CN117787369B

CN117787369B - Optical computing system, complex value detection method and data processing method

Info

Publication number: CN117787369B
Application number: CN202410210847.6A
Authority: CN
Inventors: 李红珍; 李辰; 张新; 姜金哲
Original assignee: Suzhou Metabrain Intelligent Technology Co Ltd
Current assignee: Suzhou Metabrain Intelligent Technology Co Ltd
Priority date: 2024-02-27
Filing date: 2024-02-27
Publication date: 2024-05-10
Anticipated expiration: 2044-02-27
Also published as: CN117787369A

Abstract

The application relates to an optical computing system, a complex value detection method and a data processing method, wherein the system comprises a controller, a detector and two optical networks in mirror symmetry, the two optical networks comprise a first MZI, a second MZI and a third MZI, and the two optical networks are used for receiving input signal light with the same polarization; the first MZI is used for dividing the signal light into a plurality of paths of signal light to be processed; the second MZI is used for carrying out phase modulation on the signal light to be processed to obtain calculated signal light; the third type of MZI is used for generating reference light according to signal light to be processed, and interfering the reference light with calculated signal light to obtain detection signal light, wherein the phase difference of the reference light of the two optical networks is pi/2; the controller is used for configuring the same interference parameters for the second type MZIs at the mirror symmetry positions of the two optical networks after acquiring the complex value calculation instruction; the detector is used for complex value detection. The complex value detection efficiency can be improved by adopting the system.

Description

Optical computing system, complex value detection method and data processing method

Technical Field

The present disclosure relates to the field of optical computing, and in particular, to an optical computing system, a complex value detection method, and a data processing method.

Background

Optical computing refers to techniques and equipment that employ optical methods to implement arithmetic processing and data transmission. The Optical Neural Network (ONN) is a novel network formed by utilizing optical technologies, such as optical connection technology and optical device technology, and is mainly used for realizing a neural network model by utilizing optical computing technology. The optical neural network can increase the deep learning speed by nearly two orders of magnitude, as it outsources matrix multiplication operations, which are widely used in conventional neural networks, to application specific photonic integrated circuits (Photonic Integrated Circuits, PIC). One architecture of the dedicated PIC uses a Mach-zehnder interferometer (Mach-Zehnder Interferometer, MZI) as a core element, and a plurality of MZI cascades form a complex network. The special PIC can realize complex value ONN operation which cannot be executed by the traditional digital electronic platform, namely, the amplitude and the phase are coded at the same time.

An important function of the dedicated PIC in the complex-valued ONN operation is to implement complex-valued detection, including intensity detection and coherent detection, to obtain the amplitude and phase of the output optical signal, respectively. However, the conventional technical scheme uses an on-chip coherent detection line for coherent detection in complex-value detection, and two-step detection is required to obtain phase data, so that complex-value detection efficiency is not high.

Disclosure of Invention

In view of the above, it is desirable to provide an optical computing system, a complex value detection method, and a data processing method that can improve the complex value detection efficiency.

In a first aspect, the present application provides an optical computing system, including a controller, a detector, and two optical networks mirror-symmetrical in a topological relation, where the two optical networks each include a first MZI, a second MZI, and a third MZI, and the two optical networks are configured to receive input signal light with the same polarization; the first MZI is used for dividing the signal light into a plurality of paths of signal light to be processed and providing the signal light to the second MZI and the third MZI; the second type of MZI is used as a computing node for carrying out phase modulation on the signal light to be processed to obtain computing signal light, and the computing signal light is provided for the third type of MZI; the third type of MZI is used for generating reference light according to the signal light to be processed, and interfering the reference light with the calculated signal light to obtain detection signal light, wherein the phase difference of the reference light output by the third type of MZI of the two optical networks is pi/2;

the controller is used for configuring the same interference parameters for the second type MZIs at the mirror symmetry positions of the two optical networks after acquiring the complex value calculation instruction; the detector is used for respectively receiving the detection signal lights output by the two optical networks during complex value calculation to carry out complex value detection so as to determine the complex value calculation result.

In some embodiments, in the topology, multiple layers are included in sequence in a direction from the near to the far of the symmetry line in two optical networks, with the number of MZI layers decreasing in sequence.

In some embodiments, the number of MZI layers per layer is sequentially reduced by one in the direction from the near to the far of the symmetry line for both optical networks.

In some embodiments, one MZI closest to the symmetry line in the two optical networks is taken as a first MZI, and a third MZI includes the output side MZI and the first MZI, and the one MZI furthest from the symmetry line is taken as a last MZI; starting from the last MZI, each MZI comprises a first MZI and an output side MZI, starting from the last MZI, and starting from the last MZI, each MZI further comprises at least one second MZI, wherein the first MZI and the output side MZI are respectively positioned at the head end and the tail end of each layer, and the second MZI is positioned in the middle of each layer.

In some embodiments, the number of first-type MZI in each layer MZI is the same as the number of output-side MZI.

In some embodiments, the number of first-type MZIs and the number of output-side MZIs in each layer of MZIs are both 1.

In some embodiments, the interference parameters include an amount of phase modulation of a first interference arm of the MZI and an amount of phase modulation of a second interference arm of the MZI in the two optical networks, the amount of phase modulation of the second interference arm being the same, the amount of phase modulation of the first interference arm of the output side MZI in the mirror symmetric position being set to be the same, the amount of phase modulation of the second interference arm being the same.

In some embodiments, the amount of phase modulation of the first interference arm of the first layer MZI of one of the two optical networks is set to 0 and the amount of phase modulation of the first interference arm of the first layer MZI of the other network is set to pi/2.

In some embodiments, the detector comprises a photo detector connected to the output side MZI of the two optical networks, respectively, and a data processing device connected to the photo detector; the photoelectric detector is used for respectively converting detection signal light output by the two optical networks into photocurrent; the data processing device is used for determining the phase difference between the input signal light and the calculated signal light according to the photocurrents corresponding to the two optical networks so as to perform complex value detection and obtain a complex value calculation result.

