Hybrid photon neural network device and classification method

Abstract

The invention relates to the field of optical calculation and optical signals, in particular to a hybrid photon neural network device and a classification method. Comprising a multi-channel tunable light source, a modulator array, an FPGA and a 4 * 4 MZI array. A DFB-SA array and a PD array; the output end of the multi-channel tunable light source is connected with the input end of the modulator array, the FPGA is connected with the radio frequency end of the modulator array, the output end of the modulator array is connected with the input end of the 4 * 4 MZI array, the output end of the 4 * 4 MZI array is connected with the input end of the DFB-SA array, and the output end of the DFB-SA array is connected with the input end of the PD array. According to the invention, the phase parameters of the 4 * 4 MZI array can be configured through an efficient gradient descent algorithm, so that the device has advantages in the aspect of flexible simulation of neural network synapses; in addition, the device adopts a distributed feedback semiconductor laser array with a saturated absorption region, and the nonlinear activation function of photon neurons is realized by fully utilizing the threshold characteristic of the distributed feedback semiconductor laser array.

Description

Technical Field

[0001] The present invention relates to the field of optical computing and optical signals, and in particular to a hybrid photonic neural network device and a classification method. Background Art

[0002] At present, the limitations of the traditional von Neumann computing architecture in improving computing performance are becoming increasingly prominent, mainly reflected in the delay and energy consumption problems caused by the frequent movement of data between the processor and memory, which limits the improvement of overall computing efficiency.

[0003] In this context, neuromorphic computing has emerged as a computing model that simulates the working mode of biological neural networks. It deeply simulates the precise organizational structure and complex operating mechanism of the human brain's nervous system, thereby achieving the ability to process complex information and learn tasks. Compared with the processing method of the von Neumann architecture, neuromorphic computing has higher parallelism and fault tolerance, and can more efficiently cope with large-scale data sets and complex computing challenges.

[0004] As a branch of neuromorphic computing, photonic neural networks have even better effects. Photonic neural networks use photons as a medium for information transmission and use optical devices to simulate the functions of neurons and synapses. Because photons have the characteristics of high speed, low power consumption and easy integration, they not only greatly improve the speed of information processing and reduce energy consumption, but also provide the possibility of realizing more highly integrated neural network systems.

[0005] At present, the research on photonic neural networks mainly focuses on optical synapses and optical neurons. However, in the process of large-scale application, how to build a complete and efficient photonic neural network architecture to ensure seamless connection and close collaboration between photonic synapses and photonic neurons is an urgent problem that people in this field need to solve. Summary of the invention

[0006] In response to the problems mentioned in the prior art, the present invention proposes a hybrid photonic neural network device and a classification method. In the device, a 4×4 MZI array is used as a photonic synapse, and a distributed feedback semiconductor laser with a saturated absorption zone is used as a photonic neuron to jointly construct a hybrid photonic neural network. The hybrid photonic neural network is used as an inference framework for neuromorphic applications, and weight parameters can be flexibly configured, which is easy to implement.

[0007] In order to achieve the above object, the present invention adopts the following technical solution: The present invention discloses a hybrid photonic neural network device, comprising a multi-channel tunable light source, a modulator array, an FPGA, a 4×4 MZI array, a DFB-SA array and a PD array; The output end of the multi-channel tunable light source is connected to the input end of the modulator array, the FPGA is connected to the RF end of the modulator array, the output end of the modulator array is connected to the input end of the 4×4 MZI array, the output end of the 4×4 MZI array is connected to the input end of the DFB-SA array, and the output end of the DFB-SA array is connected to the input end of the PD array.

[0008] As a further improvement of the present invention, the 4×4 MZI array is a four-channel Mach-Zehnder interferometer array, and the DFB-SA array is a distributed feedback semiconductor laser array with a saturated absorption region; A 4×4 MZI array is used as a photonic synapse and a DFB-SA array is used as a photonic neuron to jointly construct a hybrid photonic neural network.

[0009] As a further improvement of the present invention, the hybrid photonic neural network can be flexibly configured according to weight parameters.

[0010] As a further improvement of the present invention, in a 4×4 MZI array, the phase parameters of the 4×4 MZI array are configured by a gradient descent algorithm, and the gradient descent algorithm comprises the following steps: S1, randomly initialize the phase of the phase shifter in the 4×4 MZI array; S2. Define the cost function CF to measure the difference between the network state and the target state; S3, through the forward propagation and back propagation process, the phase of the phase shifter is continuously adjusted to make the CF converge, so as to realize the simulation of the weight matrix; S4, forward propagation obtains the output result according to the current phase of the phase shifter, and reverse propagation adjusts the phase value according to the gradient information of CF on the phase of the phase shifter; S5. Approach the optimal weight configuration through iteration.

