A high-speed optical pulse neural network computing chip architecture for classification tasks

Abstract

The application discloses a high-speed optical pulse neural network computing chip architecture for a classification task, comprising a signal loading unit, including a picosecond light source and a high-speed broadband modulator; the picosecond light source is coupled to an on-chip system and is encoded through the high-speed broadband modulator; a full connection unit distributes the encoded pulse signals to different neurons through a multilayer multimode interferometer, the synapses of the neurons are phase change material weight units, linear weighting of the pulse of each wavelength channel is carried out, and the activation unit of the neuron is realized by a graphene-silicon-based integrated photonic crystal microcavity; a wavelength division device is used to combine the output pulses of each neuron and input to a next full connection unit or an output unit; and the output unit outputs signals through connection to a high-speed detector. The application realizes all-optical computing, reduces the overall power consumption and delay of the network, and improves the integration of a single neuron.

Description

Technical Field

[0001] This invention relates to the field of optical neural networks, and more particularly to a high-speed optical pulse neural network computing chip architecture for classification tasks. Background Technology

[0002] Neural networks, inspired by the information processing mechanisms of biological nervous systems, are powerful machine learning models. However, traditional electron-based artificial neural networks face bottlenecks in terms of computational speed and energy consumption, limited to 1 TOPs / W / mm². 2 Within this range. Compared to neural networks in traditional von Neumann architecture electronic computers, optical neural networks (ONNs) leverage the unique advantages of photons, such as wide bandwidth, low power consumption, high parallelism, and high speed, enabling efficient logical computation and matrix operations. This opens up broad prospects for applications in artificial intelligence, including image recognition, audio classification, and phase transition system analysis, with the potential for ultra-fast processing speed and lower power consumption.

[0003] Current on-chip optical computing architectures, based on modulated continuous-wave light, suffer from low power density, making it difficult to activate the nonlinear properties of materials and achieve all-optical nonlinear activation functions. In 2019, a research team at the University of Münster in Germany pioneered an all-optical integrated and scalable neuromorphic framework for building spiking neural networks, demonstrating that the system can perform typical artificial intelligence pattern recognition tasks. By using ultrafast pulsed light, instantaneous power density can be increased without exceeding the material's thermal damage threshold, while effectively stimulating its nonlinear properties. Therefore, pulsed light is very suitable for realizing all-optical computing architectures. However, current spiking neural network architectures suffer from slow speed, high power consumption, and poor parallelism, hindering large-scale general-purpose expansion. Summary of the Invention

[0004] The purpose of this invention is to address the problems of existing technologies, such as the lack of advantages in response speed and power consumption compared to electrical chips, as well as poor parallelism, by proposing a high-speed optical pulse neural network computing chip architecture for classification tasks.

[0005] The objective of this invention is achieved through the following technical solution: a high-speed optical pulse neural network computing chip architecture for classification tasks, comprising; The signal loading unit includes a picosecond light source and a high-speed broadband modulator; the picosecond light source is coupled to the on-chip system and encoded by the high-speed broadband modulator; The fully connected unit distributes the encoded pulse signal to different neurons through a multi-layer multimode interferometer. The synapses of the neurons are phase change material weighting units, which linearly weight the pulses of each wavelength channel. The activation unit of the neurons is realized by a graphene-silicon-based integrated photonic crystal microcavity. Wavelength division multiplexing devices are used to merge the output pulses of each neuron and input them to the next fully connected unit or output unit. The output unit outputs signals by connecting to a high-speed detector.

[0006] Furthermore, the picosecond light source is a 100GHz high repetition rate, 150fs short pulse picosecond light source, generated by an off-chip fiber femtosecond light source or an on-chip mode-locked laser light source.

