Coherent photon pulse neural network device based on FP-SA and MZM

Abstract

The present invention discloses a coherent photon pulse neural network device based on FP-SA and MZM, comprising: a DQPSK MZM for receiving two electrical pulse signals, performing electro-optical conversion on the two electrical pulse signals through two MZMs and simultaneously assigning a weight to each converted photon pulse signal; summing the two weighted photon pulse signals through a phase shifter with zero phase shift and an optical coupler at the output end for combining, thereby obtaining a first optical signal representing the weighted summation result; the first optical signal enters through an optical circulator, the FP-SA performs a nonlinear response on the first optical signal to implement a nonlinear operation, and a second optical signal representing the result of the nonlinear operation is output externally through the optical circulator. Based on the device provided by the present invention, a photon pulse neural network with a higher processing linear rate, easy industrial application, and reconfigurable weight parameters can be realized.

Description

Technical Field

[0001] The present invention belongs to the field of optical computing, and in particular relates to a coherent photon pulse neural network device based on FP-SA and MZM. Background Art

[0002] As Moore's Law reaches its limits, traditional computing architectures are increasingly unable to meet the demands of continued growth in computing power. In particular, with the rapid development of artificial intelligence, even the computing power of integrated circuits is struggling to cope with the rapidly increasing volume of data. Neuromorphic computing promises to transcend traditional computing, achieving higher computing power and lower power consumption. Several neuromorphic processors, such as TrueNorth, Spinnaker, and Loihi, have already been developed and demonstrated remarkable computing capabilities.

[0003] Due to the inherent advantages of light in high speed, broadband and low electromagnetic interference, neuromorphic computing has been rapidly and effectively developed in the field of optics, and photonic neural computing has become a promising approach to overcome the von Neumann bottleneck.

[0004] To date, a number of coherent photon pulse neural networks have been theoretically and experimentally realized based on discrete devices and integrated solutions. For example, coherent photon pulse neural networks based on MZI (Mach–Zehnder Interferometer), coherent photon pulse neural networks based on hybrid silicon waveguides, and coherent photon pulse neural networks based on PCM (Phase change material, waveguide integrated phase change material).

[0005] Among them, the coherent photon pulse neural network implemented based on MZI has a low processing linear rate and has strict requirements on the phase parameters of MZI; the coherent photon pulse neural network implemented based on hybrid silicon waveguide has high requirements on device preparation processes such as doping ratio, and has a small tolerance error limit, making it difficult to achieve industrial application; and the coherent photon pulse network device implemented based on PCM cannot be reconstructed once its weight parameters are set, and its utilization efficiency is low.

[0006] Therefore, further exploration is needed to develop photonic pulse neural networks with higher processing linear rates, ease of industrial application, and reconfigurable weight parameters. Summary of the Invention

[0007] In order to solve the above problems existing in the prior art, the present invention provides a coherent photon pulse neural network device based on FP-SA (a two-stage Fabry-Perot laser with a saturable absorber) and MZM (Mach-Zehnder modulator).

[0008] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0009] A coherent photon pulse neural network device based on FP-SA and MZM, comprising: DQPSK MZM, optical circulator and FP-SA; wherein,

[0010] The DQPSK MZM is configured to: receive two external electrical pulse signals; perform electrical-to-optical conversion on the two electrical pulse signals through the two MZMs contained therein, and simultaneously assign weights to each converted photon pulse signal; sum the two weighted photon pulse signals through a phase shifter having a zero phase shift connected after either of the two MZMs and an optical coupler at the output end for combining, to obtain a first optical signal representing the weighted summation result; wherein the weights are assigned to the photon pulse signals by configuring the bias voltages of the MZMs;

[0011] The first optical signal enters the FP-SA through the optical circulator;

[0012] The FP-SA is configured to perform a nonlinear response on the first optical signal to implement a nonlinear operation, and output a second optical signal representing a result of the nonlinear operation through the optical circulator.

[0013] Optionally, the two electrical pulse signals arrive at the DQPSK MZM at non-simultaneous times.