[0011] As a further improvement of the present invention, the specific process of step S2 is: define the cost function CF as the theoretical output With the actual output The similarity between them is expressed as follows: .

[0012] As a further improvement of the present invention, the theoretical output and the actual output The expressions are as follows:

[0013]

[0014] Where: represents the target weight matrix to be simulated; represents the actual weight matrix; Represents the input matrix of the settings.

[0015] As a further improvement of the present invention, the PD array is also connected to an oscilloscope OSC.

[0016] A classification method of a hybrid photonic neural network device, the classification method is to classify irises.

[0017] As a further improvement of the present invention, the multi-channel tunable light source is used to generate an optical carrier signal with constant power, the FPGA is used to generate an electrical signal with iris sample data characteristics, and the modulator array modulates the optical carrier signal according to the electrical signal to generate optical modulation signals of specific intensity corresponding to information of different characteristic dimensions, so that the optical modulation signal adapts to the calculation requirements of iris classification; The optical modulated signal enters the 4×4 MZI array and performs linear weighted operations according to the preset weight matrix to form a weighted pulse signal. When the DFB-SA array receives the weighted pulse signal, it generates a nonlinear response. The PD array converts the optical signal output by the DFB-SA array into an electrical signal, which is output as the final iris classification result.

[0018] As a further improvement of the present invention, the information of the characteristic dimensions includes iris sepal length, sepal width, petal length, and petal width.

[0019] Compared with the prior art, the present invention has achieved the following technical effects: The present invention discloses a hybrid photonic neural network device. By setting up a 4×4 Mach-Zehnder interferometer array, its phase parameters can be configured through an efficient gradient descent algorithm, so that the device has advantages in flexibly simulating neural network synapses; in addition, the device adopts a distributed feedback semiconductor laser array with a saturated absorption zone, and fully utilizes its threshold characteristics to realize the nonlinear activation function of photonic neurons.

[0020] Since different application scenarios and tasks often require neural networks to have different weight matrix configurations, the hybrid photonic neural network device of the present invention can meet the above requirements by virtue of the flexibility and configurability of its weight parameters. By adjusting the phase parameters of the Mach-Zehnder interferometer array and the activation threshold of the distributed feedback semiconductor laser array with a saturated absorption zone, the device can adapt to a wide range of neuromorphic application scenarios and meet the requirements of diverse tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic diagram of the structure of a hybrid photonic neural network device provided by an embodiment of the present invention; Figure 2is the power-current curve of DFB-SA under the reverse bias voltage of -3V; Figure 3 An experimental scheme diagram for verifying the performance of the hybrid photonic neural network device provided in an embodiment of the present invention; Figure 4 Diagram of the training process for configuring the target weight matrix for a 4×4 MZI array using the gradient descent algorithm; Figure 5 This is a diagram of the results of an iris classification experiment implemented by the device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0023] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0024] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0025] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, an electrical connection, or a communication; it can be a direct connection, or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0026] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0027] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0028] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0029] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0030] Various structural schematic diagrams of the embodiments disclosed in the present invention are shown in the accompanying drawings. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0031] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0032] To build an inference framework for neuromorphic applications, we further implement a photonic neural network with flexibly configurable weight parameters, such as Figure 1As shown, the present invention proposes a hybrid photonic neural network device, including a multi-channel tunable light source, a modulator array, an FPGA, a 4×4MZI array; a DFB-SA array and a PD array; the output end of the multi-channel tunable light source is connected to the input end of the modulator array, the FPGA is connected to the RF end of the modulator array, the output end of the modulator array is connected to the input end of the 4×4MZI array, the output end of the 4×4MZI array is connected to the input end of the DFB-SA array, and the output end of the DFB-SA array is connected to the input end of the PD array.

[0033] Specifically in the embodiment, FPGA is a field programmable gate array; 4×4 MZI array is a four-channel Mach-Zehnder interferometer array, which serves as a photonic synapse; DFB-SA is a distributed feedback semiconductor laser with a saturated absorption region, which serves as a photonic neuron; and 4×4 MZI array and DFB-SA array constitute a hybrid photonic neural network.

[0034] The DFB-SA laser in the embodiment is a photon neuron with typical neuron characteristics such as pulse threshold. The DFB-SA is composed of two parts: a gain region and a saturation absorption region, and these two regions are electrically isolated from each other; the gain region is controlled by adjusting the gain current, and the saturation absorption region is controlled by adjusting the reverse bias voltage. Figure 2 As shown, when the temperature is controlled at 25°C and the reverse bias voltage is set to -3V, the experimental test shows that the threshold current of DFB-SA is 27mA.

