A neural network accelerator based on a multimode interferometer and coherent detection

Abstract

The application provides a neural network accelerator based on a multimode interferometer and coherent detection, relates to the field of optical neural networks, and comprises a semiconductor laser, a multimode interferometer module, a coupler module and a balanced detector module, wherein the semiconductor laser module provides an optical signal; the multimode interferometer module divides the optical signal into multiple paths and adjusts the phase of the optical signal, and outputs a processing optical signal and a reference optical signal; the coupler module is used for realizing coherent detection of the processing optical signal and the reference optical signal, and coupling the processing optical signal and the reference optical signal into the same waveguide; and the balanced detector module performs beat frequency on the processing optical signal and the reference optical signal transmitted in the same waveguide, and converts the beat frequency optical signal into an electrical signal as an input signal of the next layer of an artificial neural network. The application introduces a multimode interferometer, introduces an electrode in the multimode interference area, and realizes a neural network in a complex domain in combination with coherent detection, so that the fitting capacity of the network is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical neural networks, and in particular to a neural network accelerator based on a multimode interferometer and coherent detection. BACKGROUND

[0002] Current optical neural network implementation schemes can be mainly divided into a cascaded Mach-Zehnder interferometer scheme and a wavelength-based scheme. The cascaded Mach-Zehnder interferometer scheme is based on the principle of matrix decomposition, which decomposes a matrix into the product of three matrices, and the three matrices are composed of the transmission matrices of three different Mach-Zehnder interferometer cascaded networks. The wavelength-based scheme uses wavelength division multiplexing technology to connect multiple micro-ring modulators, and each micro-ring modulator corresponds to a different resonant wavelength. Adjusting the electrical signal of a micro-ring modulator only changes the optical signal of the resonant wavelength of the micro-ring modulator.

[0003] The complexity of the devices required by the cascaded Mach-Zehnder interferometer scheme is in a square relationship with the dimension of the implemented matrix, and the size of the Mach-Zehnder interferometer device is relatively large, on the order of several hundred microns. Therefore, it is difficult to implement a larger scale integration using the coherent scheme. Although the scheme based on micro-ring modulators can integrate a larger system due to the smaller size of the devices used, the resonant wavelength of the micro-ring modulator usually needs to be controlled by adding an additional heater due to the sensitivity of the micro-ring to the process. Therefore, the scheme based on micro-ring modulators consumes more energy. SUMMARY

[0004] To solve the above problems, a neural network accelerator based on a multimode interferometer and coherent detection is proposed. By introducing a process error-insensitive multimode interferometer and introducing multiple electrodes in the multimode interference region of the multimode interferometer, matrix operations can be completed with smaller sizes, and a programmable optical neural network can be obtained, which realizes a more compact structure and increases the fitting ability of the network.

[0005] The present application proposes a neural network accelerator based on a multimode interferometer and coherent detection, which includes a semiconductor laser, a multimode interferometer module, a coupler module and a balanced detector module, wherein,

[0006] The semiconductor laser, the multimode interferometer module, the coupler module and the balanced detector module are connected in sequence;

[0007] The semiconductor laser module provides an optical signal for the multimode interferometer module;

[0008] The multimode interferometer module divides the optical signal into multiple paths and adjusts the phase of the optical signal, and outputs a processed optical signal and a reference optical signal to the coupler module;

[0009] The coupler module is used to realize coherent detection of the processing optical signal and the reference optical signal, and couple the processing optical signal and the reference optical signal to the same waveguide;

[0010] The balanced detector module beats the processing optical signal and the reference optical signal transmitted in the same waveguide, and converts the beat optical signal into an electrical signal as an input signal of the next layer of artificial neural network.

[0011] Optionally, the semiconductor laser includes:

[0012] A distributed feedback semiconductor laser;

[0013] A distributed Bragg semiconductor laser.

[0014] Optionally, the multi-mode interferometer module includes:

[0015] A first multi-mode interferometer module divides the optical signal into N+1 paths, determines N paths of the processing optical signal and the reference optical signal of the N+1 path;

[0016] A second multi-mode interferometer module, the second multi-mode interferometer module is a multi-mode interferometer with N inputs and N outputs, used to perform matrix operation and output the processing optical signal;

[0017] A first phase shifter module for acting on the input waveguide of the second multi-mode interferometer module to adjust the phase of the processing optical signal, wherein the electrical signal of the first phase shifter module is the value of the input vector in the neural network;

[0018] A second phase shifter module for acting on the output waveguide of the second multi-mode interferometer module to adjust the transmission matrix;

[0019] A third phase shifter module for adjusting the phase of the reference optical signal.