[0014] Optionally, configuring the bias voltage of the MZM to assign a weight to the photon pulse signal includes:

[0015] According to the weight ratio W1:W2 to be assigned to the two photon pulse signals, the bias voltages corresponding to the two MZMs are searched from pre-measured experimental data when the ratio of the power outputs of the two MZMs is equal to W1:W2;

[0016] According to the query result, the bias voltages of the two MZMs are configured.

[0017] Optionally, the FP-SA is configured to operate below a Q-switched pulse threshold;

[0018] The Q-switched pulse threshold is defined as: the minimum gain region current at which the FP-SA is stimulated to produce a Q-switched pulse state under the condition that the FP-SA has no external light input and operates at a given reverse bias voltage.

[0019] Optionally, the reverse bias voltage has a value range of -6V to 0V.

[0020] Optionally, the device further comprises: two radio frequency amplifiers;

[0021] The radio frequency amplifier is used to amplify the electrical pulse signal to drive the MZM.

[0022] Optionally, the device further comprises: an adjustable optical attenuator and a polarization controller;

[0023] The adjustable optical attenuator is used to adjust the optical power of the first optical signal;

[0024] The polarization controller is used to adjust the polarization state of the first optical signal.

[0025] Optionally, the MZM is a titanium-diffused lithium niobate single-drive MZM.

[0026] The present invention also provides a coherent photon pulse neural network chip based on FP-SA and MZM, which includes any of the above-mentioned coherent photon pulse neural network devices based on FP-SA and MZM.

[0027] The present invention also provides an optical computing method, comprising:

[0028] Prepare any one of the above-mentioned coherent photon pulse neural network devices based on FP-SA and MZM, and determine the weight parameters of the device;

[0029] Determining and configuring bias voltages of two MZMs included in the DQPSK MZM in the device according to the weight parameters;

[0030] Two electrical pulse signals are input to the DQPSK MZM to implement the following optical calculations using the device:

[0031] The DQPSK MZM performs electro-optical conversion on the two electrical pulse signals and simultaneously assigns a weight to each converted photon pulse signal; wherein, the weights corresponding to the two photon pulse signals respectively constitute the weight parameters; a phase shifter with a zero phase shift and an optical coupler for combining at the output end, connected after any one of the two MZMs, sums the two weighted photon pulse signals to obtain a first optical signal representing the weighted summation result; the first optical signal is sent to the FP-SA by the optical circulator; the FP-SA performs a nonlinear response on the first optical signal to implement a nonlinear operation, and a second optical signal representing the nonlinear operation result is output through the optical circulator.

[0032] In the coherent photon pulse neural network device based on FP-SA and MZM provided by the present invention, the DQPSK MZM receives two external electrical pulse signals, performs electro-optical conversion on the two electrical pulse signals through its two contained MZMs, and simultaneously assigns a weight to each converted photon pulse signal. A phase shifter with a zero phase shift connected after either of the two MZMs and an optical coupler for combining the two signals at the output end are used to sum the two weighted photon pulse signals to obtain a first optical signal representing the weighted summation result. This first optical signal enters the FP-SA through the optical circulator, causing the FP-SA to perform a nonlinear response to the first optical signal to implement a nonlinear operation. A second optical signal representing the result of the nonlinear operation is output externally through the optical circulator. The present invention uses the DQPSK MZM as a photonic synapse to implement linear weighted summation calculations in the coherent photon pulse neural network. The DQPSK MZM, FP-SA, and optical circulator are all in-production components, so the device is compatible with existing production processes and has the conditions for industrial application. The weights of the coherent photon pulse neural network can be configured by configuring the bias voltages of the two MZMs in the DQPSK MZM, making the device weight reconfigurable. Furthermore, because the MZM is an active device, it can achieve a higher processing linear rate than the passive MZI.

[0033] In summary, based on the device provided by the present invention, a photon pulse neural network with higher processing linear rate, easy industrial application and reconfigurable weight parameters can be realized.

