Silicon-based light-controlled nonlinear activation unit, method and photonic neural network

Abstract

The application provides a silicon-based light-controlled nonlinear activation unit, method and photonic neural network, relates to the technical field of silicon-based photonics, and comprises a photodetector, an electro-optic intensity modulator and a bias / interface circuit. A control signal light is converted into a driving electrical signal by the photodetector, directly drives a silicon-based PIN phase modulation arm, and performs intensity domain nonlinear modulation on independent data signal light. The electro-optic intensity modulator combines segmented electrodes and programmable bias to form a single-threshold or double-threshold nonlinear mapping relationship. The bias circuit and the temperature control module ensure that the device maintains a stable operating point under environmental disturbance. The activation unit can be deployed in a cascaded manner in a multi-layer photonic neural network, and a power management module is introduced between layers to dynamically match the control light and data signal light power, thereby ensuring the consistency of the cross-layer nonlinear response. The scheme is compatible with the CMOS process, has the characteristics of low power consumption, scalability and easy programming, and is suitable for on-chip deep photonic neural networks and efficient edge intelligent computing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of silicon-based photonics, and in particular to a silicon-based optical control nonlinear activation unit, method and photonic neural network. BACKGROUND

[0002] In a photonic neural network, in addition to linear computing units that undertake matrix multiplication and addition, nonlinear activation units are also critical. They are responsible for introducing nonlinear mapping, enabling the network to express complex functions, and are indispensable modules for achieving efficient photonic computing. However, current implementations still rely on traditional electronic devices to complete activation in the electrical domain, which forces the computing link to convert between optical and electrical, making it difficult to support high-density all-optical integration and ultra-low latency processing on a chip.

[0003] In existing research on optical nonlinear activation, programmable design often requires additional bias and external control circuits, which increases system complexity and energy consumption, and also limits the potential for on-chip large-scale integration. At the same time, the trigger threshold of many optical activation schemes is approximately proportional to the input power: when the threshold is high, the network needs to invest more optical power to ensure that the unit enters the target operating area. For a multi-layer cascaded photonic neural network, this not only offsets the attenuation caused by on-chip propagation and device insertion loss, but also raises the overall energy budget, limiting the stackable depth and practicality.

[0004] Therefore, although the related field has made some progress, significantly reducing the activation threshold while maintaining programmability and adapting to single-layer and multi-layer propagation is still a core challenge in this field. Stable working point control, low complexity bias and interface schemes also need to be further explored to balance process realizability and system scalability. SUMMARY

[0005] To overcome the defects of the prior art, the technical problem to be solved by the present application is to provide a silicon-based optical control nonlinear activation unit, method and photonic neural network, which adopts the following technical solutions:

[0006] In one aspect, the present application provides a silicon-based optical control nonlinear activation unit, comprising:

[0007] A photodetector for receiving a control signal light from a linear computing unit and outputting an electrical signal;

[0008] An electro-optical intensity modulator electrically connected to the photodetector, comprising a silicon-based PIN phase modulation arm, for intensity modulating a data signal light under the action of the electrical signal;

[0009] A bias and interface circuit for providing operating bias for the photodetector and the electro-optical intensity modulator and coupling the electrical signal;

[0010] The electro-optical intensity modulator maps phase change into intensity change by phase modulation matching interference or resonance transmission function, and a driving electrical signal obtained by converting the control signal light by the photodetector is loaded to the PIN phase modulation arm to implement nonlinear modulation of light intensity on the data signal light.

[0011] Further improved, at least one of a graphene layer or a III-V semiconductor layer is hetero-integrated on the waveguide of the PIN phase modulation arm to improve the refractive index modulation efficiency and reduce the activation threshold.

[0012] Further improved, the electrode of the PIN phase modulation arm is of a segmented structure, including a first carrier injection segment and a second carrier injection segment driven by independent electrodes, for forming a double-threshold window and obtaining a mapping relationship between the output light power and the control signal light power of the segmented nonlinearity.

[0013] Further improved, the electro-optical intensity modulator integrates a micro-heating electrode and a temperature sensing resistor to form a closed-loop temperature control module, for controlling the wavelength detuning of the resonance / interference operating point within ±5pm and the phase drift within ±0.02π.

[0014] Further improved, the P-pole of the photodetector is electrically connected to the P-pole of the PIN phase modulation arm, the N-pole of the photodetector is connected to the positive pole of a DC power supply through a current-limiting resistor, and the N-pole of the PIN phase modulation arm is grounded through a bypass capacitor, forming a reverse series topology, the electrical time constant τ of which is determined by the current-limiting resistor and the bypass capacitor and is in the range of 10ns-1µs.

