CN116029343A - Photon pulse neural network implementation method based on MRR and phase change material - Google Patents

Abstract

The invention provides a photon pulse neural network implementation method based on MRR and phase change materials, which simulates the characteristics of an optical synaptic array and constructs a PMRR unit for replacing the optical synaptic array; simulating the optical pulse transmission characteristics of the optical pulse neurons, and optimizing the PMRR units to obtain optimized PMRR units; an optical synaptic array is simulated by using PMRR units, an optical pulse neuron is simulated by using optimized PMRR units, and a beam splitter is simulated by a wavelength division multiplexing selector, so that a photon pulse neural network based on MRR and phase change materials is obtained. The PMRR unit designed by the invention can realize the weighted calculation function of optical synapse on input pulse, and can also realize the function of photon pulse neuron, and the two can realize the feasibility of photon pulse neural network through integration. Compared with the prior art, the method has the characteristics of easier implementation, easier integration, low complexity, strong expandability and the like.

Description

Photon pulse neural network implementation method based on MRR and phase change material

Technical Field

The invention belongs to the technical field of optical pulse network identification, and particularly relates to a photon pulse neural network implementation method based on MRR and phase change materials.

Background

Nowadays, artificial Neural Networks (ANNs) have achieved significant achievements in various application fields such as pattern recognition, object tracking, and the like. However, the computational model involved in artificial neural networks is becoming more complex, and thus it is necessary to build a high computational power and low power consumption computing platform. Traditional computers based on von neumann architecture are limited by the separation of memory and processor units, which prevents efficient data transfer between them. Third generation neural networks, i.e., impulse neural networks (SNNs), are more brain-like than traditional neural networks, and SNNs have higher computational efficiency and lower power consumption.

To date, integration schemes have been validated numerically by simulation or experimentation on optical synapses, e.g., based on MRR weight libraries, PCM, semiconductor optical amplifier SOAs. The optical pulse neuron scheme based on discrete devices and integrated schemes also has a photonic pulse neuron based on PCM, a photonic pulse neuron based on a micro-column laser, and a photonic pulse neuron based on an integrated distributed feedback semiconductor laser.

However, photon implementation is still in the start-up phase. Graphene materials embedded in silicon waveguides have been successfully used to simulate photonic neurons to improve the efficiency of nonlinear effects. However, graphene is not easy to prepare and poor reproducibility remains a significant challenge. The weight scheme based on the SOA has larger occupied area and is not beneficial to high integration. The output power based on micropillar laser neurons is relatively low and when applied to multi-layer or deep optical pulse neural networks, additional amplification may be required to compensate for the loss. Photon pulse neurons based on integrated distributed feedback semiconductor lasers require photoelectric conversion, which increases the complexity and power consumption of the system.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a photon pulse neural network implementation method based on MRR and phase change materials. The technical problems to be solved by the invention are realized by the following technical scheme:

the invention provides a photon pulse neural network implementation method based on MRR and phase change materials, which comprises the following steps:

step 1: acquiring structural information of an optical pulse neural network;

the structure information comprises neuron components, connection relations of the neurons and connection relations among different constituent units of the optical pulse neural network, wherein the optical pulse neural network comprises an input layer, a beam combiner, a beam splitter, an optical synapse array and an output layer;

step 2: simulating the magnitude of optical synapse control light transmittance in the optical synapse array, and constructing a PMRR unit for replacing the optical synapse array;

the PMRR unit comprises an MRR unit and a PCM phase change unit, wherein the PCM phase change unit is positioned on an annular waveguide of the MRR unit;

step 3: simulating optical pulse transmission characteristics of optical pulse neurons in an input layer and an output layer, and optimizing the PMRR unit to obtain an optimized PMRR unit;

step 4: simulating optical synaptic weights in the optical synaptic arrays using the PMRR units, simulating an output layer and optical pulse neurons of the output layer using optimized PMRR units, and simulating the beam combiner by a wavelength division multiplexing selector constructed by an MRR unit to obtain a photon pulse neural network based on the MRR and phase change materials.

