed, and the constraints of "lightweight and small size" of the on-board load are adapted by eliminating redundant devices such as balanced detectors, backup optical links, and electronic subtraction circuits.

[0154] Power consumption and delay reduction: The system power consumption and weighted operation delay can be reduced without the real-time calibration algorithm of the balanced detector and the electronic domain subtraction operation.

[0155] Reliability and scalability improvement: The number of devices and link nodes is reduced, and the system failure rate is reduced; only the micro-ring and modulator array need to be increased to support different sizes of optical neural network models.

[0156] The realization of the above technical advantages is essentially that the present application breaks through the core bottleneck of the existing optical neural network (PNN) negative weight implementation scheme "hardware redundancy, high complexity, and dependence on additional operation", and the key is to build a new technical logic of "weight size and sign decoupling control" and "no redundant device signal extraction": on the one hand, the traditional scheme relies on an indirect method of additional link or offline operation, directly relying on the slope characteristics of the linear region of the transmission curve of the electro-optical modulator (such as MZM, EAM), and completing the physical control of the weight sign through the static working point (Q point) bias, without the need for differential link or mathematical transformation; on the other hand, replace the multi-channel detection and electronic subtraction circuit with "single-channel photodetector + capacitor direct current isolation", directly extract the AC weighted signal containing positive and negative signs, and reduce the device redundancy and electronic domain noise in principle. This "de-reliance and strong integration" design from the technical root not only is the core innovation of the present application that distinguishes it from existing technologies, but also provides stable logical support for the flexible landing of subsequent multi-scene combination schemes.

[0157] The above further describes the present application in conjunction with specific / preferred embodiments, and cannot be deemed to limit the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, they can make several substitutions or variations to the described embodiments, and these substitutions or variations shall be deemed to fall within the protection scope of the present application. In the description of the present application, the description of the terms "an embodiment", "some embodiments", "a preferred embodiment", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are contained in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In the case of no mutual contradiction, those skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples. Although the embodiments of the present application and their advantages have been described in detail, it should be understood that various changes, substitutions and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A system for implementing positive and negative weighting in the optical domain, characterized in that, include: A multi-wavelength light source array is used to provide multiple optical signals of different wavelengths, with each wavelength corresponding to one input channel; An adjustable bias photoelectric modulator array corresponds one-to-one with the multi-wavelength light source array. Each modulator is configured with an independent bias circuit to bias the static operating point in the positive slope linear region or the negative slope linear region of the transmission curve, so as to load the input electrical signal and realize the sign control of positive and negative weights. The multi-wavelength micro-ring weighting unit is used to perform transmittance weighting on the modulated multi-wavelength optical signal, and the weight is controlled by the transmittance of the micro-ring. The photoelectric detection unit uses a single-channel photoelectric detector to receive weighted multi-wavelength composite optical signals and convert them into electrical signals; The capacitor blocking DC and signal synthesis unit is used to filter out the DC component in the electrical signal and synthesize the AC weighted signal of each channel, and output the real number weighted sum.

2. The system according to claim 1, characterized in that, The adjustable bias photoelectric modulator is a Mach-Zehnder modulator, which controls the operating point to be located in the positive slope linear region or the negative slope linear region of the transmission curve by bias voltage, so that the output light intensity increases or decreases as the input electrical signal increases, respectively, to correspond to the positive weight or negative weight sign expression.

3. The system according to claim 1, characterized in that, The adjustable bias photoelectric modulator is an electroabsorption modulator, which is controlled by bias voltage to be in a low absorption state or a high absorption state, corresponding to the positive slope region and the negative slope region respectively, so that the output light intensity increases or decreases as the input electrical signal increases.

4. The system according to any one of claims 1 to 3, characterized in that, The system also includes an optical beam splitter, which is disposed between the adjustable bias optoelectronic modulator array and the multi-wavelength micro-ring weighting unit, for splitting the modulated multi-wavelength optical signal and guiding it to each independent micro-ring resonator in the micro-ring weighting unit.

5. The system according to any one of claims 1 to 3, characterized in that, The multi-wavelength micro-ring weighting unit adopts a silicon-based micro-ring array structure. Each micro-ring is designed with a radius according to the corresponding wavelength. By changing the resonant state of the micro-ring, the transmittance of each wavelength optical signal is independently adjusted to achieve attenuation weighting of each optical signal.

6. The system according to any one of claims 1 to 4, characterized in that, The photodetector unit also includes a transimpedance amplifier for converting the current signal output by the photodetector into a voltage signal, the voltage signal containing a DC bias component and an AC modulation component caused by the modulator's static operating point.

7. The system according to any one of claims 1 to 5, characterized in that, The DC blocking capacitor in the signal synthesis unit is connected in series with the DC blocking capacitor at the output of the photodetector to filter out the DC bias component in the voltage signal, retaining only the AC weighted component related to the positive and negative signs of the input features and weights. The AC components of each channel are then superimposed by the synthesis circuit to output a real weighted sum. The capacitor blocking DC and the real-valued weighted sum signal output by the signal synthesis unit are used as input to a nonlinear activation unit to complete the full operation of a single layer in the optical neural network.

8. A method for implementing positive and negative weighting in the optical domain using the system according to any one of claims 1-6, characterized in that, Includes the following steps: It provides a multi-wavelength light source array, outputting optical signals of multiple wavelengths, with each wavelength corresponding to an input feature channel; By using an adjustable biased opto-modulator array, the static operating point of each modulator is biased in the positive slope linear region or the negative slope linear region of the transmission curve, so as to correspond to the sign control of positive or negative weights respectively. The input electrical signal is loaded onto the AC port of each modulator, and the modulated optical signal is transmitted to the multi-wavelength micro-ring weighting unit for transmittance weighting. The weight is independently controlled by the micro-ring transmittance. The system receives the weighted multi-wavelength composite optical signal through a single-channel photodetector and converts it into an electrical signal containing both DC and AC components. The current signal is converted into a voltage signal using a transimpedance amplifier; By using capacitors to block DC and filter out the DC component in the voltage signal, the AC weighted component containing only symbolic information is extracted. The AC weighted signals of all channels are synthesized, and the real-valued weighted sum of the optical neural network is output.

9. The method according to claim 8, characterized in that, The step of biasing the modulator in the negative slope linear region specifically includes: setting the modulator operating point in the linear region of the falling edge of the transmission curve by biasing the voltage, so that the output light intensity decreases as the input electrical signal increases, thereby achieving waveform reversal and expressing negative weight.

10. The method according to claim 8 or 9, characterized in that, The specific steps of the micro-ring weighting unit in weighting the optical signal include: dynamically adjusting the refractive index of the micro-ring through thermal modulation, electro-optic modulation, or material phase change to continuously change its transmittance and achieve continuous control of the weight magnitude.

Images