[0015] Further improved, the biasing and interface circuit includes a voltage scanning unit and an amplification limiting unit, the voltage scanning unit scans the bias of the PIN phase modulation arm in the range of 0-5V and writes the lock value, so as to set the operating point and select the target nonlinear curve interval; and the amplification limiting unit is used to limit the peak value of the electrical signal generated by the photodetector.

[0016] Another aspect of the present application provides a photonic neural network, including at least one layer of the silicon-based light-controlled nonlinear activation unit as proposed in any one of the above, and a silicon-based linear matrix calculation unit optically connected before and after the silicon-based light-controlled nonlinear activation unit; wherein each layer of the silicon-based light-controlled nonlinear activation unit receives the control signal light of the previous layer, nonlinearly modulates the data signal light provided by an independent light source, and outputs to the next layer of linear matrix calculation unit.

[0017] Further improved, it also includes an interlayer power management module for dynamically setting the power of the data signal light of the corresponding independent light source according to the power of the control signal light of each layer, so that the power of the data signal light and the power of the control signal light are positively correlated and fall within the target threshold interval, to compensate for cascade attenuation and improve the consistency of the multi-layer output curve.

[0018] The application also provides an optical control nonlinear activation method, comprising:

[0019] S1: converting the control signal light from the linear computing unit into an electrical signal;

[0020] S2: processing the electrical signal to generate a driving signal, wherein the driving signal forms a double-threshold operating window through threshold programmable segmented driving or equivalent control;

[0021] S3: using the carrier dispersion effect to load the driving signal to the PIN phase modulation arm to intensity modulate the data signal light provided by the independent light source, and obtain a nonlinear activation output optical signal;

[0022] S4: implementing closed-loop control on the operating point of the PIN phase modulation arm, and outputting the nonlinear activation output optical signal to the next linear computing unit.

[0023] Further improved, in step S2, the bias parameter of the segmented driving is searched and the lock value is written in the range of 0-5V through programmable voltage scanning, so as to select the target operating point of single threshold or double threshold, so that the output curve can be repeatedly switched between monotonic nonlinearity and segmented nonlinearity without changing the device structure.

[0024] Compared with the prior art, the application has the following beneficial effects:

[0025] Firstly, by splitting the control signal light and the data signal light, the control light is directly driven to the phase modulation arm after photoelectric conversion, and the data light is modulated and output in the intensity modulator, realizing the physical decoupling of the control link and the data link, which not only shortens the electro-optical conversion path, reduces the homologous power consumption and signal crosstalk, but also significantly reduces the threshold power and link transmission delay of the nonlinear activation, so that the unit is more suitable for on-chip photonic neural networks requiring multi-layer cascade.

[0026] Secondly, by using the PIN phase modulation arm with segmented electrodes and locking the bias point through the voltage scanning unit, the activation function curve can be flexibly switched between single threshold and double threshold modes without changing the physical structure, so as to obtain different segmented nonlinear mapping relationships. This programmability enables the neural network to dynamically reconstruct the activation function according to the specific task requirements, greatly improving the adaptability and parameter adjustment efficiency of the network.

[0027] Thirdly, in the present application, the electro-optic intensity modulator integrates the micro-heating electrode and the temperature sensing resistor to form a closed-loop temperature control module, so that the wavelength detuning of the resonance / interference operating point is controlled within ±5pm and the phase drift is controlled within ±0.02π, thereby controlling the wavelength detuning and the phase drift of the operating point within a very small range and effectively suppressing the influence of thermal drift on the nonlinear response. Meanwhile, the reverse series topology is adopted, in which the P-pole of the photodetector is electrically connected to the P-pole of the PIN phase modulation arm, the N-pole of the detector is connected to the positive pole of the power supply through a current-limiting resistor, and the N-pole of the phase arm is connected to the ground through a bypass capacitor, so that the electrical time constant τ is located in the interval of 10ns-1µs, effectively limiting the current overshoot and suppressing high-frequency noise, and the combination of the two guarantees the high consistency and repeatability of the output response of the nonlinear activation unit.