The invention has the beneficial effects that:

the invention provides a photon pulse neural network implementation method based on MRR and phase change materials, which comprises the steps of obtaining structural information of an optical pulse neural network; simulating the regulation weight of the optical synaptic array, and constructing a PMRR unit for replacing the optical synaptic array; simulating optical pulse transmission characteristics of optical pulse neurons in an input layer and an output layer, and optimizing the PMRR unit to obtain an optimized PMRR unit; simulating optical synaptic weights in the optical synaptic arrays using the PMRR units, simulating optical pulse neurons of an input layer and an output layer using optimized PMRR units, and simulating the beam combiner by a wavelength division multiplexing selector constructed by an MRR unit to obtain a photon pulse neural network based on the MRR and phase change materials. The PMRR unit designed by the invention can realize the calculation function of the optical synapse regulation weight on input pulse, and can also realize the function of photon pulse neurons, and the combination of the two can realize the feasibility of the photon pulse neural network. Compared with the prior art, the method has the characteristics of easier implementation, easier integration, low complexity, strong expandability and the like.

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Drawings

FIG. 1 is a schematic diagram of a PMRR unit-based optical pulse neural network;

FIG. 2 is a schematic cross-sectional view of an MRR unit and GST unit embedded in the MRR;

FIG. 3 is a graph showing the transmittance of PMRR units versus the amorphization degree;

FIG. 4 is a schematic diagram of a PMRR simulated synapse array;

FIG. 5 is a schematic of PMRR simulated optical impulse neuron responses;

FIG. 6 is a schematic diagram of a fully connected optical impulse neural network;

FIG. 7 is a graph showing the results of the time evolution of the output peaks during training of 12 input modes;

FIG. 8 is a result diagram of a clean picture and a noisy picture;

fig. 9 is a graph of the result of pattern recognition of an input picture in a 12-input pattern test.

Detailed Description

The present invention will be described in further detail with reference to specific examples, but embodiments of the present invention are not limited thereto.

Before describing the present invention, the technical concept of the present invention will be described first.

Ge ₂ Sb ₂ Te ₅ (GST) is one of the most mature phase change memory materials at present. The GST material has the advantages of high crystallization speed and large difference of photoelectric properties between amorphous (a-GST) and crystalline (c-GST). The nature of GSTs shows the ability to achieve photonic synapses. Silicon-based optoelectronics is an ideal platform for photonic neural networks because it is complementary toMetal Oxide Semiconductor (CMOS) compatibility has the advantages of low cost, commercial maturity, ease of integration with electronic devices, etc. Mach-Zehnder interferometers (MZIs) and microring resonators (MRRs) are two dominant methods of implementing photonic neural networks. The MRR has the advantages of small volume, easy integration, strong expandability and the like, so that the applicant combines the phase change material with the MRR to simulate the optical pulse neural network.

The following describes the specific procedure of the scheme of the present invention in detail.

Referring to fig. 1 and fig. 2, the implementation method of the photonic pulse neural network based on the MRR and the phase change material provided by the invention includes:

step 1: acquiring structural information of an optical pulse neural network;

the structure information comprises neuron components, connection relations of the neurons and connection relations among different constituent units of the optical pulse neural network, and the optical pulse neural network comprises an input layer, a beam combiner, a beam splitter, an optical synaptic array and an output layer;

referring to fig. 1, (a) is a diagram of an optical pulse neural network in the prior art. Photon SNN consists of N input (PRE) neurons, M output (POST) neurons, and n×m interconnected synapses. As shown in fig. 1 (a), the information is encoded in peak form. The input pulse sequence is multiplied by the synaptic weight and the weighted sum signal is received by the post-synaptic neuron. The internal state of a neuron is called "membrane potential", which is weighted by synapses and compared to a threshold at each time step. Once the membrane potential reaches a threshold, the neuron will output a peak.

Step 2: simulating the magnitude of optical synapse control light transmittance in the optical synapse array, and constructing a PMRR unit for replacing the optical synapse array;

the PMRR unit comprises an MRR unit and a PCM phase change unit, wherein the PCM phase change unit is positioned on an annular waveguide of the MRR unit; the phase change material of the PCM phase change cell comprises a GST phase change material.