[0034] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of a coherent photon pulse neural network device based on FP-SA and MZM provided by an embodiment of the present invention;

[0036] Figure 2 It is a structural diagram of DQPSK MZM;

[0037] Figure 3 A micrograph of the chip of FP-SA is shown in FIG;

[0038] Figure 4 yes Figure 3 The spectrum of FP-SA shown in Figure 1 is free running;

[0039] Figure 5 yes Figure 3 The timing diagram of the FP-SA being excited to the Q-switched pulse state;

[0040] Figure 6 yes Figure 3 The power-current curves of the FP-SA at different reverse bias voltages are shown;

[0041] Figure 7 yes Figure 1 A diagram showing the working principle of the device shown;

[0042] Figure 8 An exemplary modulation response curve of a DQPSK MZM is shown;

[0043] Figure 9 Schematic diagram of another coherent photon pulse neural network device based on FP-SA and MZM provided by an embodiment of the present invention;

[0044] Figure 10 This is a schematic structural diagram of another coherent photon pulse neural network device based on FP-SA and MZM provided by an embodiment of the present invention;

[0045] Figure 11 The figure shows an experimental platform constructed to verify the performance of the device provided by the embodiment of the present invention;

[0046] Figure 12 The verification result of the coherent photon pulse neural network implemented by the device provided by the embodiment of the present invention is shown to have the time domain accumulation characteristic;

[0047] Figure 13 The verification result of the threshold characteristic of the coherent photon pulse neural network implemented by the device provided by the embodiment of the present invention is shown;

[0048] Figure 14 The verification result of the coherent photon pulse neural network implemented by the device provided by the embodiment of the present invention is shown to have a refractory period characteristic. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0050] In order to further explore the photon pulse neural network with higher processing linear rate, easy to realize industrial application and reconfigurable weight parameters, the embodiment of the present invention provides a coherent photon pulse neural network device based on FP-SA and MZM, such as Figure 1 As shown in the figure, the device includes a DQPSK MZM, an optical circulator, and an FP-SA. The DQPSK MZM is a differential quadrature phase-shift keying Mach-Zehnder modulator, and the FP-SA is a two-stage Fabry-Perot laser formed by introducing a saturable absorber (SA) into a Fabry-Perot (FP) cavity, referred to as an FP-SA.

[0051] The DQPSK MZM is used to receive two external electrical pulse signals V in1 and V in2 ; The two electrical pulse signals are respectively converted into optical signals through the two MZMs contained in the device, and a weight is assigned to each converted photon pulse signal; the two weighted photon pulse signals are summed by a phase shifter with 0 phase shift connected after any one of the two MZMs and an optical coupler for combining at the output end, to obtain a first optical signal S1 representing the weighted summation result; wherein the weight is assigned to the photon pulse signal by configuring the bias voltage (V1 and V2) of the two MZMs.

[0052] Specifically, the structure of DQPSK MZM is as follows Figure 2 As shown, it contains two MZMs, two optical couplers OC, and a phase shifter PS with zero phase shift connected to either of the two MZMs. The DQPSK MZM is equipped with two RF input ports, two voltage bias ports, one optical input port, and one optical output port. The two RF input ports are used to receive two external electrical pulse signals V in1 and V in2 The two electrical pulse signals enter the two MZMs for electro-optical modulation. The two voltage bias ports are used to provide bias voltages for the two MZMs. By configuring the sizes of V1 and V2, the weight parameters of the photon pulse neural network can be modulated onto the amplitude envelopes of the photon pulse signals output by the two MZMs. The optical input port is used to receive the external input continuous optical carrier CW. The continuous optical carrier CW entering the DQPSK MZM is split into two paths by the optical coupler OC1 and sent to the two MZMs respectively. After passing through the two MZMs, both electrical pulse signals are modulated into photon pulse signals, and both photon pulse signals are simultaneously assigned weights. Then, a phase shifter PS is used to generate a zero phase shift, achieving constructive interference between the two weighted photon pulse signals. These signals are then combined by the optical coupler OC2 at the output end to achieve a summation operation. The resulting first optical signal S1, representing the weighted summation result, is output from the optical output end of the DQPSK MZM to the outside of the DQPSK MZM.