[0035] In this embodiment, the 4×4 MZI array is used as a photon synapse to simulate weights with flexibly adjustable parameters. Specifically, the phase of the phase shifter is configured using a gradient descent algorithm. The gradient descent algorithm mainly includes: (1) Randomly initialize the phase of the phase shifters in the 4×4 MZI array; (2) Define the cost function CF to measure the difference between the network state and the target state; (3) Digital forward propagation obtains the output result based on the phase of the phase shifter; (4) Backward propagation adjusts the phase value according to CF; (5) Repeat (3) and (4) to approach the optimal weight configuration through multiple rounds of iterations.

[0036] In practical applications, the way to adjust the phase value of the phase shifter is to adjust the voltage applied to the electrode. Specifically, it includes: (1) Randomly initialize the loaded voltage value; (2) Define the cost function CF as the theoretical output With the actual output The similarity between them, where the CF expression is as follows:

[0037] Theoretical output and the actual output The expressions are as follows:

[0038]

[0039] Where: represents the target weight matrix to be simulated; represents the actual weight matrix, which is composed of the voltages applied to the electrodes of the 4×4 MZI array; Represents the input matrix of the settings.

[0040] (3) Forward propagation to obtain actual output . (4) Backward propagation adjusts the voltage according to the calculation result of CF. The current voltage Increase step length or reduce ; (5) Repeat (3) and (4) to make CF approach the target value 1 after multiple rounds of iterations.

[0041] It can be understood that different tasks correspond to different weight matrices. By using this gradient descent algorithm to complete training, the corresponding weight matrix can be adapted for each task, and flexible configuration of weight parameters can be achieved, thereby making the hybrid photonic neural network device suitable for a variety of application scenarios.

[0042] In one embodiment, the hybrid photonic neural network device provided by the embodiment of the present invention can be used to classify irises, that is, the task of the photonic neural network includes: classifying irises. The iris classification task mentioned here is to distinguish various types of characteristic data of irises. The processed data comes from the iris data set, such as information on four characteristic dimensions of iris sepal length, sepal width, petal length, and petal width. These data are integrated and converted into a specific format. The purpose of classification can be related to the actual task of the photonic neural network, that is, to accurately identify and distinguish the specific categories to which different iris samples belong. With regard to specific category subdivision rules and other contents, the embodiment of the present invention maintains flexibility and does not make rigid regulations.

[0043] like Figure 3As shown, an example of the present invention consists of four tunable light sources TL1, TL2, TL3, TL4, four modulators MZM1, MZM2, MZM3, MZM4, a field programmable gate array FPGA, four erbium-doped fiber amplifiers EDFA1, EDFA2, EDFA3, EDFA4, seven polarization controllers PCs, a 4×4 MZI array, a DFB-SA array, three photodetectors PD1, PD2, PD3 and an oscilloscope OSC.

[0044] In this embodiment, four TLs output a continuous optical carrier of 16 dBm; information of four characteristic dimensions of iris flower samples, namely, sepal length, sepal width, petal length, and petal width, is input into the FPGA to generate an electrical signal having the data characteristics of the iris flower sample; the modulator array modulates the optical carrier signal according to the electrical signal to generate an optical modulation signal; the EDFA is used to adjust the power of the optical modulation signal; a polarization controller PC is placed on each of the four channels to adjust the polarization state of the optical signal; the optical signal output by the PC enters the 4×4 MZI array to load weight information, performs linear weighted operations, and forms a weighted pulse signal; the output port of the 4×4 MZI array is connected to the PC, and three PCs are used to control the polarization states of the optical signals in the three channels respectively so as to match them with the DFB-SA; the optical signal output by the PC generates a nonlinear response after entering the DFB-SA array; the output ends of the DFB-SA array are respectively connected to the PDs, and the three PDs convert the received optical signals into electrical signals respectively, and the final iris flower classification results are observed on the oscilloscope.

[0045] The voltages of MZM1, MZM2, MZM3, MZM4 and the phase shifters in the 4×4 MZI array are provided by a voltage source, and the temperatures of the 4×4 MZI array and the DFB-SA array are controlled at the same time.

[0046] The experimental results are as follows: In the embodiment of the present invention, the training process of configuring the target weight matrix using the gradient descent algorithm for the 4×4 MZI array is as follows: Figure 4 As shown in the figure, the broken line shows how CF changes with the number of iterations. As the number of iterations increases, CF shows an overall upward trend and approaches the target value of 1. This shows that during each iteration, the voltage loaded on the phase shifter is constantly adjusted, so that the similarity between the actual output and the theoretical output continues to increase. After multiple rounds of iterations, the 4×4 MZI array gradually approaches the target weight configuration to achieve iris classification.