[0007] Furthermore, the signal loading unit includes a multi-channel signal loading unit or a spatiotemporally misaligned multiplexed signal loading unit; the multi-channel loading unit splits the input optical pulse, with each wavelength corresponding to a high-speed modulator, and uses multiple high-speed modulators to encode the input value of the signal, which is then combined into a single waveguide. The spatiotemporal misaligned multiplexing signal loading unit also includes an inverse design wavelength division multiplexing device. The optical pulses coupled to the chip are encoded by a high-speed broadband modulator. The inverse design wavelength division multiplexing device splits the encoded optical pulses and then performs spatiotemporal misaligned encoding through waveguide delay before merging them into a single waveguide.

[0008] Furthermore, the inverse design wavelength division multiplexing (WDM) device includes an input waveguide, multiple output port waveguides, and a two-dimensional device functional region located therebetween. The two-dimensional device functional region is set on a silicon-based insulator material platform, and the material distribution of the functional region is determined by an inverse design method: the functional region is discretized into several 20nm*20nm grid cells, and the equivalent dielectric constant of each grid cell varies between two or more preset material values. Iterative optimization is performed with port transmission and crosstalk as the target, thereby obtaining a wavelength division multiplexing or demultiplexing response that satisfies the target wavelength set.

[0009] Furthermore, the phase change material in the neuron is configured with a weighting unit corresponding to each wavelength channel, and is configured as an adjustable transmission type or an adjustable interference type structure. The phase change material weighting unit of the adjustable transmission type structure includes: an on-chip silicon waveguide, a phase change material film covering the evanescent field region of the waveguide, and a thermally or photo-induced modulation structure for phase state writing or erasing. By using the different complex refractive indices of the phase change material in different phase states, the weighting unit can transmit light pulses through the waveguide at different amplitudes, thereby realizing the adjustable weight value.

[0010] Furthermore, the activation unit of the neuron includes: The system comprises a silicon-based photonic crystal defect microcavity structure, an input / output waveguide coupled to the microcavity, and a graphene material layer disposed in the strong field localization region of the microcavity. Electrodes are placed on both sides of the graphene to adjust its carrier concentration, thereby controlling the nonlinear response intensity and threshold. The photonic crystal microcavity is used to provide high field enhancement and high energy density localization, enabling picosecond light pulses to obtain ultrafast nonlinear effects within the microcavity. The graphene material has ultrafast carrier dynamics processes, which are used to generate nonlinear absorption or refractive index change responses on the picosecond scale, thereby making the output nonlinearly mapped to the input, realizing the activation function in the neural network.

[0011] Furthermore, the connection method between the weight units and activation units in the neuron includes: The activation unit maps the activation result to a preset regeneration wavelength, and then combines it with the outputs of other neurons through a wavelength division multiplexing (WDM) device to enter the next layer, or completes nonlinear mapping on the same carrier wavelength and directly enters the next layer.

[0012] On the other hand, the specification also provides a classification method for the aforementioned architecture, including: After preprocessing the data features to be processed, they are mapped to the modulation amount of the on-chip optical signal. In the m-channel spatiotemporal misalignment multiplexing signal loading layer, the femtosecond laser continuously generates N picosecond optical pulses (P1,-P2) according to the optical signal modulation amount. N ), and then coupled to the on-chip system; The signal is loaded by the signal loading layer and sent to the fully connected layer with picosecond response nonlinear activation capability; In a fully connected layer with picosecond response and nonlinear activation capability, pulses are distributed by cascaded MMIs in the signal distribution layer to m different regenerative signal neurons with linear weights and nonlinear activation functions for processing. Each neuron is divided into a synapse and an activation part. The pulse is sent to the WDM and divided into m different wavelengths, which are weighted differently and finally merged into one channel. The pulse passes through an active nonlinear amplifier, the new pulse is nonlinearly activated, and is then combined with the output pulses of other neurons in the signal combining layer via WDM; The pulse is transmitted to the next fully connected layer of the network, which has picosecond-response nonlinear activation capability; After passing through N fully connected layers, the pulse is sent to the signal output layer for final weighting and activation. Finally, the signal is sent directly to the high-speed detector for output.