[0053] In practical applications, the method of configuring the bias voltage of the MZM to assign weights to the photon pulse signal specifically includes: based on the ratio W1:W2 of the weights to be assigned to the two photon pulse signals, querying from pre-measured experimental data the bias voltages corresponding to the two MZMs in the DQPSKMZM when the ratio of the power outputs of the two MZMs is equal to W1:W2; then, configuring the bias voltages of the two MZMs according to the query results.

[0054] The process of pre-measuring the experimental data includes iterating over the bias voltages V1 and V2 of the two MZMs in the DQPSK MZM to obtain the output power ratio P1:P2 of the two MZMs for each combination of V1 and V2, thus forming the experimental data. Therefore, when the weight ratio W1:W2 is desired for the two photon pulse signals, the experimental data can be used to find the P1:P2 that is equal to W1:W2, thereby obtaining the corresponding V1 and V2.

[0055] Continue to see Figure 1 As shown, a first optical signal S1 enters the FP-SA through an optical circulator. Specifically, the first optical signal S1 enters the input port of the optical circulator and then enters the FP-SA through the output port of the optical circulator. The optical circulator is a three-port device that also has a coupling port. In practice, this coupling port can be connected to a matched load.

[0056] The FP-SA is configured to perform a nonlinear response on the first optical signal to implement a nonlinear operation, and output a second optical signal representing the result of the nonlinear operation through the optical circulator.

[0057] The FP-SA is a PIN structure grown based on AlGaInAs (aluminum gallium indium arsenide) / InP (indium phosphide) materials. The microstructure of the chip is as follows: Figure 3 As shown in the figure, the laser consists of a gain region and a saturation absorption region, with an electrical isolation region between the two regions. Each region is provided with electrodes. The gain region electrode is connected to an external bias current (gain region current), and the saturation absorption region electrode is connected to an external reverse bias voltage (reverse bias voltage). External light enters the laser and is excited inside the laser to produce laser light output. Figure 3 The square icon is used to indicate the tail of the laser. The light reflectivity of the laser tail end face is about 95%. The triangle icon is used to indicate the light output direction of the laser. The light reflectivity of the laser light outlet end face is about 30%.

[0058] Figure 4 The output spectrum of FP-SA during free operation is shown, wherein the horizontal axis represents the wavelength and the vertical axis represents the output power of FP-SA. It can be seen from its output spectrum that FP-SA is a multi-longitudinal mode laser.

[0059] Preferably, the FP-SA is configured to operate below the Q-switched pulse threshold, i.e., the gain region current of the FP-SA is configured to be below the Q-switched pulse threshold. Here, the Q-switched pulse threshold is defined as the minimum gain region current that causes the FP-SA to be stimulated into a Q-switched pulse state (i.e., a self-pulsing state) when there is no external light input to the FP-SA and the FP-SA is operating at a given reverse bias voltage. Figure 5The timing diagram shows the FP-SA working under a certain reverse bias voltage and being stimulated to produce a Q-switched pulse state, wherein the horizontal axis represents time and the vertical axis represents the voltage value corresponding to the output power of the FP-SA after photoelectric conversion.

[0060] In practical applications, when FP-SA is used as a neuron, the current in the gain region of FP-SA should not exceed the Q-switched pulse threshold corresponding to the currently set reverse bias voltage, and should be sufficient to enable FP-SA to enter the excited state. Therefore, before starting to use FP-SA as a neuron, you can test the power and current curves of FP-SA under different reverse bias voltages (such as Figure 6 As shown), a reverse bias voltage is selected from it, and then the Q-switched pulse threshold is determined based on the reverse bias voltage, and the gain region current of the FP-SA is determined accordingly. In addition, Figure 6 It can be seen that the larger the reverse bias voltage is, the larger the Q-switched pulse threshold of the FP-SA is.

[0061] Preferably, the reverse bias voltage of the FP-SA ranges from -6V to 0V.