In some embodiments, two outputs of an output side MZI in each layer MZI are each connected to one photodetector, and two photodetectors connected to the same output side MZI are connected in balance.

In some embodiments, the system further comprises a grating coupler, through which the two optical networks are respectively connected to the photodetectors.

In some embodiments, the optical computing system further comprises a fourth class of MZI that connects the first layer MZI of the two optical networks.

In some embodiments, the number of fourth class MZI is 1 more than the number of second class MZI in the first layer MZI.

In some embodiments, the interference parameter of the fourth class MZI is set to 0.

In some embodiments, the controller is further configured to configure different interference parameters for at least two different fourth-class MZI after the real-valued calculation instructions are obtained.

In some embodiments, the controller is further configured to configure different interference parameters for at least one pair of MZI of the second type at mirror symmetric locations of the two optical networks after obtaining the real-valued calculation instruction; the detector is also used for respectively receiving the detection signal lights output by the two optical networks to carry out real value detection during real value calculation so as to determine the result of the real value calculation.

In some embodiments, when the total number of inputs of the second type of MZI of the network is the sum of the numbers of the first type of MZI and the third type of MZI and the number of the second type of MZI are as follows:

wherein, is greater than or equal to 2.

In some embodiments, the optical computing system further comprises a light source and an optical splitter connected to the light source, the optical splitter further being connected to the two optical networks, respectively, wherein the light source is configured to transmit the original signal light to the optical splitter; the optical splitter is used for splitting the original signal light into two paths of input signal light with the same polarization and providing the two paths of input signal light to the two optical networks.

In a second aspect, the present application provides a complex value detection method, which is applied to the optical computing system according to any embodiment of the first aspect, and the complex value detection method includes: acquiring a first photoelectric current value which is a value of photocurrent converted from detection signal light output by one of two optical networks; acquiring a second photoelectric current value which is a value of photocurrent converted by signal light output by the other of the two optical networks; and determining the phase difference between the input signal light and the calculated signal light according to the first photoelectric current value and the second photoelectric current value so as to perform complex value detection and obtain a complex value calculation result.

In some embodiments, determining the phase difference between the reference light and the calculated signal light from the first and second light current values includes: calculating the ratio of the first photoelectric current value to the second photoelectric current value; a phase difference between the input signal light and the calculated signal light is determined based on the ratio.

In a third aspect, the present application provides a data processing method applied to the optical computing system as any one of the embodiments in the first aspect, the data processing method comprising: acquiring data to be processed; modulating input signal light according to data to be processed, and inputting the input signal light into two optical networks trained in advance to obtain detection signal light corresponding to the two optical networks; determining a phase difference between input signal light and calculated signal light according to the detection signal light corresponding to the two optical networks;

and determining a data processing result according to the phase difference between the input signal light and the calculated signal light.

In a fourth aspect, the present application provides a training method of an optical neural network, which is applied to an optical computing system according to any embodiment of the first aspect, where the training method of the optical neural network includes: acquiring sample data for training; modulating input signal light according to sample data, inputting two initialized optical networks in an optical computing system, and obtaining detection signal light corresponding to the two optical networks; determining a phase difference between input signal light and calculated signal light according to the detection signal light corresponding to the two optical networks;

Determining a data processing result according to a phase difference between the input signal light and the calculation signal light; determining a calculated value of the loss function according to the data processing result of each round; adjusting interference parameters of MZIs in the two optical networks according to the calculated value of the loss function until the calculated value of the loss function meets the termination condition; and acquiring and storing interference parameters of each MZI in two optical networks in the optical computing system when the termination condition is met, and terminating training.

The optical computing system, the complex value detection method and the data processing method respectively input the input signal light with the same polarization into two optical networks which are in mirror symmetry in topological relation, respectively separate the input signal light into corresponding computing signal light and reference signal light through the two optical networks, further interfere the corresponding computing signal light with the reference light to obtain interfered detecting signal light, and the interference parameters of the second MZI at the symmetrical position are the same when complex value detection is carried out, so that the separated signal light executes the decomposition of the identical unitary weight matrix when passing through the second MZI, the computing signal light output by the two optical networks is the same, and the phase difference of the reference light output by the third MZI of the two optical networks is pi/2, so that after the two optical networks are correspondingly interfered, the detecting signal light is converted into photocurrent, and the photocurrent of one network is identical with that of the other networkIn proportion to the photocurrent of another networkProportional to the ratio of the total number of the components, wherein,For inputting signal light and calculating phase difference of signal light, the corresponding photocurrent of two optical networks can be used for extractingTo achieve complex value detection. Compared with the prior art, when complex-valued detection is performed, two detection and extraction of phase information are required, the optical computing system of the application can output two paths of detection signal light simultaneously through two optical networks when complex-valued detection is performed, and obtain the photoelectric current values corresponding to the two paths of detection signal light simultaneously, so that the optical computing system can extract according to the photoelectric currents corresponding to the two optical networksThe efficiency of complex value detection is improved.

Drawings

FIG. 1 is a block diagram of an optical computing system in some embodiments;

FIG. 2 is a schematic diagram of the internal structure of a MZI in some embodiments;

FIG. 3 is a schematic diagram of the topology of two optical networks in some embodiments;

FIG. 4 is a schematic diagram of a topology of a MZI network according to the prior art;

Fig. 5 is a schematic diagram of the topology of two optical networks in other embodiments;

FIG. 6 is a schematic diagram of the topology of two optical networks including a fourth type of MZI in some embodiments;

FIG. 7 is a block diagram of an optical computing system in alternative embodiments;

Fig. 8 is a schematic diagram of the topology of two optical networks in other embodiments;

FIG. 9 is a flow chart of a complex value detection method in some embodiments;

FIG. 10 is a flow chart of a method of data processing in some embodiments;

FIG. 11 is a flow chart of a training method of an optical neural network in some embodiments;

Fig. 12 is an internal block diagram of a controller in some embodiments.