[0013] Furthermore, the signal loading performed by the signal loading layer includes: The input optical pulse is directly modulated, and then each pulse is divided into m optical pulses of different wavelengths through an integrated wavelength division multiplexing device. The optical signal pulses of different wavelengths are delayed by m delay lines of different lengths to form time division multiplexing of the signal. The encoded pulses are combined into a waveguide. Alternatively, the input light pulse can be split into beams, with each wavelength corresponding to a high-speed modulator. Multiple high-speed modulators can be used to encode the input value of the signal, and the encoded signals can be combined into a single waveguide.

[0014] The beneficial effects of this invention are: 1. This invention uses pulsed light as the signal light, which is advantageous for achieving all-optical activation by exciting the nonlinearity of the device with high instantaneous power. The computation speed of the pulsed light neural network reaches 10 ps, ​​the power consumption of a single multiplication operation can reach 1.875 fJ, and the computing power density can reach 2.13 × 10⁻⁶. 3 TOPs / mm 2 The computing power energy efficiency density can reach 0.71×10 6 TOPs / W / mm 2 It offers a 2-4 order of magnitude improvement over continuous optical neural networks and a significant 2-7 order of magnitude improvement over electronic computing chips.

[0015] 2. This invention utilizes an inverse design wavelength division multiplexer to achieve wavelength division multiplexing, and performs spatiotemporal misalignment coding through waveguide delay. It also achieves multiplexing in the time and frequency domains, thereby improving the parallelism and computational efficiency of the network.

[0016] 3. This invention introduces an inverse-design wavelength division multiplexer, phase change material weighting units, and graphene-silicon-based integrated photonic crystal microcavity nonlinear activation units into the network, achieving all-optical computing, reducing the overall power consumption and latency of the network, and increasing the integration density of a single neuron to 30×25 μm. 2 . Attached Figure Description

[0017] Figure 1 A schematic diagram of a high-speed optical pulse neural network computing chip architecture for classification tasks provided in this embodiment of the invention, using a channel signal loading layer; Figure 2 This is a schematic diagram of a high-speed optical pulse neural network computing chip architecture for classification tasks provided in an embodiment of the present invention, using a spatiotemporally misaligned multiplexed signal loading layer. Detailed Implementation

[0018] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0019] like Figure 1As shown, the present invention provides a high-speed optical pulse picosecond pulse optical neural network computing chip architecture for classification tasks, which consists of a loading layer, a fully connected layer with picosecond response nonlinear activation capability, and an output layer.

[0020] The loading layer includes a high-repetition-rate picosecond light source and a high-speed broadband modulator, and may also incorporate time-division misalignment units. The high-repetition-rate (100 GHz) short-pulse (150 fs) picosecond light source is generated off-chip (fiber femtosecond light source) or on-chip (mode-locked laser light source), coupled to the on-chip system, and then encoded by a balanced broadband high-speed modulator (100 GHz). After passing through an on-chip broadband filter, the pulse is split into multiple beams at 1 nm intervals by an inverse-designed wavelength division multiplexing (WDM) device, then time-spatial misalignment encoding is performed through waveguide delay, and finally combined into a single waveguide by the inverse-designed WDM device.

[0021] If time-division misalignment units are not added, the input optical pulses are split into beams, with each wavelength corresponding to a high-speed modulator. Multiple high-speed modulators are used to encode the input values ​​of the signals, and the encoded signals are then combined into a single waveguide.