[0062] See also Figure 7 As shown, in the coherent photon pulse neural network implemented by the device provided by the embodiment of the present invention, two electrical pulse signals are used to simulate the output of the presynaptic neuron, DQPSK MZM is used as the photon synapse to implement the linear weighted summation (represented by SUM) calculation in the coherent photon pulse neural network, and FP-SA is used as the postsynaptic neuron (represented by POST) to implement the nonlinear operation in the coherent photon pulse neural network.

[0063] Among them, in order to better simulate the characteristics of signal transmission in neural networks, the two electrical pulse signals V in1 and V in2 Can be configured to arrive non-simultaneously DQPSK MZM, for example Figure 7 In the circuit, there is an electrical delay in one of the electrical pulse signals. Of course, even if the two electrical pulse signals V in1 and V in2 Configured to arrive at the DQPSK MZM simultaneously, the coherent photon pulse neural network device provided by the embodiment of the present invention can still be used to perform optical computing.

[0064] To achieve a higher processing linear rate, a Ti:LiNbO3 (titanium-diffused lithium niobate) DQPSK MZM with a modulation speed of up to 22.5 Gbaud can be used. All of these MZMs are single-drive titanium-diffused lithium niobate MZMs. This DQPSK MZM uses a high-speed traveling-wave MZM for linear modulation and a phase shifter for summation, thereby achieving a higher processing linear rate. Figure 8Figure 2 shows the modulation response curve of this DQPSK MZM; among them, sub-figure (a) shows the output power change trend of the DQPSK MZM when V1 = 4.5V and V2 traverses from 0V to 20V, sub-figure (b) shows the output power change trend of the DQPSK MZM when V2 = 4.5V and V1 traverses from 0V to 30V, and sub-figure (c) shows the output power change trend of the DQPSK MZM when V2 = V1 and the bias voltage traverses from 0V to above 20V.

[0065] In one embodiment, when the power of the electrical pulse signal is low, such as Figure 9 As shown, the coherent photon pulse neural network device provided by the embodiment of the present invention may also include: two radio frequency amplifiers EA; the radio frequency amplifiers are used to amplify the electrical pulse signals input from the outside to drive the MZM.

[0066] In another embodiment, Figure 10 As shown, the coherent photon pulse neural network device provided in this embodiment of the present invention may further include: a variable optical attenuator (VOA) and a polarization controller (PC). The variable optical attenuator (VOA) is used to adjust the optical power of the first optical signal S1, and the polarization controller (PC) is used to adjust the polarization state of the first optical signal, so that the optical power and polarization state of the first optical signal entering the FP-SA through the optical circulator match the input of the FP-SA.

[0067] exist Figure 1 Based on the coherent photon pulse neural network device shown in FIG, the embodiment of the present invention also provides another coherent photon pulse neural network device, which includes Figure 9 The two RF amplifiers EA in the device shown also include Figure 10 The device shown includes a variable optical attenuator (VOA) and a polarization controller (PC).

[0068] In the process of implementing the present invention, the inventors conducted a series of experiments to verify the performance of the coherent photon pulse neural network implemented by the device provided in the embodiment of the present invention. The experimental process and results are described in detail below.

[0069] Build an experimental platform, such as Figure 11 As shown in the figure, DQPSK MZM is used as a photon synapse to realize the linear computing function of the coherent photon pulse neural network, and FP-SA is used as a photon neuron to realize the nonlinear computing function of the coherent photon pulse neural network. The two channels of the arbitrary waveform generator AWG (model Tektronix AWG70001A) generate two electrical pulse signals V in1 and V in2, one of the electrical pulse signals is delayed relative to the other, in order to simulate the signal transmission mechanism in a real neural network. In addition, two RF amplifiers EA are used to amplify the two electrical pulse signals to drive the two MZMs of the DQPSK MZM. Here, two electrical pulse signals V in1 and V in2 The FP-SA is used to simulate the outputs of two presynaptic neurons, while the POST is simulated using a tunable laser TL (AQ2200-136 TLS tunable light source module) operating in a continuous wave (CW) state to output a continuous optical carrier. A polarization controller, PC2, adjusts the polarization state of the TL output to match the DQPSK MZM. Within the DQPSK system, the TL output is split into two identical copies, which are then fed into two MZMs. The continuous optical carrier performs electro-optical modulation on the two electrical pulse signals, generating photon pulses. Simultaneously, weight information is added to the photon pulses using the bias voltages of the two MZMs, generating two weighted photon pulses. After modulation by the two MZMs, the two weighted photon pulses are weighted and summed based on the interference between the coherent light fields. The signals are then combined at the DQPSK output. The first optical signal representing the weighted summation result is then injected into the FP-SA. At this time, the variable optical attenuator (VOA) can be used to adjust the injection intensity of the first optical signal, and the polarization controller PC2 can be used to adjust the polarization state of the first optical signal injected into the FP-SA to match the FP-SA.