Referring to fig. 2, the MRR unit of the present invention is composed of 2 straight waveguides and 1 annular waveguide, the annular waveguide is located in the middle of the 2 straight waveguides, one straight waveguide is an input end waveguide and the other is an output end waveguide; the PCM phase change unit is covered on the annular waveguide; the PCM phase change unit heats the phase change material of the PCM phase change unit by light pulse passing through the straight waveguide to cause phase change, so that the PCM phase change unit can be converted between crystalline state and amorphous state; the PCM phase change units with different crystallization degrees can control the intensity of light in the input straight waveguide of the MRR, so as to realize different synaptic intensities.

The PMRR cell is composed of 2 straight waveguides, 1 annular waveguide and 1 PCM cell, as shown in fig. 2 (a), wherein the PCM cells covered on the annular waveguide are heated by the light pulse passing through the waveguide, so that the PCM cells can be converted between crystalline and amorphous states, and PCM cells with different crystallization degrees can control the light intensity of Drop ports of the MRR, so as to realize different synaptic intensities.

The invention is not limited to GST phase change materials, and all phase change materials capable of realizing the invention belong to the protection scope of the invention, and the phase change principle of the invention is described below by taking GST phase change materials as examples.

Refractive index data corresponding to GST materials with different amorphous degrees are imported into Lumerical software. The relationship between transmittance and amorphization was used to simulate synaptic behavior. Time domain finite difference (FDTD) simulations were performed on GST cells with 11 different degrees of amorphization (0, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%). The transmission characteristics of the MRR are shown in fig. 3 (a). As expected, the transmittance of each device decreases with increasing amorphization. The relation between the transmittance and the degree of amorphization was observed at a fixed wavelength (1.58913 μm), as shown in fig. 3 (b). It can be seen more clearly that the transmittance is positively correlated with the degree of amorphization.

As shown in fig. 4, the synaptic weight is represented by the transmittance of the PMRR cell. To achieve large-scale computation, MRR cell cascades of different radii are used to represent different synapses, which are then arranged in parallel to form an array of optical synapses.

Step 3: simulating optical pulse transmission characteristics of optical pulse neurons in an input layer and an output layer, and optimizing the PMRR unit to obtain an optimized PMRR unit;

referring to fig. 1 (b), compared with the PMRR cell before the optimization, the PMRR cell after the optimization further includes a control straight waveguide disposed on the PCM phase change cell, where the control straight waveguide is used for writing optical pulses, and controlling the timing at which the probe pulse of the input waveguide is detected, so as to output a peak signal through the output waveguide, thereby implementing the functions of the pulse neurons in the input layer and the output layer.

Further, the control straight waveguide is used for writing light pulses, and if the heat of the light pulses reaches a threshold value, the PCM phase change unit is in an amorphized state; in an amorphized state, the detection pulses of the input end waveguide are coupled through the annular waveguide, and the peak signal is output by the output end waveguide; after the output end waveguide outputs the peak signal, the control straight waveguide is used for resetting the PCM phase change unit to enable the PCM phase change unit to be in a crystallization state; in the crystallization state, the detection pulse is strongly coupled to the MRR unit through evanescent wave coupling by the input end waveguide and is absorbed by the phase change material, so that the output end waveguide does not output a pulse signal.

As shown in fig. 5, the neurons operate in alternating "write" and "read" cycles. The GST unit on the MRR is initially in a crystalline state. With an incident "write" pulse, the GST cell begins to be partially amorphized. In the "read" phase of the neuron, the MRR cannot transmit a probe pulse due to partial amorphization of the GST material. Until the GST material is completely amorphized (threshold is reached), a detection pulse is detected at the "Drop" port of the MRR. Once the neuron triggers, a "reset" pulse will reset the state of the device to an initial state, as shown in fig. 5, panel (b). When the GST unit is in the crystalline state, a suitable probe pulse is strongly coupled to the MRR through the "input" waveguide by evanescent coupling and absorbed by the GST material, so that no output pulse (peak) is observed. However, if the total instantaneous power of the weighted input pulses from the presynaptic neurons is high enough to switch the GST cell to an amorphous state, then a probe pulse will be output from the "Drop" port. Since the switching of the GST cell state occurs when a certain threshold power is reached, the neuron will only produce an output peak when the weighted sum of the input powers exceeds the threshold. Thus, the system naturally mimics the basic function of a biological neuron.