[0028] Fourthly, in the present application, the graphene or III-V material is heterogeneously integrated, which effectively improves the refractive index modulation efficiency and further reduces the activation threshold; meanwhile, the interlayer power management module at the system level can dynamically match the optical power and compensate for the cascade attenuation, so as to realize excellent scalability while ensuring low power consumption. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0030] Figure 1 A structure schematic diagram of a silicon-based light-controlled nonlinear activation unit provided by the present application;

[0031] Figure 2 A circuit schematic diagram of the electrical connection relationship between the photodetector and the PIN phase modulation arm in a silicon-based light-controlled nonlinear activation unit provided by the present application;

[0032] Figure 3 A flowchart of a light-controlled nonlinear activation method provided by the present application. DETAILED DESCRIPTION

[0033] In order to facilitate the understanding of those skilled in the art, the structure of the present application will be further described in detail in combination with the embodiments and the drawings:

[0034] In the description of the present application, the terms "first", "second" are only used for description purpose, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. The orientation or positional relationship indicated by the terms "part", "side", "end" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the indicated device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application.

[0035] As shown in Figure 1 A silicon-based light-controlled nonlinear activation unit, comprising:

[0036] A photodetector for receiving a control signal light from the linear calculation unit and outputting an electrical signal;

[0037] An electro-optical intensity modulator electrically connected to the photodetector, comprising a silicon-based PIN phase modulation arm for intensity modulation of data signal light under the action of the electrical signal;

[0038] A bias and interface circuit for providing operating bias for the photodetector and the electro-optical intensity modulator and realizing coupling of the electrical signal;

[0039] Wherein, the electro-optical intensity modulator maps the phase change to the intensity change through the phase modulation cooperating with the interference or resonance transfer function, the driving electrical signal obtained by converting the control signal light through the photodetector is loaded to the PIN phase modulation arm to implement the optical intensity nonlinear modulation on the data signal light.

[0040] As shown in Figure 1As shown, in a specific embodiment, the above-mentioned silicon-based light-controlled nonlinear activation unit is arranged between two adjacent linear matrix calculation units. The control signal light enters the photodetector through the on-chip coupling structure to complete the photoelectric conversion, and the output driving electrical signal is directly loaded to the PIN phase modulation arm under the action of the bias and interface circuit; the data signal light is injected into the electro-optic intensity modulator by an independent light source or a pre-allocated signal channel, and intensity modulation is realized in the high slope region of the transfer function thereof. In order to reduce crosstalk and power competition, the control signal light and the data signal light preferably use separate input and independent waveguides; as a preferred embodiment, a wave separation organization with different center wavelengths, such as 1310 nm or 1550 nm, can be set on the input side, and the physical decoupling is realized by a demultiplexer in the chip. The photodetector and the PIN phase modulation arm are arranged in close proximity on the layout to reduce parasitic and improve direct drive efficiency. The electro-optic intensity modulator can use a Mach-Zehnder interference structure or a micro-ring resonant structure: the former introduces intensity variation through the phase difference of the two arms, and the latter changes the transmission intensity by shifting the resonant condition, both of which meet the mechanism requirements of phase modulation and intensity mapping. The static operating point is preferably placed in the high slope region of the transfer function to obtain significant intensity variation under small phase disturbance, or it can be placed in the near threshold or compression region according to the application to form different nonlinear patterns.

[0041] In a specific embodiment, at least one of a graphene layer or a III-V semiconductor layer is heterogeneously integrated on the waveguide of the PIN phase modulation arm to improve the refractive index modulation efficiency and reduce the activation threshold.

[0042] As a preferred embodiment, the PIN phase modulation arm is locally windowed to remove the coating layer above it, a graphene thin layer is deposited or transferred above the waveguide and an ohmic contact is formed, or a III-V functional layer is introduced above the waveguide by bonding / epitaxy transfer, so that the electrically induced refractive index change has a stronger overlap with the optical field mode. The hetero layer is preferably arranged in the core area of the phase-sensitive segment, covering part of the total length of the phase arm, so as to improve the modulation efficiency while controlling the process complexity and loss. This material enhancement scheme does not change the unit link, and is only an optional implementation to reduce the driving and threshold power.

[0043] In a specific embodiment, the electrodes of the PIN phase modulation arm are of a segmented structure, including a first carrier injection segment and a second carrier injection segment driven by independent electrodes, for forming a double-threshold window and obtaining a mapping relationship between the output optical power and the control signal optical power of the segmented nonlinearity.