[0070] The bias voltages of MZM1 and MZM2 and the bias of the phase shifter PS are provided by a DC power supply. The reverse bias voltage of the FP-SA is connected to an external voltage source, and the gain region current is provided by an external laser diode controller. The laser diode controller can also control the operating temperature of the FP-SA.

[0071] To facilitate measurement and analysis, an optical coupler (OC) was used to couple the first optical signal, which was then connected to an optical power meter (PM) for optical power measurement. Another optical coupler (OC) was used to split the FP-SA output into two paths for analysis. One path was connected to an optical spectrum analyzer (OSA) (Advantest Q8384) for spectral analysis, while the other path was converted to an electrical signal using a photodetector (PD) (Agilent / HP11982a). This signal was then connected to an oscilloscope (OSC) (Keysight DSOV334A) for timing analysis.

[0072] The experimental process and results are as follows:

[0073] Experiment 1: Input electrical pulse signals to the two MZMs of the DQPSK MZM respectively. The electrical pulse signal input to MZM1 is as follows: Figure 12 As shown in the neutron diagram (a), the electrical pulse signal input to MZM2 is as follows: Figure 12 As shown in the neutron diagram (b), the electrical pulse signal of this path has an electrical delay relative to the electrical pulse signal in the input terminal MZM1. The electrical domain observation result of the weighted summation of these two electrical pulse signals after DQPSKMZM is as follows: Figure 12 As shown in the neutron diagram (c), there are two closely spaced weak pulses, and two weak perturbation pulses of the same power located before and after the closely spaced pulses. By adjusting the above-mentioned electrical delay, the peak interval between the two closely spaced weak perturbation pulses is set to 0.6ns, thereby triggering the FP-SA to generate a neuron response output spike, as shown in Figure 2. Figure 12 As shown in sub-figure (d) in the figure, the pulses before and after this group of closely spaced pulses do not generate response pulses. This shows that although a single weak pulse cannot reach the spike threshold, two closely spaced weak pulses are integrated in time and therefore exceed the threshold. This shows that the coherent optical neural network implemented based on the embodiment of the present invention has a time integration characteristic, also known as a time domain accumulation characteristic.

[0074] Furthermore, in order to prove that the results of Experiment 1 above are not accidental and reproducible, 100 identical electrical pulse signals were continuously applied to the coherent photon pulse neural network device to stimulate it. The experimental results are as follows: Figure 12 As shown in sub-figure (e) in , the horizontal axis represents the repetition period of the electrical pulse signal applied to the coherent photon pulse neural network, and the vertical axis represents the time axis of a single cycle; it can be seen that the same spike response is obtained in each cycle.

[0075] Experiment 2: Input the following to MZM1: Figure 13 The electric pulse signal shown in the neutron diagram (a) is input to MZM2 as shown in Figure 13 The electric pulse signal shown in the neutron image (b) is the result of weighted summation of these two electric pulse signals after DQPSK MZM. Figure 13 Neutron image (c) shows three pulses with different powers. After these three pulses were injected into the FP-SA, only the second pulse with higher power triggered the FP-SA neuron to produce a spike response, while the first and third pulses did not produce a response spike because they did not exceed the triggering threshold of the FP-SA. This shows that the coherent optical neural network implemented based on the embodiment of the present invention has threshold characteristics.