Detailed Description

In order to make the technical scheme and advantages of the present disclosure more apparent, the embodiments of the present disclosure and related technical contents are further described in detail below with reference to the accompanying drawings and the description. It should be understood that the following description is only for illustrating the technical solutions of the embodiments of the present disclosure, and is not intended to limit more possible implementations of the present disclosure.

It is noted that the relational terms such as "first," "second," and the like, herein may be used solely to distinguish one from another article, state, or action without necessarily indicating, implying a relative importance or order relationship. The terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that the object that is included may not be limited to the objects listed herein. The term "plurality" or other variants are used to denote a number of objects of two or more.

In a first aspect, as shown in fig. 1, an embodiment of the present application provides an optical computing system 10, where the optical computing system 10 includes a controller 101, two optical networks 102 (including a first optical network 1021 and a second optical network 1022) that are mirror symmetric in topology, and a detector 103, where the first optical network 1021 and the second optical network 1022 each include a first MZI, a second MZI, and a third MZI. Wherein the first optical network 1021 and the second optical network 1022 are configured to receive input signal light of the same polarization.

The first type MZI is used for dividing signal light into a plurality of paths of signal light to be processed and providing the signal light to the second type MZI and the third type MZI.

The second type of MZI is used as a computing node for carrying out phase modulation on the signal light to be processed to obtain computing signal light, and the computing signal light is provided for the third type of MZI.

The third type of MZI is used for generating reference light according to the signal light to be processed, and interfering the reference light with the calculated signal light to obtain detection signal light, wherein the phase difference of the reference light output by the third type of MZI of the two optical networks is pi/2.

The controller is used for configuring the same interference parameters for the second type of MZI at the mirror symmetry positions of the two optical networks after acquiring the complex value calculation instruction.

The detector is used for respectively receiving the detection signal lights output by the two optical networks to carry out complex value detection so as to determine the complex value calculation result.

Among them, MZI (Mach-Zehnder Interferometer ) is an interferometer that can be used to observe the relative phase shift changes generated by the different paths and the medium after splitting the light beam emitted from the individual light sources into two collimated light beams.

Specifically, the MZI of the present application includes two inputs, two outputs, two couplers, and two sets of interference arms. One of each group of interference arms is provided with an adjustable phase shifter, and the adjustable phase shifter is provided with interference parameters which can be configured or adjusted by a controller according to external electric signals. The signal light output by the output end can be controlled by modulating the interference parameter of the adjustable phase shifter. In one embodiment, a schematic diagram of the internal structure of the MZI is shown in fig. 2. In fig. 2, the MZI includes an input 21, an input 22, an output 23, and an output 24.

Referring to fig. 3, the optical computing system of the present application includes a first optical network 31 and a second optical network 32, where the first optical network 31 and the second optical network 32 are mirror symmetrical about a symmetry line 33. Each of the two optical networks includes a plurality of MZI, and MZI ports not used are connected to each other by the optical waveguide 34, and are also connected to the optical waveguide 34.

In the present application, two optical networks that are mirror symmetrical in topology refers to the mirror symmetry of the connection relationship of each MZI in the two optical networks. As shown in FIG. 3, 6 MZIs numbered 1-6 are the first type of MZI, 6 MZIs numbered 15-20 are the second type of MZI, and 8 MZIs numbered 7-14 are the third type of MZI. Wherein MZI number 1 and MZI number 4 are mirror symmetric, MZI number 15 and MZI number 20 are mirror symmetric, and so on.

Further, in fig. 3, the signal light emitted by the light source 35 is processed to obtain two paths of input signal light with the same polarization, which are respectively provided to the first optical network 31 and the second optical network 32, and the corresponding detection signal light 36 and the detection signal light 37 are respectively output after passing through the first optical network 31 and the second optical network 32.

Further, the first type MZI is mainly used for performing spectral processing on the input signal light to obtain multiple paths of signal light to be processed, and the multiple paths of signal light are provided for the second type MZI and the third type MZI. The second type of MZI is used as a computation node, and is mainly used for performing unitary weight matrix decomposition to obtain computation signal light, and the computation signal light is provided for the third type of MZI. The unitary weight matrix decomposition refers to decomposing an input n×n unitary weight matrix into a plurality of MZI implementations, where N is a positive integer greater than or equal to 2. The third type of MZI processes the received signal light to be processed into reference light, and further interferes the received calculation signal light with the reference light to obtain detection signal light.

The controller is used for setting different interference parameters for the MZIs in the network in different application scenes so as to adapt to different scene requirements. For example, in a complex-valued computing scenario, the same interference parameters are configured for the second type of MZI at mirror symmetric locations of the two optical networks. In other computing scenarios, interference parameters that are adapted to other computing scenarios may be configured by the controller.

In the present application, the complex-value detection may include intensity detection for obtaining the amplitude of the output signal light and coherent detection for obtaining the phase information of the output signal light. In general, coherent detection is typically implemented using homodyne detection techniques. The basic principle of the homodyne detection technology is that the electric field product is obtained by modulating signal light and reference light interference, and the exact position of the phase difference in the range is judged according to the cosine and sine relation of the phase difference. The homodyne detection technique requires that the signal light and the reference light come from the same light source, have the same frequency and polarization, and require four sets of interference light signals, and are converted into two sets of photocurrents by a balance detector, and the phase difference is extracted according to the ratio.

Based on the principle of the homodyne detection technology, the application can process the input signal light through two optical networks in the optical computing system, output detection signal light, further detect the photocurrents of the detection signal light corresponding to the two optical networks through the detector, and extract the phase difference between the input signal light and the calculation signal light according to the photocurrents corresponding to the two optical networks so as to realize complex value detection.