[0022] The fully connected layer with picosecond response nonlinear activation capability comprises a signal distribution layer, regenerated signal neurons with linear weights and nonlinear activation functions, and a signal combining layer. Pulses encoded by the last layer are distributed to different neurons for processing via a multi-layer MMI. Each neuron consists of a synapse and an activation function. The pulses are sent to an inversely designed wavelength division multiplexing (WDM) device and split into different wavelengths, weighted differently using a phase-change material weighting module, and then combined into a waveguide via the inverse WDM device. Next, the pulses pass through an all-optical nonlinear activator, where a new wavelength pulse is nonlinearly activated and transmitted to the next layer of the network. By cascading and changing the spacing between the splitting and WDM channels and the regenerated wavelength, the size of the fully connected layer can be arbitrarily changed to realize optical computing functional modules such as matrix compression, pooling, and transformation. After the fully connected operation is completed, the signal is sent to the output layer, a signal fully connected layer directly connected to a high-speed detector for signal output.

[0023] The reverse design WDM achieves spatial separation and beamforming of multiple wavelength channels by comprising an input waveguide, multiple output port waveguides, and a two-dimensional device functional region located therebetween. The functional region is situated on a silicon-based insulator (SOI) material platform, and its material distribution is determined through a reverse design method: the functional region is discretized into several 20nm x 20nm grid cells, with the equivalent dielectric constant of each grid cell varying between two or more preset material values ​​(e.g., silicon and silicon dioxide, or silicon and an equivalent low-refractive-index material). Iterative optimization is performed with port transmission and crosstalk as the objectives, thereby obtaining a wavelength division multiplexing / demultiplexing response that satisfies the target wavelength set {λ1…λm}.

[0024] In one implementation, the optimization objectives include: when the input is incident at wavelength λi, the transmission power at output port i is maximized, and the transmission power at other ports is minimized, to achieve demultiplexing; when using the same device or its dual structure in reverse, the wavelengths can be combined from multiple ports to the same output waveguide to achieve multiplexing. To ensure device manufacturability, minimum feature size constraints, boundary smoothness constraints, and etching deviation robustness constraints can be introduced during reverse design, allowing the resulting wavelength division multiplexing device to be fabricated on-chip using conventional photolithography and etching processes. The channel spacing of the wavelength division multiplexer can be set as needed (e.g., 1 nm), thereby providing a multi-channel carrier basis for subsequent waveguide delay coding, weight loading, and interlayer regeneration.

[0025] The phase change material weighting unit achieves linear weighting of pulses in each wavelength channel by being configured as either an tunable transmission type or an tunable interference type structure. Taking the tunable transmission type as an example, the weighting unit includes: an on-chip silicon waveguide, a phase change material thin film covering the evanescent field region of the waveguide, and a thermally or photo-induced modulation structure for phase state writing / erasing; the phase change material has different complex refractive indices in different phase states, causing the weighting unit to transmit light pulses through the waveguide at different amplitudes, thereby achieving tunability of the weight value w.

[0026] In this invention, each wavelength channel l ii A corresponding weighting unit is set, and the input pulse is divided by the inverse design wavelength division multiplexing device to form the signals of each channel (e.g., x1). l 1 x2 l 2 x3 l 3 x4 l 4 Each channel generates a weighted output through its corresponding weighting unit. Their equivalent relationship can be expressed as: the amplitude or energy of the output light pulse of the i-th channel is proportional to the input value x. i And the product of weights wi is related, that is, to achieve "y i x i l i The multiplicative weighting is then applied; subsequently, the wavelength division multiplexing (WDM) devices are designed in reverse to combine the pulses into the same waveguide, forming a weighted multi-wavelength pulse set (e.g., y1x1). l 1 y2x2 l 2 y3x3 l 3 y4x4 l 4Thus, the phase change material weighting unit realizes the physical correspondence of "weight multiplication" in the neural network, and the weight value can be written / updated by setting the phase state of the phase change material.