Step 4: simulating optical synaptic weights in the optical synaptic arrays using the PMRR units, simulating an output layer and optical pulse neurons of the output layer using optimized PMRR units, and simulating the beam combiner by a wavelength division multiplexing selector constructed by an MRR unit to obtain a photon pulse neural network based on the MRR and phase change materials.

Referring to fig. 1 (b), the photonic pulse neural network based on MRR and phase change material includes: an input unit, a WDM multiplexing selector, a beam splitter, a synaptic array unit, and an output unit;

each input unit comprises N optimized PMRR units, and each optimized PMRR unit corresponds to an optical pulse neuron of an input layer in the optical pulse neural network; the annular waveguide radii of the N optimized PMRR units are different; the WDM multiplexing selector is composed of N MRR units; the synaptic array is formed by M rows of PMRR unit arrays, and each row of PMRR unit array is formed by N optimized PMRRs of a shared input end waveguide and a shared output end waveguide; the output unit is composed of M optimized PMRR units;

the input end waveguide of each MRR unit in the WDM multiplexing selector is correspondingly connected with the output end waveguide of each PMRR unit, the output end waveguides of N MRR units in the WDM multiplexing selector are shared, the output end waveguide shared by N MRR units in the WDM multiplexing selector is connected to the input of the beam splitter, and the M-path output of the beam splitter is correspondingly connected with the input end waveguide of M rows of PMRR units in the synaptic array; the output end waveguides of the M rows of PMRR units in the synaptic array are correspondingly connected with the control straight waveguides of the optimized PMRR units of the output units.

The N-th optimized PMRR unit in the input unit is used for inputting pixel pulse signals of images through own input end waveguide, generating output pulses of the N-th wavelength according to the input pixel pulse signals when the energy of own control straight waveguide reaches a threshold value, and inputting the output pulses of the N-th wavelength into the input end waveguide of the N-th MRR in the straight WDM multiplexing selector through own output end waveguide;

the WDM multiplexing selector is used for fusing output pulses with N wavelengths into one pulse through N MRR units and inputting one pulse into the beam splitter;

the beam splitter is used for dividing one path of pulse into M paths of pulses according to the pulse energy, and inputting an mth path of pulse into an Mth PMRR unit in the synaptic array unit;

the M-th PMRR unit in the synaptic array unit is used for adjusting the weight of the input M-th pulse according to the incidence rate of the PMRR unit and generating the M-th pulse according to the result after the weight is adjusted; the Mth pulse is output to the control straight waveguide of the Mth optimized PMRR unit through the output end waveguide for weighting;

and the M-th optimized PMRR unit of the output unit is used for accumulating the energy of the input M-th pulse, generating a pulse signal when the accumulated energy reaches a threshold value, and outputting the pulse signal through an output end waveguide.

As shown in fig. 6, fig. 6 demonstrates a pattern recognition task based on 12 pointer directions of clock hour of full-connected photon SNN, and the constructed photon pulse neural network is composed of 400 presynaptic neurons (PRE) and 12 postsynaptic neurons (POST) which are connected in a full-connected manner, and weight training is performed through a ReSuMe supervised learning algorithm, and each pattern is represented by a pixel matrix with a size of 20×20, as shown in fig. 6 (b). The target output patterns of the 12 POST for each pattern are shown in fig. 6 (c). For transient mode "1", POST1 should trigger a peak while the other 11 POST remain silent. Also, for time of day mode "2" ("3" … "12"), only POST2 (POST 3 … POST 12) responds by issuing 1 pulse, while the rest of the POST remains silent.

The implementation process and implementation performance of the invention are described below through simulation experiments.