In the conventional technology, on-chip coherent detection circuits are generally used for realizing complex-value detection, and two-step detection is required to obtain phase data. Referring specifically to fig. 4, when complex-valued detection is performed on the target optical path, the original signal light emitted from the light source 41 is separated into the reference light and the signal light. The separated signal light passes through the MZI of the intermediate region and then outputs modulated signal light. The reference light is transmitted from the MZI denoted by 43 along the MZI of the edge to a certain MZI of the edge (e.g., the MZI denoted by 44 in fig. 4), and finally interferes with the task-modulated signal light at the MZI denoted by 44 and outputs the detection signal light 42, the interference condition being determined by the phase difference between the input signal light and the modulated signal lightIt is decided that different interference conditions lead to a difference in photocurrent amplitude.

Wherein, in the first step, the interference parameter of the MZI with the reference number of 43 is set to be 0, the interference parameter of the MZI with the reference number of 44 is set to be pi/2, the maximum interference occurs between the reference light and the signal light, and the photocurrent is extracted,Proportional toCosine values of (2); second step of detecting, the interference parameter of MZI with reference number 43 is set to pi/2, the interference parameter of MZI with reference number 44 is kept unchanged, and photocurrent is extracted，Proportional toAccording to the sine value of (2)AndThe ratio can be determined。

In the application, because the phase difference of the reference light output by the third MZI of the two optical networks is pi/2 and the second MZI of the mirror symmetry position of the two networks is configured with the same interference parameter, the detection signal light can be output through the upper half network and detected by the detector to obtain. Meanwhile, outputting another detection signal light through the lower half part of the network, and detecting the detection signal light through a detector to obtainFurther according toAndCan be determined to. Therefore, compared with the prior art, the application can detect and extract in one stepThe efficiency of complex value detection is improved.

With continued reference to FIG. 3, the number of MZIs per layer decreases from the symmetry line position to 7, including 4 layers in sequence.

With continued reference to FIG. 3, the number of MZIs per layer decreases one after another from the symmetry line position to the direction numbered 7.

In some embodiments, one MZI closest to the symmetry line in the two optical networks is taken as a first MZI, and a third class of MZI includes the output side MZI and the first MZI, and the one MZI furthest from the symmetry line is taken as a last MZI;

Starting from the last MZI, each MZI comprises a first MZI and an output side MZI, starting from the last MZI, and starting from the last MZI, each MZI further comprises at least one second MZI, wherein the first MZI and the output side MZI are respectively positioned at the head end and the tail end of each layer, and the second MZI is positioned in the middle of each layer.

With continued reference to FIG. 3, each network includes 4 layers, the last layer having a MZI number of 1, the second-to-last layer having a MZI number of 2, including a first type MZI and an output side MZI, the third-to-last layer having a MZI number of 3, including a first type MZI, a second type MZI and an output side MZI, and the last layer having a MZI number of 4, including a first type MZI, two second type MZIs and an output side MZI. The MZI of each layer is connected to the MZI in the previous layer and the MZI in the next layer.

As can be seen from fig. 3, the MZI arrangements in the two optical networks in the present application respectively form a triangular array, and by using such a triangular array, it is possible to implement decomposition of any nxn unitary weight matrix onto a plurality of cascaded MZI.

In the application, the calculated signal light output by each network can comprise multiple paths, the output reference light can also comprise multiple paths, and each path of calculated signal light and the reference light can finally interfere at the output MZI.

The number of the first MZIs in each layer of MZIs is the same as that of the MZIs on the output side, so that each path of calculated signal light can have corresponding reference light to interfere with the corresponding reference light, and corresponding detection signal light is obtained. So that at least one path of detection signal light can be output through each network in the application.

In some embodiments, when the total number of inputs of the second type MZI of the network isWhen the sum of the numbers of the first MZI and the third MZI isNumber of MZIs of the second typeThe relationship of (2) is as follows:

Wherein, Greater than or equal to 2.

Referring to fig. 5, as shown in fig. 5, signal light emitted from a light source 51 is converted into two paths of input signal light, which are provided to two optical networks with mirror symmetry, and the two optical networks are respectively used to output corresponding detection signal light 52 and detection signal light 53. The arrangement form of the second type MZIs in each network forms a triangular network, according to a triangular decomposition algorithm, any N×N unitary weight matrix can be decomposed into N (N-1)/2 unitary rotation matrices to be multiplied continuously, and each unitary rotation matrix corresponds to one second type MZI, so that N (N-1)/2 second type MZIs are needed for realizing the N×N unitary weight matrix for the N×N unitary weight matrix. Where N represents the total number of inputs to the second type of MZI of the network.

Illustratively, for a 3×3 unitary weight matrix, the decomposition followsWherein, the method comprises the steps of, wherein,、AndHas the following form:

,,

Wherein, Each of which isAndCan be modulated separately. Thus, for a 3×3 unitary weight matrix, each triangular network needs to configure 3 MZI of the second type, the entire network needs to configure 6 MZI of the second type, and the total number of MZI of the third type and MZI of the first type that the entire network needs to configure is 14.

In the application, the highest dimension of the unitary weight matrix executable by the second MZI can be judged according to the number of input ports and output ports of the triangular network formed by the second MZI. For example, when there are 3 pairs of input and output ports, the matrix corresponding to the highest dimension that can be performed is a3×3 unitary weight matrix.

In the application, one of each group of interference arms of the MZI is provided with an adjustable phase shifter, and the adjustable phase shifter is provided with interference parameters which can specifically comprise the phase modulation quantity of the first interference armAnd the amount of phase modulation of the second interference arm. First layer MZI of two optical networksThe difference is configured to be pi/2,Configured identically. The interference parameters at the other symmetrical positions are respectively the same.