[0044] In a specific embodiment, two sections of injection region are arranged in sequence along the light propagation direction, and each section is connected to an independent DC bias port and superimposed with the driving electrical signal on the photodetector side. By scanning the DC bias of the two sections in the range of 0-5V and locking to the target value, a segmented curve of "low power holding - medium power gain jump - high power recompression" can be constructed in the same device structure, thereby realizing a single threshold or double threshold programmable operating mode.

[0045] As a preference, the length ratio of the two sections can be selected in the range of 1:1-1:3, and the lateral coupling between the electrodes is reduced by isolation resistance or grid wiring to avoid interference caused by the fluctuation of the bias of one section on the other section; if necessary, a short passive waveguide section can be inserted between the two sections to further isolate the electrical and thermal crosstalk.

[0046] The electro-optical intensity modulator integrates a micro-heating electrode and a temperature sensing resistor to form a closed-loop temperature control module, which is used to control the wavelength detuning of the resonance or interference operating point within ±5pm and the phase drift within ±0.02π.

[0047] The temperature control module takes the phase arm adjacent temperature or reference transmission offset as the feedback quantity, and uses a PID or equivalent control strategy to adjust the heating current of the micro-heating electrode, so that the static operating point remains stable under environmental changes or self-heating conditions.

[0048] As a preference, the micro-heating electrode can be made of TiN, Poly-Si or metal thin film; the temperature sensing resistor is arranged in thermal coupling with the phase arm to improve the measurement representativeness. In the scanning optimization phase, the fixed thermal parameters are kept, and after locking, the closed-loop compensation mode is entered; when it is detected that the temperature control is close to the saturation region, the control logic can trigger a small range bias re-locking to restore the margin.

[0049] As shown in Figure 2 , the P pole of the photodetector is electrically connected to the P pole of the PIN phase modulation arm, the N pole of the photodetector is connected to the positive pole of the DC power source through a current-limiting resistor R, and the N pole of the PIN phase modulation arm is grounded through a bypass capacitor C, forming a reverse series topology, the electrical time constant τ of which is determined by the current-limiting resistor R and the bypass capacitor C, and is in the range of 10ns-1µs.

[0050] In the above embodiment, the topology makes the detector work in the reverse bias region and the phase arm work in the injection bias point, and the common P node simplifies the wiring and reduces the loop area. The RC time constant composed of the current-limiting resistor R and the bypass capacitor C defines the electrical port bandwidth, suppresses high-frequency noise and rising edge spikes, and improves the smoothness and repeatability of the nonlinear response. As a preference, the current-limiting resistor R and the bypass capacitor C are arranged close to the PIN phase arm on the layout to shorten the return path; the common node and the ground ring use low-resistance metal channels to reduce parasitic inductance; for the MZI structure, the metal counterweights can be used on both arms to weaken the phase shift caused by asymmetric parasitics.

[0051] The biasing and interface circuit includes a voltage scanning unit and an amplification limiting unit. The voltage scanning unit scans the bias of the PIN phase modulation arm in the range of 0-5V and writes a locking value, so as to set the working point and select the target nonlinear curve interval; the amplification limiting unit is used for limiting the peak value of the electric signal generated by the photodetector.

[0052] Specifically, the voltage scanning unit can be composed of a DAC, a sample and hold, and comparison logic, and works in two stages of coarse scanning and fine scanning: the coarse scanning quickly positions the threshold interval with a larger step, the fine scanning accurately determines the locking point with a small step, and the locking value is written into the holding unit for long-term operation. In order to adapt to device aging and environmental drift, the control logic can trigger a small range review within a preset period; if the threshold value window is deviated, the locking value is fine-tuned. The amplification limiting unit can use a transimpedance amplifier to convert the photoelectric current signal generated by the photodetector into a voltage signal, and cooperate with a clamping network in the subsequent stage to limit the peak value of the output voltage; another implementation is to use an active amplifier with internal automatic gain control or limiting function, which realizes peak limiting while completing signal amplification, prevents curve drift or temperature rise accumulation caused by overdrive saturation; the limiting threshold is set according to the device voltage resistance and target dynamic range, and is coordinated with the temperature control closed loop strategy. It should be understood that the selection criteria of the target working point include but are not limited to: making the output-input relationship curve closest to the shape of the target activation function (such as Sigmoid function), or obtaining the maximum nonlinear gain in a specific control optical power interval.