[0027] The all-optical nonlinear activator (ANA) can be implemented using a graphene-silicon-based integrated photonic crystal microcavity nonlinear activation unit. This activation unit includes: a silicon-based photonic crystal defect microcavity structure (e.g., a one-dimensional / two-dimensional photonic crystal nanobeam defect cavity or point defect cavity), an input / output waveguide coupled to the microcavity, and a graphene material layer disposed in the strong-field localized region of the microcavity. Electrodes are placed on both sides of the graphene to adjust its carrier concentration, thereby controlling the nonlinear response intensity and threshold. The photonic crystal microcavity provides high-field enhancement and high-energy-density localization, enabling picosecond light pulses to achieve significant ultrafast nonlinear effects within the microcavity. Graphene material exhibits ultrafast carrier dynamics, generating nonlinear absorption / refractive index changes and other responses on the picosecond scale, thus causing the output to exhibit a nonlinear mapping to the input, realizing the activation function in a neural network.

[0028] In the operation of this invention, multi-wavelength pulses, weighted and combined by a weighting unit, are input to the activation unit. When the input pulse energy is below a threshold, the output is approximately linear; when the input pulse energy is above the threshold, the output exhibits saturation / compression or gating characteristics, thereby achieving nonlinear activation. To facilitate interlayer cascading, in one embodiment, the activation unit can map the activated result to a preset regeneration wavelength and combine it with the outputs of other neurons through a wavelength division multiplexing (WDM) device before entering the next layer. In another embodiment, it can also directly enter the next layer after completing the nonlinear mapping on the same carrier wavelength, achieving all-optical inference computation.

[0029] Corresponding to the aforementioned embodiment of a high-speed optical pulse neural network computing chip architecture for classification tasks, the present invention also provides an embodiment of the execution process of the aforementioned architecture.

[0030] The data to be processed is the training dataset, and its input consists of Feature 1 and Feature 2 (which can be preprocessed feature vectors of images / audio / other signals; this embodiment uses two-dimensional features as an example). This feature vector is mapped to the modulation amount of the on-chip optical signal, which is used to drive the subsequent high-speed modulator to complete pulse coding.

[0031] As shown in Figure 1, in the four-channel signal loading layer, a femtosecond laser generates a picosecond light source, which is then coupled to the on-chip system.

[0032] pulse P The light is split into four beams of different wavelengths by an integrated wavelength division multiplexing (WDM) device. l 1 、l 2 、l3 、l 4 ).

[0033] The input value is encoded by four high-speed modulators (x1, x2, x3, x4).

[0034] Encoded pulse (x1) l 1 x2 l 2 x3 l 3 x4 l 4 Combined into a single waveguide via WDM P´ And send it to the next 4×4 fully connected layer.

[0035] In the next layer, pulse P´ The cascaded MMIs are assigned to four different neurons for processing. , , , ).

[0036] Each neuron is divided into a synapse and an activation portion. Pulses are sent to the WDM and split into different wavelengths (x1). l 1 x2 l 2 x3 l 3 x4 l 4 The y1, y2, y3, and y4 are weighted differently and then merged into a single channel (y1x1). l 1 y2x2 l 2 y3x3 l 3 y4x4 l 4 ).

[0037] The pulse passes through an active nonlinear activator (ANA), and a new pulse... It is nonlinearly activated and communicates with the output impulses of other neurons via the WDM. , , ,) merged into P〞 .

[0038] pulse P〞 It is transmitted to the next fully connected layer of the network.

[0039] After three fully connected layers, the pulseP 4 It is sent to the output layer for final weighting and activation.

[0040] Finally, the signal is sent directly to the high-speed detector for output.

[0041] Based on the above-described implementation process of the architecture, the present invention also provides embodiments of different implementation steps for different signal loading layers.

[0042] As shown in Figure 2, the data to be processed is the training dataset, and its input consists of Feature 1 and Feature 2 (which can be feature vectors of images / audio / other signals after preprocessing; this embodiment uses two-dimensional features as an example). This feature vector is mapped to the modulation amount of the on-chip optical signal, used to drive the subsequent high-speed modulator to complete pulse coding. In the m-channel spatiotemporal misaligned multiplexed signal loading layer, the femtosecond laser continuously generates N picosecond optical pulses (P1, -P2). N Then, coupled to the on-chip system, this series of optical pulses will be modulated to load the signal.