By making the difference between the phase modulation amounts of the first interference arms of the first MZIs in the two optical networks pi/2, the difference between the phases of the reference light output by the third MZIs of the two optical networks pi/2 is made. By configuring the interference parameters of the MZIs at the symmetrical positions to be identical, the calculated signal lights of the two optical networks are identical, so that the identical calculated signal lights are respectively interfered with two reference lights with pi/2 phase difference to output detection signal lights, and the detection signal light output by one network is converted into photocurrent and then has the same size as that of the photocurrentIn proportion to the detected signal light output by another network, which is converted into photocurrentProportional to the ratio of the total number of the components, wherein,For inputting signal light and calculating phase difference of signal light, extracting according to corresponding path photocurrent of two optical networksComplex value detection is realized.

Specifically, assuming that the phase modulation amount of a first interference arm of a first layer MZI of a first optical network of the two optical networks is set to 0, the output reference light is a first type reference light, and the first type reference light and the calculated signal light of the first optical network interfere to output a first type detection signal light. The phase modulation quantity of a first interference arm of a first layer MZI of a second optical network in the two optical networks is set to be pi/2, the output reference light is a second type reference light, and the second type reference light and the calculated signal light of the second optical network interfere to output a second type detection signal light.

Since the interference parameters of MZI at mirror symmetry positions of the two optical networks are the same and the input signal light of the two optical networks is the same polarized, the calculated signal light obtained by the two optical networks is the same. And because the phase difference between the first type of reference light and the second type of reference light is pi/2, the photocurrent corresponding to the first type of detection signal light is equal toIn proportion to the photocurrent corresponding to the second kind of detection signal lightIs proportional to the light current corresponding to the first type of detection signal light and the second type of detection signal lightAnd obtaining a complex value calculation result.

In some embodiments, the one MZI of the two optical networks closest to the symmetry line is the last MZI, and the system further comprises a fourth class MZI connected to the first MZI of the two optical networks.

Referring to fig. 6, the MZI numbered 64 in fig. 6 represents a fourth MZI, which connects the first MZI of two optical networks. Specifically, the signal light emitted by the light source 61 is converted into two paths of input signal light, and the two paths of input signal light are provided to two optical networks in mirror symmetry, and the two optical networks are respectively passed through to output corresponding detection signal light 62 and detection signal light 63. In the complex-value detection, the fourth type of MZI may be controlled so that the signal light of the first optical network is transmitted to the next MZI of the first optical network via the fourth type of MZI and is not transmitted to the second optical network, and the signal light of the second optical network is transmitted to the next MZI of the second optical network via the fourth type of MZI and is not transmitted to the first optical network.

With continued reference to fig. 6, the number of MZI of the first layer of the two optical networks in fig. 6 is 4, and the number of MZI of the fourth type is 3.

According to the application, the interference parameter of the fourth MZI is set to be 0, so that when complex value detection is executed, the phase of the signal light passing through the fourth MZI is not changed, and the signal light of the first optical network is not transmitted to the second optical network by controlling the fourth MZI, the signal light of the second optical network is not transmitted to the first optical network, and mutual interference among signals is reduced.

In the present application, the real value calculation instruction refers to an instruction for instructing an optical calculation system to perform real value calculation to obtain a real value calculation result. The optical computing system of the present application can be used for complex value computation as well as real value computation. Specifically, the application realizes that the optical computing system can execute complex value detection and real value detection by setting the fourth MZI, wherein the method is specifically embodied in the following two aspects: first, when complex value calculation is executed, the interference parameter of the fourth type MZI is 0, so that when the signal light passes through the fourth type MZI, the phase is not modulated, and meanwhile, the signal light of the first optical network is not transmitted to the second optical network, the signal light of the second optical network is not transmitted to the first optical network, and mutual interference between signals is reduced.

Secondly, when real-value calculation is executed, the fourth type of MZI is used for connecting the first optical network and the second optical network, so that the number of MZIs serving as calculation nodes of the whole network is increased, the number of input ends of the calculation nodes is increased, and meanwhile, interference parameters of each MZI can be configured according to requirements of different application scenes, so that real-value calculation with higher dimensionality can be executed.

In particular, the optical computing system of the present application may comprise in particular an input layer, a hidden layer and an output layer. Wherein the hidden layer may comprise a plurality of neurons, all parameters and variables of the complex valued neurons are complex valued when used in complex valued calculations, which also employ complex valued arithmetic. The input-output relationship of complex valued neurons is:

Wherein, The representation input may be complex or real,Indicating the deviation.Unitary weight matrix in complex form (corner markAndRepresenting the real and imaginary parts, respectively). The optical computing system of the present application willSplit into two MZI networks of mirror symmetry, by phase of each MZI inside the networkAndIt means that the MZI is phase modulated by an external electric signal to control the phase of signal light propagating at each MZI, thereby realizing a unitary weight matrix through a plurality of MZI.

When the optical computing system of the present application is used to perform real-valued calculations, all parameters and variables of real-valued neurons are real-valued, and real-valued arithmetic is also used for the calculations.

When performing real-value calculation, the fourth type MZI may be used as a calculation node to participate in a calculation task, where an interference parameter of the fourth type MZI may not be set to 0.

According to the application, different interference parameters are configured for at least two different fourth-class MZIs, so that the fourth-class MZIs can participate in a calculation task, and the executable unitary weight matrix dimension of the whole network is improved.

The real value detection in the application refers to detecting the intensity of the detection signal light to obtain the amplitude of the detection signal light as the result of real value calculation.

In the present application, the MZI at each symmetrical position of the two optical networks is a pair of MZI. In performing the duplication computation, the interference parameters of each pair of MZI need to be configured identically. In real-valued calculations, the interference parameters for each pair of MZI may be different. The application configures different interference parameters for at least one pair of second-class MZIs, and can improve the executable unitary weight matrix dimension of the whole network.