[0076] Similarly, in order to prove that the results of Experiment 2 above are not accidental and reproducible, 100 identical electrical pulse signals were continuously applied to the coherent photon pulse neural network device to stimulate it. The experimental results are as follows: Figure 13 As shown in the sub-figure (e) in , it can be seen that the experimental results are consistent each time.

[0077] Experiment 3: Input the following to MZM1: Figure 14 The electric pulse signal shown in the neutron diagram (a) is input to MZM1 as shown in Figure 14 The electric pulse signal shown in the neutron image (b) is the result of weighted summation of these two electric pulse signals after DQPSK MZM. Figure 14 Neutron image (c) shows two strong perturbation pulses and a set of two closely spaced strong perturbation pulses. When these three pulses are injected into the FP-SA, a single strong perturbation pulse can trigger a response spike, while four pulses only trigger three response spikes, demonstrating that the coherent optical neural network implemented based on this embodiment of the present invention has a refractory period characteristic.

[0078] Similarly, in order to prove that the results of Experiment 3 above are not accidental and reproducible, 100 identical electrical pulse signals were continuously applied to the coherent photon pulse neural network device to stimulate it. The experimental results are as follows: Figure 14 As shown in the sub-figure (e) in , it can be seen that the experimental results are consistent each time.

[0079] In an embodiment of the present invention, a DQPSK MZM is used as a photon synapse to implement linear weighted summation calculations in a coherent photon pulse neural network. The DQPSK MZM, optical circulator, and FP-SA are all in-production devices, so the device is compatible with existing production processes and has the conditions for industrial application. The weights of the coherent photon pulse neural network can be achieved by configuring the bias voltages of the two MZMs in the DQPSK MZM, so the device has the characteristic of reconfigurable weights. Moreover, since the MZM is an active device, it can achieve a higher processing linear rate than the passive MZI. Therefore, based on the device provided by the present invention, a photon pulse neural network with a higher processing linear rate, easy industrial application, and reconfigurable weight parameters can be achieved.

[0080] Based on the same inventive concept, an embodiment of the present invention also provides a coherent photon pulse neural network chip based on FP-SA and MZM, which includes any of the above-mentioned coherent photon pulse neural network devices based on FP-SA and MZM.

[0081] In actual applications, the chip has two external electrical signal input ports for receiving two external electrical pulse signals; it also includes an optical input port for receiving an external continuous optical carrier; it also includes two voltage input ports for applying bias voltage to the two MZMs in the DQPSK MZM; and it also includes an optical output port for outputting a second optical signal representing the result of the nonlinear operation output after the FP-SA response.

[0082] It is understandable that the chip can be produced or manufactured by referring to the production and manufacturing processes of DQPSK MZM and FP-SA currently in production, and the embodiment of the present invention does not limit the production process / manufacturing process.

[0083] Based on the coherent photon pulse neural network device provided in an embodiment of the present invention, an embodiment of the present invention further provides an optical computing method, including:

[0084] (1) obtaining any of the aforementioned coherent photon pulse neural network devices based on FP-SA and MZM, and determining weight parameters of the device;

[0085] (2) Determine and configure the bias voltages of the two MZMs included in the DQPSK MZM in the device according to the weight parameters;

[0086] (3) Two electrical pulse signals are input to the DQPSK MZM to implement the following optical calculation using the device: the DQPSK MZM performs electrical-optical conversion on the two electrical pulse signals and simultaneously assigns weights to each converted photon pulse signal; wherein, the weights corresponding to the two photon pulse signals constitute weight parameters; a phase shifter with a zero phase shift connected after any of the two MZMs sums the two weighted photon pulse signals to obtain a first optical signal representing the weighted summation result; the first optical signal is sent to the FP-SA by the optical circulator; the FP-SA performs a nonlinear response on the first optical signal to implement a nonlinear operation, and a second optical signal representing the nonlinear operation result is output through the optical circulator.