[0043] The signal is loaded by a high-speed modulator.

[0044] Waves other than the signal wave are filtered out by the filter.

[0045] Each pulse (P) n The light pulses are split into m different wavelengths (λ1-) by an integrated wavelength division multiplexing (WDM) device. λ m ).

[0046] Optical signal pulses of different wavelengths are delayed by m delay lines of different lengths (t, t+1, ..., t+m-1) to form a time-division multiplexed signal.

[0047] The encoded pulses are combined into a waveguide via WDM and sent to a fully connected layer with picosecond response nonlinear activation capability.

[0048] In a fully connected layer with picosecond response and nonlinear activation capability, pulses are distributed by cascaded MMIs in the signal distribution layer to m different regenerative signal neurons with linear weights and nonlinear activation functions for processing.

[0049] Each neuron is divided into a synapse and an activation portion. Pulses are sent to the WDM and split into m different wavelengths, weighted differently, and finally merged into a single channel.

[0050] The pulse passes through an active nonlinear amplifier (ANA), the new pulse is nonlinearly activated, and is then combined with the output pulses of other neurons in the signal combining layer via WDM.

[0051] The pulse is transmitted to the next fully connected layer of the network, which has picosecond-response nonlinear activation capability.

[0052] After passing through N fully connected layers, the pulse is sent to the signal output layer for final weighting and activation.

[0053] Finally, the signal is sent directly to the high-speed detector for output.

[0054] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0055] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A high-speed optical pulse neural network computing chip architecture for classification tasks, characterized in that, include; The signal loading unit includes a picosecond light source and a high-speed broadband modulator; the picosecond light source is coupled to the on-chip system and encoded by the high-speed broadband modulator; The fully connected unit distributes the encoded pulse signal to different neurons through a multi-layer multimode interferometer. The synapses of the neurons are phase change material weighting units, which linearly weight the pulses of each wavelength channel. The activation unit of the neurons is realized by a graphene-silicon-based integrated photonic crystal microcavity. Wavelength division multiplexing devices are used to merge the output pulses of each neuron and input them to the next fully connected unit or output unit. The output unit outputs signals by connecting to a high-speed detector.

2. The high-speed optical pulse neural network computing chip architecture for classification tasks according to claim 1, characterized in that, The picosecond light source is a 100GHz high repetition rate, 150fs short pulse picosecond light source, generated by an off-chip fiber femtosecond light source or an on-chip mode-locked laser light source.

3. The high-speed optical pulse neural network computing chip architecture for classification tasks according to claim 1, characterized in that, The signal loading unit includes a multi-channel signal loading unit or a spatiotemporally misaligned multiplexed signal loading unit; the multi-channel loading unit splits the input optical pulse, with each wavelength corresponding to a high-speed modulator, and uses multiple high-speed modulators to encode the input value of the signal, which is then combined into a single waveguide. The spatiotemporal misaligned multiplexing signal loading unit also includes an inverse design wavelength division multiplexing device. The optical pulses coupled to the chip are encoded by a high-speed broadband modulator. The inverse design wavelength division multiplexing device splits the encoded optical pulses and then performs spatiotemporal misaligned encoding through waveguide delay before merging them into a single waveguide.

4. The high-speed optical pulse neural network computing chip architecture for classification tasks according to claim 3, characterized in that, The inverse design wavelength division multiplexing (WDM) device includes an input waveguide, multiple output port waveguides, and a two-dimensional device functional region located therebetween. The two-dimensional device functional region is set on a silicon-based insulator material platform. The material distribution of the functional region is determined by an inverse design method: the functional region is discretized into several 20nm*20nm grid cells, and the equivalent dielectric constant of each grid cell varies between two or more preset material values. Iterative optimization is performed with port transmission and crosstalk as the target, thereby obtaining a wavelength division multiplexing or demultiplexing response that satisfies the target wavelength set.