In some embodiments, the data processing device is further connected to the controller, and the data processing device is further configured to feed back the obtained complex value calculation result to the controller, so that the controller adjusts the interference parameters of each MZI according to the complex value calculation result.

Specifically, the photoelectric detector converts the detection signal light into photocurrent, then converts the signal through the analog-to-digital converter, and transmits the converted signal to the data processing equipment, and the data processing equipment feeds the signal back to the controller after passing through the digital-to-analog converter, so that the controller adjusts the interference parameters of each MZI according to the complex value calculation result.

In some embodiments, the optical computing system further comprises a grating coupler, through which the two optical networks are respectively connected to the photodetectors.

In some embodiments, the optical computing system further comprises an optical source and an optical splitter connected to the optical source, the optical splitter further being connected to the two optical networks, respectively.

The light source is used for transmitting the original signal light to the optical splitter; the optical splitter is used for splitting the original signal light into two paths of input signal light with the same polarization and providing the two paths of input signal light to the two optical networks.

In the application, the original signal light emitted by the light source is separated into two paths of input signal light with the same polarization through the optical splitter and provided for two optical networks.

Referring to fig. 7, fig. 7 is a block diagram illustrating an optical computing system according to an embodiment. In fig. 7, an original signal light emitted by a coherent light source is separated into two paths of input signal light with the same polarization through an optical splitter, the two paths of input signal light are transmitted to an MZI optical network through a grating coupler array, the input signal light is processed through two mirror symmetry networks in the MZI optical network, detection signal light corresponding to the two networks is obtained, the detection signal light is further output to a photoelectric detector through the grating coupler array, the photoelectric detector converts the detection signal light into corresponding photocurrent, the photocurrent is input to a data processing device after passing through a digital-to-analog converter, and the data processing device extracts a phase difference between the input signal light and the calculated signal light according to the photocurrent to realize complex value calculation.

Referring to fig. 8, fig. 8 is a schematic diagram illustrating connection between two optical networks in an embodiment. In fig. 8, the first optical network 81 and the second optical network 82 are mirror symmetric about a symmetry line 83. The signal light emitted by the light source 84 is converted into two paths of input signal light, which are provided to the first optical network 81 and the second optical network 82, and the two paths of input signal light are respectively output to the corresponding detection signal light 85 and the detection signal light 86 after passing through the first optical network 81 and the second optical network 82, and the detection signal light 85 and the detection signal light 86 are respectively converted into corresponding photocurrents by the photodetector 87And。

Further, the MZI numbered 88 and the MZI numbered 89 are mirror symmetric with respect to the second type of MZI, the MZI numbered 88 and the MZI numbered 89. The number 90 is a fourth MZI, which is used to avoid that the signal light transmitted in the first optical network is transmitted to the second optical network or that the signal light transmitted in the second optical network is transmitted to the first optical network when performing complex-valued detection, and to connect the first optical network and the second optical network when performing real-valued detection, so that the whole network consisting of the first optical network and the second optical network can perform real-valued calculation with higher dimension.

In a second aspect, the present application further provides a complex value detection method, where the complex value detection method is applied to the optical computing system in any embodiment of the first aspect, as shown in fig. 9, the complex value detection method may include:

step S91, acquiring a first photoelectric current value, where the first photoelectric current value is a photocurrent converted from a detection signal light output by one of the two optical networks.

Step S92, a second optical current value is obtained, where the second optical current value is a photocurrent converted from signal light output by the other of the two optical networks.

Step S93, according to the first photoelectric value and the second photoelectric value, determining the phase difference between the reference light and the calculated signal light to perform complex value detection, and obtaining a complex value calculation result.

In the present application, the relationship between the input and output of the MZI can be expressed as:

Wherein, Is a unitary rotation matrix. The application can modulate the phase of the adjustable phase shifterAndChanges inAn arbitrary unit unitary rotation matrix is realized, and input signal light is output from two ports with power distributed by the matrix.

In the application, input signal light respectively enters a first MZI of two optical networks and is separated to obtain multiple paths of signal light to be processed, wherein part of the signal light to be processed is provided for a second MZI, and calculated signal light is obtained after passing through the second MZI, and an output electric field is assumed to be. The other part of the signal light to be processed is provided for a third type MZI of the two optical networks, the third type MZI firstly separates out the reference light, and the corresponding output electric field is assumed to beWherein, the method comprises the steps of, wherein,AndRepresenting the amplitude of the corresponding output electric field,Indicating that the same frequency is present and,AndThe phases of the calculated signal light and the input signal light are shown.

Further, the calculated signal light and the reference light in each network interfere within a certain target MZI of the third class of MZIs, provided that the amount of phase modulation of the first interference arm of this target MZI is setBy using the corresponding unit rotation matrixThe output electric field that eventually leaves both optical networks can be expressed as:

Wherein, AndThe output electric fields of the upper and lower ports of the target MZI, respectively. Each output electric field is converted into photocurrent by a corresponding photodetector, and the amplitude is proportional to the square of the output electric field, i.e

The two photodetectors connected with the target MZI are connected in a balanced manner, and then the photocurrent corresponding to one of the networks is:

（1）

the other network, i.e. the network to which the reference light is added with pi/2 phase shift, corresponds to photocurrent:

（2）

Wherein, . Based on the detectionAndThe values of (2) are combined with the expressions (1) and (2) to obtainIs a value of (2).

Wherein the present application can combine the detected values according to the ratio of the expressions (1) and (2)AndCan extract the value of (a)。

In a third aspect, the present application further provides a data processing method, where the data processing method is applied to the optical computing system in any embodiment of the first aspect, as shown in fig. 10, where the data processing method includes: step S101, obtaining data to be processed.

Step S102, modulating input signal light according to data to be processed, and inputting the input signal light into two optical networks trained in advance to obtain detection signal light corresponding to the two optical networks.