1. The MRR array implements a weighted function verification of the linear computation.

The coded pulse sequence was weighted with the configured 400 x 12PMRR array instead of the optical synapses and weighted and responded by post-synaptic neurons, resulting in the simulation test results shown in fig. 7. In fig. 7 (a) - (d), the evolution of pattern 1 over 4 different POST is shown, and it is found that POST1 can eventually produce a pulse at a certain time while the other POST remains silent as training iterates. The training process for the other 11 input modes in the POST is shown in FIGS. 7 (e) - (o). And after training convergence, obtaining a weight matrix. The test was then performed using 12 clean modes and noise modes, respectively, as shown in fig. 8. These 12 POSTs can output peaks that are exactly identical to the corresponding patterns shown in FIG. 9. For example, for mode "1", POST1 triggers a spike, while POSTs2-12 remain stationary. The result shows that the mode can be accurately classified by using the trained weight. The feasibility of linear calculations using 400 x 12MRR arrays was demonstrated, enabling hardware implementation of optical synapses as photonic pulse neural networks.

2. MRR and PCM integration realizes verification of photon pulse neural network.

The silicon optical MRR and the PCM based on SOI can realize photon pulse neural network, can be realized as hardware of optical synapses and optical neurons, execute linear calculation and nonlinear calculation functions in the pulse neural network, and have the feasibility of completing mode identification and target tracking tasks.

The invention provides a photon pulse neural network implementation method based on MRR and phase change materials, which comprises the steps of obtaining structural information of an optical pulse neural network; simulating the characteristics of light pulse neurons in an input layer and an output layer to accumulate light energy so as to emit light pulses, and constructing PMRR units for replacing the light pulse neurons; simulating the light pulse transmission characteristics of the light synaptic arrays for adjusting weight, optimizing the PMRR units, and constructing a synaptic array by using the optimized PMRR units; simulating optical synaptic weights in the optical synaptic arrays by using the transmittance of the PMRR units, simulating optical pulse neurons of an input layer and an output layer by using optimized PMRR units, and simulating the beam combiner by using a wavelength division multiplexing selector constructed by an MRR unit to obtain a photon pulse neural network based on the MRR and phase change materials. The PMRR unit designed by the invention can realize the accumulated excitation function of the optical pulse neurons on the input pulse, and can also realize the function of adjusting the weight of the optical synapse array, and the integrated realization of the feasibility of the photon pulse neural network. Compared with the prior art, the method has the characteristics of easier implementation, easier integration, low complexity, strong expandability and the like.

Furthermore, the terms "first," "second," and the like, are used for descriptive purposes only and are not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include one or more such feature. In the description of the present invention, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.

Although the present application has been described herein in connection with various embodiments, other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed application, from a review of the figures, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the "a" or "an" does not exclude a plurality.

The foregoing is a further detailed description of the invention in connection with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the invention, and these should be considered to be within the scope of the invention.

Claims (7)

1. The photon pulse neural network implementation method based on the MRR and the phase change material is characterized by comprising the following steps of:

step 1: acquiring structural information of an optical pulse neural network;

the structure information comprises neuron components, connection relations of the neurons and connection relations among different constituent units of the optical pulse neural network, wherein the optical pulse neural network comprises an input layer, a beam combiner, a beam splitter, an optical synapse array and an output layer;

step 2: simulating the magnitude of optical synapse control light transmittance in the optical synapse array, and constructing a PMRR unit for replacing the optical synapse array;

the PMRR unit comprises an MRR unit and a PCM phase change unit, wherein the PCM phase change unit is positioned on an annular waveguide of the MRR unit;

step 3: simulating optical pulse transmission characteristics of optical pulse neurons in an input layer and an output layer, and optimizing the PMRR unit to obtain an optimized PMRR unit;

step 4: simulating optical synaptic weights in the optical synaptic arrays using the PMRR units, simulating an output layer and optical pulse neurons of the output layer using optimized PMRR units, and simulating the beam combiner by a wavelength division multiplexing selector constructed by an MRR unit to obtain a photon pulse neural network based on the MRR and phase change materials.

2. The method for realizing the photonic pulse neural network based on the MRR and the phase change material according to claim 1, wherein the MRR unit is composed of 2 straight waveguides and 1 annular waveguide, the annular waveguide is positioned in the middle of the 2 straight waveguides, and one straight waveguide is an input end waveguide and the other straight waveguide is an output end waveguide; the PCM phase change unit is covered on the annular waveguide; the PCM phase change unit heats the phase change material of the PCM phase change unit by light pulse passing through the straight waveguide to cause phase change, so that the PCM phase change unit can be converted between crystalline state and amorphous state; the PCM phase change units with different crystallization degrees can control the intensity of light in the input straight waveguide of the MRR, so as to realize different synaptic intensities.