5. The high-speed optical pulse neural network computing chip architecture for classification tasks according to claim 1, characterized in that, The phase change material in the neuron is configured with a weighting unit corresponding to each wavelength channel, and is set as an adjustable transmission type or adjustable interference type structure. The phase change material weighting unit of the adjustable transmission type structure includes: an on-chip silicon waveguide, a phase change material film covering the evanescent field region of the waveguide, and a thermally or photo-induced modulation structure for phase state writing or erasing. By using the different complex refractive indices of the phase change material in different phase states, the weighting unit can transmit light pulses through the waveguide at different amplitudes, thereby realizing the adjustable weight value.

6. The high-speed optical pulse neural network computing chip architecture for classification tasks according to claim 1, characterized in that, The activation unit of the neuron includes: The system comprises a silicon-based photonic crystal defect microcavity structure, input and output waveguides coupled to the microcavity, and a graphene material layer disposed in the strong field localization region of the microcavity. Electrodes are placed on both sides of the graphene to adjust its carrier concentration, thereby controlling the nonlinear response intensity and threshold. The photonic crystal microcavity is used to provide high field enhancement and high energy density localization, enabling picosecond light pulses to obtain ultrafast nonlinear effects within the microcavity. The graphene material has ultrafast carrier dynamics processes, which are used to generate nonlinear absorption or refractive index change responses on the picosecond scale, thereby making the output nonlinearly mapped to the input, realizing the activation function in the neural network.

7. The high-speed optical pulse neural network computing chip architecture for classification tasks according to claim 1, characterized in that, The connection methods of the weight units and activation units in the neuron include: The activation unit maps the activation result to a preset regeneration wavelength, and then combines it with the outputs of other neurons through a wavelength division multiplexing (WDM) device to enter the next layer, or completes nonlinear mapping on the same carrier wavelength and directly enters the next layer.

8. A classification method based on the architecture described in any one of claims 1-7, characterized in that, include: After the data features to be processed are preprocessed, they are mapped to the modulation amount of the on-chip optical signal. In the m-channel spatiotemporal misalignment multiplexing signal loading layer, the femtosecond laser continuously generates N picosecond optical pulses according to the modulation amount of the optical signal, and then couples them to the on-chip system. The signal is loaded by the signal loading layer and sent to the fully connected layer with picosecond response nonlinear activation capability; In a fully connected layer with picosecond response and nonlinear activation capability, pulses are distributed by cascaded MMIs in the signal distribution layer to m different regenerative signal neurons with linear weights and nonlinear activation functions for processing. Each neuron is divided into a synapse and an activation part. The pulse is sent to the WDM and divided into m different wavelengths, which are weighted differently and finally merged into one channel. The pulse passes through an active nonlinear amplifier, the new pulse is nonlinearly activated, and is then combined with the output pulses of other neurons in the signal combining layer via WDM; The pulse is transmitted to the next fully connected layer of the network, which has picosecond-response nonlinear activation capability; After passing through N fully connected layers, the pulse is sent to the signal output layer for final weighting and activation. Finally, the signal is sent directly to the high-speed detector for output.

9. The method according to claim 8, characterized in that, The signal loading performed by the signal loading layer includes: The input optical pulse is directly modulated, and then each pulse is divided into m optical pulses of different wavelengths through an integrated wavelength division multiplexing device. The optical signal pulses of different wavelengths are delayed by m delay lines of different lengths to form time division multiplexing of the signal. The encoded pulses are combined into a waveguide. Alternatively, the input light pulse can be split into beams, with each wavelength corresponding to a high-speed modulator. Multiple high-speed modulators can be used to encode the input value of the signal, and the encoded signals can be combined into a single waveguide.

Images