Step S103, the phase difference between the input signal light and the calculated signal light is determined according to the detection signal light corresponding to the two optical networks.

Step S104, determining a data processing result according to the phase difference between the input signal light and the calculated signal light.

The data to be processed herein may be any data that needs to be processed by the optical computing system. The data to be processed here may be image data, audio data, or the like, for example. Accordingly, the data processing result here may be an image category, a prediction result of audio data, or the like.

In step S102, the phase and amplitude of the input signal light may be determined according to the data to be processed, and the input signal light may be further modulated according to the phase and amplitude of the input signal light.

Further, the specific implementation manners of step S102 to step S103 may refer to the detailed descriptions of the respective embodiments of the first aspect and the second aspect, which are not repeated herein.

Further, the data processing result is determined according to the phase difference between the input signal light and the calculated signal light, and the data processing result corresponding to each phase difference can be determined according to a preset mapping relation.

In a fourth aspect, the present application further provides a training method of an optical neural network, which is applied to the optical computing system in any embodiment of the first aspect, as shown in fig. 11, where the training method of the optical neural network includes:

Step S111, sample data for training is acquired.

Step S112, modulating input signal light according to the sample data, inputting the two initialized optical networks in the optical computing system, and obtaining detection signal light corresponding to the two optical networks.

Step S113, determining a phase difference between the input signal light and the calculated signal light according to the detected signal light corresponding to the two optical networks.

Step S114, determining a data processing result according to the phase difference between the input signal light and the calculated signal light; and determining the calculated value of the loss function according to the data processing result of each round.

Step S115, adjusting the interference parameters of the MZIs in the two optical networks according to the calculated value of the loss function until the calculated value of the loss function meets the termination condition.

Step S116, the interference parameters of each MZI in the two optical networks in the optical computing system when the termination condition is met are obtained and stored, and training is terminated.

According to the application, the optimal network models of the two optical networks can be obtained through repeated iterative training. Specifically, the present application may construct a loss function in advance, and set termination conditions in advance. The loss function may be constructed according to actual requirements or actual application scenarios, and is not specifically limited herein. The termination condition may include that the number of iterative training reaches a preset number and/or that the value of the loss function for a consecutive preset number is smaller than a preset threshold.

Further, when the calculated value of the loss function corresponding to the current round does not meet the termination condition, the interference parameters of each MZI in the two optical networks are modulated according to the calculated value of the loss function, new sample data are obtained after modulation, and the next training is performed until the calculated value of the loss function meets the termination condition, and the training is terminated.

It should be understood that, although the steps in the flowcharts of fig. 9-11 are shown in order as indicated by the arrows, these steps are not necessarily performed in order as indicated by the arrows. The steps illustrated in fig. 9-11, as well as the steps involved in other embodiments, are not strictly limited to the order of execution unless explicitly stated herein, and may be performed in other orders. Moreover, at least some of the steps of the foregoing embodiments may include multiple sub-steps or stages that are not necessarily performed at the same time, but may be performed at different times, nor do the order in which the sub-steps or stages are performed necessarily performed in sequence, but may be performed alternately or alternately with at least a portion of the sub-steps or stages of other steps or other steps.

In some embodiments, the internal architecture of the controller of the light computing system may be as shown in FIG. 12. The controller includes a processor, a memory, and a network interface connected by a system bus. Wherein the processor of the controller is configured to provide computing and control capabilities. The memory of the controller includes a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface of the controller is used for communicating with an external terminal through network connection. The computer program, when executed by a processor, implements the steps performed by the controller in any of the embodiments herein.

In some embodiments, the internal structure of the detector of the optical computing system may also take the structure shown in FIG. 12. The function of each internal structure of the detector may refer to the above description of the internal structure of the controller, and will not be described herein.

It will be appreciated by those skilled in the art that the structure shown in fig. 12 is merely a block diagram of a portion of the structure associated with an embodiment of the present disclosure and is not limiting of the computer device to which an embodiment of the present disclosure may be applied, and that a particular computer device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.

Those skilled in the art will appreciate that implementing all or part of the above-described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed may comprise the steps of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in embodiments provided by the present disclosure may include non-volatile and/or volatile memory. The nonvolatile memory can include Read Only Memory (ROM), programmable ROM (PROM), electrically Programmable ROM (EPROM), electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double Data Rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (SYNCHLINK) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), among others.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the disclosure.

The foregoing examples merely represent embodiments of the present disclosure, which are described in more detail and detail, but are not to be construed as limiting the scope of the present disclosure. It should be noted that variations and modifications can be made by those skilled in the art without departing from the spirit of the disclosure, which are within the scope of the disclosure. Accordingly, the scope of the present disclosure should be determined from the following claims.

Claims

1. An optical computing system comprising a controller, a detector, and two optical networks mirror symmetric in topology, each comprising a first type of MZI, a second type of MZI, and a third type of MZI, wherein,

The two optical networks are used for receiving input signal light with the same polarization;

The first MZI is used for dividing the signal light into a plurality of paths of signal light to be processed and providing the signal light to the second MZI and the third MZI;

the second type of MZI is used as a computing node and is used for carrying out phase modulation on the signal light to be processed to obtain computing signal light, and the computing signal light is provided for the third type of MZI;

the third type MZI is used for generating reference light according to the signal light to be processed, and interfering the reference light with the calculated signal light to obtain detection signal light, wherein the phase difference of the reference light output by the third type MZI of the two optical networks is pi/2;

The controller is used for configuring the same interference parameters for the second type MZI at the mirror symmetry positions of the two optical networks after acquiring the complex value calculation instruction;

the detector is used for respectively receiving the detection signal lights output by the two optical networks during complex value calculation to carry out complex value detection so as to determine the complex value calculation result.