3. The method according to claim 2, wherein compared with the PMRR cell before the optimization, the PMRR cell after the optimization further comprises a control straight waveguide disposed on the PCM phase change cell, the control straight waveguide is used for writing optical pulses, and controlling the timing of detecting the pulse of the input waveguide, and outputting peak signals through the output waveguide, so as to realize the functions of the pulse neurons in the input layer and the output layer.

4. The method for implementing a photonic pulse neural network based on MRR and phase change material according to claim 3, wherein the control straight waveguide is used for writing optical pulse, and if the heat of the optical pulse reaches a threshold value, the PCM phase change cell is in an amorphized state; in an amorphized state, the detection pulses of the input end waveguide are coupled through the annular waveguide, and the peak signal is output by the output end waveguide; after the output end waveguide outputs the peak signal, the control straight waveguide is used for resetting the PCM phase change unit to enable the PCM phase change unit to be in a crystallization state; in the crystallization state, the detection pulse is strongly coupled to the MRR unit through evanescent wave coupling by the input end waveguide and is absorbed by the phase change material, so that the output end waveguide does not output a pulse signal.

5. The method for implementing a photonic pulse neural network based on MRR and phase change material according to claim 3, wherein the photonic pulse neural network based on MRR and phase change material comprises: an input unit, a WDM multiplexing selector, a beam splitter, a synaptic array unit, and an output unit;

each input unit comprises N optimized PMRR units, and each optimized PMRR unit corresponds to an optical pulse neuron of an input layer in the optical pulse neural network; the annular waveguide radii of the N optimized PMRR units are different; the WDM multiplexing selector is composed of N MRR units; the synaptic array is formed by M rows of PMRR unit arrays, and each row of PMRR unit array is formed by N optimized PMRRs of a shared input end waveguide and a shared output end waveguide; the output unit is composed of M optimized PMRR units;

the input end waveguide of each MRR unit in the WDM multiplexing selector is correspondingly connected with the output end waveguide of each PMRR unit, the output end waveguides of N MRR units in the WDM multiplexing selector are shared, the output end waveguide shared by N MRR units in the WDM multiplexing selector is connected to the input of the beam splitter, and the M-path output of the beam splitter is correspondingly connected with the input end waveguide of M rows of PMRR units in the synaptic array; the output end waveguides of the M rows of PMRR units in the synaptic array are correspondingly connected with the control straight waveguides of the optimized PMRR units of the output units.

6. The method for realizing the photonic pulse neural network based on the MRR and the phase change material according to claim 5, wherein,

the N-th optimized PMRR unit in the input unit is used for inputting pixel pulse signals of images through own input end waveguide, generating output pulses of the N-th wavelength according to the input pixel pulse signals when the energy of own control straight waveguide reaches a threshold value, and inputting the output pulses of the N-th wavelength into the input end waveguide of the N-th MRR in the straight WDM multiplexing selector through own output end waveguide;

the WDM multiplexing selector is used for fusing output pulses with N wavelengths into one pulse through N MRR units and inputting one pulse into the beam splitter;

the beam splitter is used for dividing one path of pulse into M paths of pulses according to the pulse energy, and inputting an mth path of pulse into an Mth PMRR unit in the synaptic array unit;

the M-th PMRR unit in the synaptic array unit is used for adjusting the weight of the input M-th pulse according to the incidence rate of the PMRR unit and generating the M-th pulse according to the result after the weight is adjusted; the Mth pulse is output to the control straight waveguide of the Mth optimized PMRR unit through the output end waveguide for weighting;

and the M-th optimized PMRR unit of the output unit is used for accumulating the energy of the input M-th pulse, generating a pulse signal when the accumulated energy reaches a threshold value, and outputting the pulse signal through an output end waveguide.

7. The method of any one of claims 1-6, wherein the phase change material of the PCM phase change cell comprises a GST phase change material.

Images