[0047] In the embodiment of the present invention, the signal input into the 4×4 MZI array is as follows: Figure 5(a1), (a2), and (a3) ​​are three different types of iris samples. Considering that the iris classification task is a three-classification task, the signals of the first three output ports of the 4×4 MZI array are selected as the output results. The output results are shown in Figure 5 (b1), (b2), (b3); the output of the 4×4 MZI array is polarized by the PC and then enters the DFB-SA array, using its threshold characteristics to achieve nonlinear response. Specifically, if the signal received by the DFB-SA reaches the threshold, a pulse is generated, that is, the neuron generates an action potential after receiving a certain intensity of stimulation, otherwise no pulse is emitted. For each sample, it is classified according to the situation that causes the DFB-SA array to respond. The output of the DFB-SA array is shown in Figure 5 As shown in (c1), (c2), and (c3), it can be seen that the spike pulse successfully distinguishes the three types of irises and is consistent with the corresponding classification in theory.

[0048] The present invention discloses a hybrid photonic neural network device. By setting up a 4×4 Mach-Zehnder interferometer array, its phase parameters can be configured through an efficient gradient descent algorithm, so that the device has advantages in flexibly simulating neural network synapses; in addition, the device adopts a distributed feedback semiconductor laser array with a saturated absorption zone, and fully utilizes its threshold characteristics to realize the nonlinear activation function of photonic neurons.

[0049] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0050] In addition, it should be understood that although this specification is described in accordance with the implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation modes that can be understood by those skilled in the art. The above content is only to illustrate the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A hybrid photonic neural network device, characterized in that: Includes multi-channel tunable light source, modulator array, FPGA, 4×4 MZI array; DFB-SA array and PD array; The output end of the multi-channel tunable light source is connected to the input end of the modulator array, the FPGA is connected to the RF end of the modulator array, the output end of the modulator array is connected to the input end of the 4×4 MZI array, the output end of the 4×4 MZI array is connected to the input end of the DFB-SA array, and the output end of the DFB-SA array is connected to the input end of the PD array.

2. A hybrid photonic neural network device according to claim 1, characterized in that: The 4×4 MZI array is a four-channel Mach-Zehnder interferometer array, and the DFB-SA array is a distributed feedback semiconductor laser array with a saturated absorption region; A 4×4 MZI array is used as a photonic synapse and a DFB-SA array is used as a photonic neuron to jointly construct a hybrid photonic neural network.

3. A hybrid photonic neural network device according to claim 2, characterized in that: The hybrid photonic neural network can be flexibly configured according to the weight parameters.

4. A hybrid photonic neural network device according to claim 2, characterized in that: In the 4×4 MZI array, the phase parameters of the 4×4 MZI array are configured by a gradient descent algorithm, which includes the following steps: S1, randomly initialize the phase of the phase shifter in the 4×4 MZI array; S2. Define the cost function CF to measure the difference between the network state and the target state; S3, through the forward propagation and back propagation process, the phase of the phase shifter is continuously adjusted to make the CF converge, so as to realize the simulation of the weight matrix; S4, forward propagation obtains the output result according to the current phase of the phase shifter, and reverse propagation adjusts the phase value according to the gradient information of CF on the phase of the phase shifter; S5. Approach the optimal weight configuration through iteration.

5. A hybrid photonic neural network device according to claim 4, characterized in that: The specific process of step S2 is: define the cost function CF as the theoretical output The actual output The similarity between them is expressed as follows: 。 6. A hybrid photonic neural network device according to claim 5, characterized in that: Theoretical output and the actual output The expressions are as follows: Where: represents the target weight matrix to be simulated; represents the actual weight matrix; Represents the input matrix of the settings.

7. A hybrid photonic neural network device according to claim 1, characterized in that: The PD array is also connected to an oscilloscope OSC.

8. A classification method for a hybrid photonic neural network device according to any one of claims 1 to 7, characterized in that: The classification method is to classify irises.

9. The classification method of a hybrid photonic neural network device according to claim 8, characterized in that: The multi-channel tunable light source is used to generate an optical carrier signal with constant power, the FPGA is used to generate an electrical signal with the characteristics of iris flower sample data, and the modulator array modulates the optical carrier signal according to the electrical signal to generate optical modulation signals of specific intensity corresponding to information of different characteristic dimensions, so that the optical modulation signal is adapted to the calculation requirements of iris flower classification; The optical modulated signal enters the 4×4 MZI array and performs linear weighted operations according to the preset weight matrix to form a weighted pulse signal. When the DFB-SA array receives the weighted pulse signal, it generates a nonlinear response. The PD array converts the optical signal output by the DFB-SA array into an electrical signal, which is output as the final iris classification result.

10. The classification method of a hybrid photonic neural network device according to claim 9, characterized in that: The information of the characteristic dimensions includes the iris sepal length, sepal width, petal length, and petal width.

Images