2. The system of claim 1, wherein in the topology, the two optical networks comprise multiple MZI in sequence from near to far from the symmetry line, each layer having a decreasing number of MZI in sequence.

3. The system of claim 2, wherein the number of MZI layers per layer decreases one after the other in a direction from the near to the far direction of the two optical networks from the symmetry line.

4. The system of claim 2, wherein a layer of MZI closest to the symmetry line in the two optical networks is a first layer of MZI, the third class of MZI comprises an output side MZI and the first layer of MZI, and a layer of MZI furthest from the symmetry line is a last layer of MZI;

starting from the last MZI, each layer of MZI comprises the first MZI and the output side MZI, starting from the last MZI, each layer of MZI further comprises at least one second MZI, starting from the last MZI, and the first MZI and the output side MZI are respectively positioned at the head end and the tail end of each layer, and the second MZI is positioned in the middle of each layer.

5. The system of claim 4, wherein the number of MZI of the first type and the number of output side MZI in each layer MZI are the same.

6. The system of claim 5, wherein the number of MZI of the first type and the number of output side MZI in each layer MZI are each 1.

7. The system of claim 4, wherein the interference parameters include a phase modulation amount of a first interference arm of the MZI and a phase modulation amount of a second interference arm of the MZI, a difference between the phase modulation amounts of the first interference arm and the second interference arm of the first layer MZI in the two optical networks is pi/2, the phase modulation amounts of the second interference arm are identical, the phase modulation amounts of the first interference arm of the output side MZI at the mirror-symmetrical position are set to be identical, and the phase modulation amounts of the second interference arm are also identical.

8. The system of claim 7, wherein the amount of phase modulation of the first interference arm of the last layer MZI of one of the two optical networks is set to 0 and the amount of phase modulation of the first interference arm of the last layer MZI of the other network is set to pi/2.

9. The system of claim 4, wherein the detector comprises a photodetector connected to the output side MZI of the two optical networks, respectively, and a data processing device connected to the photodetector; wherein,

The photoelectric detector is used for respectively converting the detection signal light output by the two optical networks into photocurrent; the data processing device is used for determining the phase difference between the input signal light and the calculated signal light according to the photocurrents corresponding to the two optical networks so as to perform complex value detection and obtain a complex value calculation result.

10. The system of claim 9, wherein two outputs of the output side MZI in each layer MZI are each connected to one of the photodetectors, and two photodetectors connected to the same output side MZI are connected in balance.

11. The system of claim 9, further comprising a grating coupler, wherein the two optical networks are each connected to the photodetector through the grating coupler.

12. The system of claim 4, further comprising a fourth type of MZI connecting the first layer MZI of the two optical networks.

13. The system of claim 12, wherein the number of fourth class MZI is 1 more than the number of second class MZI in the first layer MZI.

14. The system of claim 12, wherein the interference parameter of the fourth class MZI is set to 0.

15. The system of claim 12, wherein the controller is further configured to configure different interference parameters for at least two different fourth-class MZI after the real-valued calculation instructions are obtained.

16. The system of claim 1, wherein the controller is further configured to configure different interference parameters for at least one pair of MZI of a second type at mirror symmetric locations of the two optical networks after obtaining the real-valued calculation instructions;

the detector is also used for respectively receiving the detection signal lights output by the two optical networks during real-value calculation to carry out real-value detection so as to determine the result of the real-value calculation.

17. The system of claim 1, wherein when the total number of inputs of the MZI of the second type of network is N, the sum y of the number of MZI of the first type and the third type of MZI is related to the number x of MZI of the second type as follows:

x＝N(N-1)

Wherein N is greater than or equal to 2.

18. The system of claim 1, further comprising a light source and an optical splitter connected to the light source, the optical splitter further being connected to the two optical networks, respectively, wherein the light source is configured to deliver the original signal light to the optical splitter;

The optical splitter is used for splitting the original signal light into two paths of input signal light with the same polarization and providing the two paths of input signal light to the two optical networks.

19. A complex-valued detection method, characterized in that the complex-valued detection method is applied to the optical computing system according to any one of claims 1 to 18, the complex-valued detection method comprising:

Acquiring a first photoelectric value, wherein the first photoelectric value is a value of photoelectric current converted by detection signal light output by one of the two optical networks;

acquiring a second photoelectric value, wherein the second photoelectric value is a value of photoelectric current converted by signal light output by the other of the two optical networks;

and determining a phase difference between the input signal light and the calculated signal light according to the first photoelectric current value and the second photoelectric current value so as to perform complex value detection and obtain a complex value calculation result.

20. The method of claim 19, wherein determining the phase difference between the reference light and the calculated signal light based on the first and second photo current values comprises:

calculating the ratio of the first photoelectric value to the second photoelectric value;

And determining a phase difference between the input signal light and the calculated signal light according to the ratio.

21. A data processing method, wherein the data processing method is applied to the optical computing system according to any one of claims 1 to 18, the data processing method comprising:

Acquiring data to be processed;

Modulating input signal light according to the data to be processed, and inputting the two optical networks which are trained in advance to obtain detection signal light corresponding to the two optical networks;

Determining a phase difference between the input signal light and the calculated signal light according to the detection signal light corresponding to the two optical networks;

22. A method of training an optical neural network for use in an optical computing system as claimed in any one of claims 1 to 18, the method comprising:

Acquiring sample data for training;

Modulating input signal light according to the sample data, inputting two initialized optical networks in an optical computing system, and obtaining detection signal light corresponding to the two optical networks;

Determining a data processing result according to the phase difference between the input signal light and the calculated signal light; determining a calculated value of the loss function according to the data processing result of each round;

adjusting interference parameters of MZIs in the two optical networks according to the calculated value of the loss function until the calculated value of the loss function meets a termination condition;

And acquiring and storing interference parameters of each MZI in two optical networks in the optical computing system when the termination condition is met, and terminating training.