[0087] It is understandable that, for the optical computing method embodiment, since its implementation process has been described in detail in the device embodiment, the description here is relatively simple, and the relevant parts can be referred to the partial description of the device embodiment.

[0088] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0089] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0090] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art can understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims.

[0091] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A coherent photon pulse neural network device based on FP-SA and MZM, characterized in that: include: DQPSK MZM, optical circulator and FP-SA; among them, The DQPSK MZM is configured to: receive two external electrical pulse signals; perform electrical-to-optical conversion on the two electrical pulse signals through the two MZMs contained therein, and simultaneously assign weights to each converted photon pulse signal; sum the two weighted photon pulse signals through a phase shifter having a zero phase shift connected after either of the two MZMs and an optical coupler at the output end for combining, to obtain a first optical signal representing the weighted summation result; wherein the weights are assigned to the photon pulse signals by configuring the bias voltages of the MZMs; The first optical signal enters the FP-SA through the optical circulator; The FP-SA is configured to perform a nonlinear operation on the first optical signal by performing a nonlinear response, and output a second optical signal representing a result of the nonlinear operation through the optical circulator; Among them, DQPSK MZM acts as a photon synapse to realize linear weighted summation calculation in coherent photon pulse neural network.

2. The coherent photon pulse neural network device according to claim 1, characterized in that: The two electrical pulse signals arrive at the DQPSK MZM at non-simultaneous times.

3. The coherent photon pulse neural network device according to claim 1, characterized in that: The method of configuring the bias voltage of the MZM to assign a weight to the photon pulse signal includes: According to the weight ratio W1:W2 to be assigned to the two photon pulse signals, the bias voltages corresponding to the two MZMs are searched from pre-measured experimental data when the ratio of the power outputs of the two MZMs is equal to W1:W2; According to the query result, the bias voltages of the two MZMs are configured.

4. The coherent photon pulse neural network device according to claim 1, characterized in that: The FP-SA is configured to operate below a Q-switched pulse threshold; The Q-switched pulse threshold is defined as: the minimum gain region current at which the FP-SA is stimulated to produce a Q-switched pulse state under the condition that the FP-SA has no external light input and operates at a given reverse bias voltage.

5. The coherent photon pulse neural network device according to claim 4, characterized in that: The reverse bias voltage has a value range of -6V to 0V.

6. The coherent photon pulse neural network device according to claim 1, characterized in that: Also includes: two RF amplifiers; The radio frequency amplifier is used to amplify the electrical pulse signal to drive the MZM.

7. The coherent photon pulse neural network device according to claim 1, characterized in that: Also includes: Variable optical attenuators and polarization controllers; The adjustable optical attenuator is used to adjust the optical power of the first optical signal; The polarization controller is used to adjust the polarization state of the first optical signal.

8. The coherent photon pulse neural network device according to claim 1, characterized in that: The MZM is a titanium-diffused lithium niobate single-drive MZM.

9. A coherent photon pulse neural network chip based on FP-SA and MZM, characterized in that: It comprises a coherent photon pulse neural network device based on FP-SA and MZM as described in any one of claims 1 to 8.

10. A light computing method, characterized in that: include: Prepare a coherent photon pulse neural network device based on FP-SA and MZM as described in any one of claims 1 to 8, and determine the weight parameters of the device; Determining and configuring bias voltages of two MZMs included in the DQPSK MZM in the device according to the weight parameters; Two electrical pulse signals are input to the DQPSK MZM to implement the following optical calculations using the device: The DQPSK MZM performs electro-optical conversion on the two electrical pulse signals and simultaneously assigns a weight to each converted photon pulse signal; wherein, the weights corresponding to the two photon pulse signals respectively constitute the weight parameters; a phase shifter with a zero phase shift and an optical coupler for combining at the output end, connected after any one of the two MZMs, sums the two weighted photon pulse signals to obtain a first optical signal representing the weighted summation result; the first optical signal is sent to the FP-SA by the optical circulator; the FP-SA performs a nonlinear response on the first optical signal to implement a nonlinear operation, and a second optical signal representing the nonlinear operation result is output through the optical circulator.

Images