[0020] Optionally, the first multi-mode interferometer module is an N+1 path beam splitter composed of a 1×N+1 multi-mode interferometer or a plurality of 1×2 multi-mode interferometers.

[0021] Optionally, the first phase shifter module, the second phase shifter module and the third phase shifter module include:

[0022] A thermo-optic phase shifter;

[0023] An electro-optic phase shifter.

[0024] Optionally, the second multi-mode interferometer module includes:

[0025] A via hole for connecting the multi-mode waveguide region and the electrode;

[0026] the electrode, for applying an electrical signal to the multimode interference region through the via hole, changing the carrier concentration near the via hole, thereby changing the refractive index of the multimode interference region near the via hole, thereby changing the transmission matrix of the second multimode interferometer module;

[0027] a transition region from a single-mode waveguide to a multimode waveguide, for reducing the loss of the to-be-processed signal light signal transmission.

[0028] Optionally, the coupler module comprises:

[0029] a cross waveguide, for transmitting the two crossed processing light signals and the reference light signal along respective propagation directions;

[0030] a coupler first sub-module, the coupler first sub-module being a 1x2 directional coupler, for coupling part of the energy of the reference light signal as a light signal for coherent detection of each of the processing light signals;

[0031] a coupler second sub-module, for coupling the reference light signal and the processing light signal into the same waveguide.

[0032] Optionally, the coupler second sub-module is a 2x2 directional coupler or is composed of a 2x2 multimode interferometer.

[0033] The technical solutions provided by the embodiments of the present application bring at least the following beneficial effects:

[0034] By introducing a process error-insensitive multimode interferometer and introducing a plurality of electrodes in the multimode interference region of the multimode interferometer, matrix operation is completed with smaller size, a programmable optical neural network is obtained, a more compact structure is realized, and a neural network in a complex domain is realized in combination with coherent detection, thereby enhancing the fitting capability of the network.

[0035] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0036] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, taken in conjunction with the following drawings, in which:

[0037] Figure 1 is a block diagram of a neural network accelerator based on a multimode interferometer and coherent detection according to an exemplary embodiment of the present application;

[0038] Figure 2 is a structural architecture of a neural network accelerator based on a multimode interferometer and coherent detection according to an exemplary embodiment of the present application;

[0039] Figure 3 is a structure architecture of a neural network expansion accelerator based on a multimode interferometer and coherent detection according to an example embodiment of the present application;

[0040] Figure 4 is a structure diagram of a second multimode interferometer module in a neural network accelerator based on a multimode interferometer and coherent detection according to an example embodiment of the present application;

[0041] Figure 5 is a structure diagram of each waveguide in a neural network accelerator based on a multimode interferometer and coherent detection according to an example embodiment of the present application. DETAILED DESCRIPTION

[0042] Embodiments of the present application are described in detail below with reference to the accompanying drawings, in which examples of embodiments are shown, wherein the same or similar reference numbers represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0043] Figure 1 is a block diagram of a neural network accelerator 100 based on a multimode interferometer and coherent detection according to an example embodiment of the present application, including a semiconductor laser 110, a multimode interferometer module 120, a coupler module 130, and a balanced detector module 140, wherein,

[0044] The semiconductor laser 110, the multimode interferometer module 120, the coupler module 130, and the balanced detector module 140 are connected in sequence;

[0045] The semiconductor laser module 110 provides an optical signal for the multimode interferometer module 120;

[0046] The multimode interferometer module 120 divides the optical signal into multiple paths and adjusts the phase of the optical signal, and outputs a processed optical signal and a reference optical signal to the coupler module 130;

[0047] The coupler module 130 is used to realize coherent detection of the processed optical signal and the reference optical signal, and couple the processed optical signal and the reference optical signal into the same waveguide;

[0048] The balanced detector module 140 beats the processed optical signal and the reference optical signal transmitted in the same waveguide, and converts the beat optical signal into an electrical signal as an input signal of the next layer of the artificial neural network.

[0049] The following is based on Figure 2 The structure of each module is described in detail.