[0053] Another aspect of the present application provides a photonic neural network, which includes at least one layer of the silicon-based light-controlled nonlinear activation unit as proposed above, and a silicon-based linear matrix calculation unit optically connected before and after the silicon-based light-controlled nonlinear activation unit; wherein each layer of the silicon-based light-controlled nonlinear activation unit receives the control signal light of the previous layer, and outputs the data signal light provided by an independent light source to the next layer of the linear matrix calculation unit after nonlinear intensity modulation.

[0054] The activation unit and the linear matrix calculation unit are sequentially spliced by interlayer interconnection waveguides to form a deep structure, and the output port of each layer and the input port of the lower layer are connected by low-loss transition taper and large-radius bending. Each layer of data signal light is provided by an independent light source, and the power is set under the control of the interlayer power management module, so that each layer enters its nonlinear target interval; the light signal output by the activation unit is directly used as the input of the next layer of linear operation. In order to reduce the interlayer thermal crosstalk and back reflection, isolation grooves can be arranged around the resonant structure, and a micro-inclination or surface roughening is used at the interface to suppress the reflection return.

[0055] The interlayer power management module is used for dynamically setting the power of the data signal light of the corresponding independent light source according to the power of the control signal light, so that the power of the data signal light is positively correlated with the power of the control signal light and falls within a target threshold interval, to compensate for cascade attenuation and improve the consistency of the multi-layer output curve.

[0056] In the above embodiment, the interlayer power management module can work in an open-loop lookup table mode or a closed-loop calibration mode. In the open-loop mode, the system establishes a mapping table of control power-target data power when it is shipped or power-on self-test; during operation, the corresponding data signal light power is set according to the control power of each layer. In the closed-loop mode, a small amount of monitoring light can be introduced between the layers, and the on-chip monitoring PD is used to feedback the output error and fine-tune the light source drive, so that the actual output falls within the target threshold window. The two modes can work together with the temperature control closed loop to ensure the consistency of the cross-layer curve position and slope.

[0057] As shown in Figure 3 The present application also provides a light control nonlinear activation method, comprising the following steps:

[0058] S1: converting the control signal light from the linear calculation unit into an electrical signal. The control signal light is coupled into a photodetector to complete the conversion from light to electricity. To ensure the quality of the driving electrical signal, the bias and interface circuit provide stable reverse bias and low noise loop on the PD side, and the electrical signal is subjected to baseline drift suppression and bandwidth shaping when necessary, so that its amplitude and spectrum are adapted to subsequent direct drive loading.

[0059] S2: processing the electrical signal to generate a driving signal, and forming a double-threshold operating window through threshold programmable segmented driving or equivalent control. The voltage scanning unit optimizes the two DC bias ports of the segmented electrode, scans in the range of 0-5V according to the preset strategy (coarse scan-fine scan), selects the target operating point of single threshold or double threshold according to the relationship curve of output optical power-control power, and writes the lock value. The processing logic allows different lock levels to be set for the two segments, to form a gain mutation in the medium power area and construct a compression or relinearity section in the high power area. To offset temperature drift and aging, a periodic micro-relocking process can be provided.

[0060] S3: using the carrier dispersion effect to load the driving signal to the PIN phase modulation arm, to modulate the data signal light provided by the independent light source in intensity, and obtain a nonlinear activation output light signal. The driving signal is directly loaded to the PIN phase modulation arm, causing changes in refractive index and optical path, which are converted into intensity changes of the data signal light through interference or resonant transfer function. In the MZI structure, intensity modulation is achieved by changing the phase difference between the two arms; in the micro-ring structure, the transmission is significantly changed by shifting the resonance center. The static operating point is set according to the target curve shape to ensure that the desired nonlinear response is obtained within the target power interval.

[0061] S4: The working point of the PIN phase modulation arm is closed-loop controlled, and the nonlinear active output optical signal is output to the next linear calculation unit. The temperature sensing resistor and the micro-heating electrode form a feedback loop to compensate for environmental fluctuations and self-heating effects in real time, so that the working point is kept within the aforementioned locking interval; the output light is sent into the next linear unit through the waveguide to complete the interlayer cascade. If necessary, a small-angle anti-reflection or local coating can be provided at the output port to further stabilize the interlayer coupling.

[0062] In step S2, the bias parameters of the segmented drive are searched and the locking value is written by programmable voltage scanning in the range of 0-5V to select the target working point of single threshold or double threshold, so that the output curve can be repeatedly switched between monotonic nonlinearity and segmented nonlinearity without changing the device structure.