[0050] As Figure 2 shown, the LD is a semiconductor laser 110, which provides a light source for the structure of the application. Among them, the semiconductor laser includes a distributed feedback semiconductor laser and a distributed Bragg semiconductor laser, and other models of lasers are selected according to actual needs.

[0051] In a possible embodiment, the semiconductor laser 110 is a distributed feedback semiconductor laser.

[0052] Among them, the multimode interferometer module 120 further includes:

[0053] The first multimode interferometer module divides the optical signal into N+1 paths, determines N paths of the to-be-processed optical signal and the reference optical signal of the N+1 path;

[0054] The second multimode interferometer module is an N-input and N-output multimode interferometer, which is used for performing matrix operation and outputting the processed optical signal.

[0055] The first phase shifter module is used for acting on the input waveguide of the second multimode interferometer module and adjusting the phase of the to-be-processed signal, wherein the electrical signal of the first phase shifter module is the value of the input vector in the neural network.

[0056] The second phase shifter module is used for acting on the output waveguide of the second multimode interferometer module and adjusting the transmission matrix.

[0057] The third phase shifter module is used for adjusting the phase of the reference optical signal.

[0058] In a possible embodiment, as Figure 2 shown, the MMI1 is the first multimode interferometer module, the MMI2 is the second multimode interferometer module, the PS1-PS4 are the first phase shifter modules, the PS5-PS8 are the second phase shifter modules, and the PS9 is the third phase shifter module. The following is a specific description and analysis.

[0059] As shown in the Figure 2 embodiment of the application, when N is 4, the MMI1 divides the optical signal output by the semiconductor laser into 5 paths, wherein 4 paths of the optical signal are input to the MMI2 through the PS1, the PS2, the PS3 and the PS4 respectively, and then transmitted to the coupler module 130 through the PS5, the PS6, the PS7 and the PS8 respectively. The fifth path of the optical signal is directly transmitted to the coupler module 130 through the PS9.

[0060] In the embodiment of the application, the MMI1 is an N+1 path beam splitter composed of a 1xN+1 multimode interferometer or a plurality of 1x2 multimode interferometers, and the types of the PS1-PS9 include but are not limited to thermo-optic phase shifters and electro-optic phase shifters.

[0061] In addition, the second multi-mode interferometer module MMI2 is described in detail.

[0062] The specific structure of the MMI2 is shown in Fig. 3, where MMR is a multi-mode interference region, Taper1-Taper8 are transition regions from single-mode waveguide to multi-mode waveguide for reducing the loss of optical signal transmission, Met is an electrode composed of metal for connecting electrical signal, and Via is a via hole connecting the multi-mode waveguide region and the electrode composed of metal. Figure 4

[0063] In the embodiment of the present application, based on the self-imaging principle, the optical signal entering the multi-mode interference region will periodically reappear, and the transmittance of the optical signal at the output port can be changed by adjusting the length and width of the multi-mode interference region. Since the self-imaging principle depends on the refractive index of the material, the refractive index of the multi-mode interference region can be changed by the free carrier dispersion effect. The electrical signal applied to the electrode acts on the multi-mode interference region through the via hole, changes the carrier concentration near the via hole, and thus changes the refractive index of the multi-mode interference region near the via hole. Therefore, the transmission matrix of the entire multi-mode interferometer is changed.

[0064] It should be noted that the number of MMI2 in the second multi-mode interferometer module is not fixed, and the number of PS5-PS8 in the second phase shifter module is also not fixed, Figure 2 but the most basic structure proposed in the embodiment of the present application is expanded according to the actual situation in the application, and the same structure as MMI2 and PS5-PS8 is added to change the structure of the accelerator, which also belongs to the protection scope of the present application.

[0065] In a possible embodiment, the accelerator proposed in the present application is expanded, and the structure of the expanded accelerator is shown in Fig. 4. Figure 2 Figure 3

[0066] As shown in Fig. 5, the coupler module 130 includes cross waveguides Cross1-Cross4, coupler first sub-modules DC1-DC4, and coupler second sub-modules Coupler1-Coupler4. Figure 2 Among them, Cross1-Cross4 makes two crossed optical signals transmit along their respective propagation directions, that is, the processed optical signal and the reference optical signal are transmitted to DC1-DC4 along their respective propagation directions.

[0067]

[0068] DC1-DC4 are all 1x2 directional couplers, and the energy of the reference optical signal is coupled as the optical signal for coherent detection of each processed optical signal.​​​​

[0069] Coupler1-Coupler4 couple the reference light signal with the processed light signal that has passed through MMI2 and phase shifter into the same waveguide.

[0070] In one possible embodiment, as shown in FIG. 1, Wg1-wg5 are waveguides for light signals, Cross1 makes the light signal in wg1 transmit to wg3, the light signal in wg2 transmit to wg4, DC1 couples part of the light signal in wg4 to wg5 Figure 5

[0071] In addition, Coupler1-Coupler4 are 2x2 directional couplers, or are composed of 2x2 multimode interferometers.

[0072] As shown in FIG. 1, the balanced photodetector module 140 includes balanced photodetectors BPD1-BPD4. Figure 2

[0073] In addition, as shown in FIG. 1, in the embodiment of the present application, the black straight lines and the black curved parts connected by each module are single-mode waveguide structures. Figure 2

[0074] The embodiment of the present application realizes a more compact structure by introducing a process error-insensitive multimode interferometer and a plurality of electrodes in the multimode interference region of the multimode interferometer to complete matrix operation with a smaller size, and realizes a neural network in a complex domain by combining coherent detection, thereby enhancing the fitting capability of the network.

[0075] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, the steps described in the present disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions of the present disclosure can be achieved, which are not limited herein.

[0076] The above detailed description does not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.​​​

Claims

1. A neural network accelerator based on a multi-mode interferometer and coherent detection, characterized in that, The semiconductor laser, the multimode interferometer module, the coupler module and the balanced detector module are sequentially connected. The semiconductor laser provides an optical signal for the multimode interferometer module. The multimode interferometer module divides the optical signal into multiple paths, adjusts the phase of the optical signal, and outputs a processed optical signal and a reference optical signal to the coupler module. The coupler module is used to realize coherent detection of the processed optical signal and the reference optical signal, and couple the processed optical signal and the reference optical signal into the same waveguide. The balanced detector module beats the processed optical signal and the reference optical signal transmitted in the same waveguide, and converts the beat optical signal into an electrical signal as an input signal of the next layer of artificial neural network. The multimode interferometer module comprises: A first multimode interferometer module divides the optical signal into N+1 paths, determines N processed optical signals and an N+1 reference optical signal. A second multimode interferometer module is an N-input N-output multimode interferometer, used to perform matrix operation and output a processed optical signal. A first phase shifter module is used to act on the input waveguide of the second multimode interferometer module to adjust the phase of the processed optical signal, and the electrical signal of the first phase shifter module is the value of the input vector in the neural network. A second phase shifter module is used to act on the output waveguide of the second multimode interferometer module to adjust the transmission matrix. A third phase shifter module is used to adjust the phase of the reference optical signal. The second multimode interferometer module comprises: A through hole is used to connect the multimode waveguide region and the electrode. The electrode is used to apply an electrical signal to the multimode interference region through the through hole to change the carrier concentration near the through hole, thereby changing the refractive index of the multimode interference region near the through hole, and thereby changing the transmission matrix of the second multimode interferometer module. A single-mode waveguide to multimode waveguide transition region is used to reduce the loss of the processed optical signal transmission. The semiconductor laser comprises:

2. The neural network accelerator of claim 1, wherein, A distributed feedback semiconductor laser; A distributed Bragg semiconductor laser. The first multimode interferometer module is an N+1 path beam splitter composed of a 1×N+1 multimode interferometer or multiple 1×2 multimode interferometers.

3. The neural network accelerator of claim 1, wherein, The first phase shifter module, the second phase shifter module and the third phase shifter module comprise:

4. The neural network accelerator of claim 1, wherein, A thermo-optic phase shifter; An electro-optic phase shifter. The coupler module comprises:

5. The neural network accelerator of claim 1, wherein, A cross waveguide makes the processed optical signal and the reference optical signal cross and transmit along their respective propagation directions; A coupler first sub-module is a 1×2 directional coupler, used to couple part of the energy of the reference optical signal as the optical signal for coherent detection of each processed optical signal; A coupler second sub-module couples the reference optical signal and the processed optical signal into the same waveguide. ​ 6. The neural network accelerator of claim 5, wherein, The coupler second sub-module is a 2x2 directional coupler or consists of a 2x2 multimode interferometer. The coupler second sub-module is a 2x2 directional coupler or consists of a 2x2 multimode interferometer.

Images