[0063] To improve reliability, the scanning can use a reference verification parallel strategy: one curve is used for optimization, and the other curve is used for verification of curve stability after locking with a small perturbation. Whether the locking value needs to be fine-tuned is judged by the differential result. The locking data can be stored in the on-chip non-volatile area or the upper controller cache to support power-off memory and batch consistency. If it is detected during operation that the temperature control enters saturation or the limiting amplitude is frequently triggered, the control logic can trigger a small-scale compensation scan to update the locking point and record the log for subsequent maintenance and tracing.

[0064] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A silicon-based optically controlled nonlinear activation unit, characterized by, Comprising: a photodetector for receiving control signal light from a linear computing unit and outputting an electrical signal; an electro-optical intensity modulator electrically connected to the photodetector, comprising a silicon-based PIN phase modulation arm for intensity modulation of data signal light under the action of the electrical signal; a bias and interface circuit for providing operating bias for the photodetector and the electro-optical intensity modulator and realizing coupling of the electrical signal; wherein the electro-optical intensity modulator maps phase changes to intensity changes through a phase modulation cooperating interference or resonance transfer function, and the driving electrical signal obtained by converting the control signal light through the photodetector is loaded to the PIN phase modulation arm to implement optical intensity nonlinear modulation on the data signal light.

2. The unit of claim 1, wherein, At least one of a graphene layer or a III-V semiconductor layer is heterogeneously integrated on the waveguide of the PIN phase modulation arm to improve refractive index modulation efficiency and reduce the activation threshold.

3. The unit of claim 1, wherein, The electrodes of the PIN phase modulation arm are of a segmented structure, comprising a first carrier injection segment and a second carrier injection segment driven by independent electrodes, for forming a double-threshold window and obtaining a mapping relationship between the output optical power and the control signal light power of the segmented nonlinearity.

4. The unit of claim 3, wherein, The electro-optical intensity modulator integrates a micro-heating electrode and a temperature sensing resistor to form a closed-loop temperature control module for controlling the wavelength detuning of the resonance / interference operating point within ±5pm and the phase drift within ±0.02π.

5. The unit of claim 1, wherein, The P pole of the photodetector is electrically connected to the P pole of the PIN phase modulation arm, the N pole of the photodetector is connected to the positive pole of a direct current power supply through a current limiting resistor, and the N pole of the PIN phase modulation arm is grounded through a bypass capacitor, forming a reverse series topology, the electrical time constant τ of which is determined by the current limiting resistor and the bypass capacitor and is in the range of 10ns-1µs.

6. The unit of claim 5, wherein, The bias and interface circuit includes a voltage scanning unit and an amplification limiting unit, the voltage scanning unit scans the bias of the PIN phase modulation arm within the range of 0-5V and writes a lock value, thereby setting the operating point and selecting the target nonlinear curve interval; The amplification limiting unit is used to limit the peak value of the electrical signal generated by the photodetector.

7. A photonic neural network, characterized by Comprising at least one layer of the silicon-based light-controlled nonlinear activation unit as claimed in any one of claims 1-6, and a silicon-based linear matrix computing unit optically connected before and after it; wherein each layer of the silicon-based light-controlled nonlinear activation unit receives the control signal light of the previous layer, nonlinearly intensity modulates the data signal light provided by an independent light source, and outputs to the next layer of linear matrix computing unit.

8. The photonic neural network of claim 7, wherein, Further comprising an interlayer power management module for dynamically setting the power of the data signal light of the corresponding independent light source according to the power of each layer of control signal light, so that the power of the data signal light and the power of the control signal light remain positively correlated and fall within the target threshold interval, to compensate for cascade attenuation and improve the consistency of the multi-layer output curve.

9. A light controlled nonlinear activation method, characterized by, Comprising: S1: converting the control signal light from a linear computing unit into an electrical signal; S2: processing the electrical signal to generate a driving signal, the driving signal forming a double-threshold working window through threshold-programmable segmented driving or equivalent control; S3: loading the driving signal to a PIN phase modulation arm through carrier dispersion effect, intensity-modulating a data signal light provided by an independent light source to obtain a nonlinearly activated output light signal; S4: implementing closed-loop control on the working point of the PIN phase modulation arm and outputting the nonlinearly activated output light signal to a next linear calculation unit.

10. The method of claim 9, wherein, In step S2, the bias parameter of the segmented driving is searched and the lock value is written through programmable voltage scanning in the range of 0-5V to select the target working point of single threshold or double threshold, so as to repeatedly switch the output curve between monotonic nonlinearity and segmented nonlinearity without changing the device structure.

Images