CN111338423A

CN111338423A - Optical device

Info

Publication number: CN111338423A
Application number: CN201811559763.4A
Authority: CN
Inventors: 涂敏海; 李腾浩
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2018-12-19
Filing date: 2018-12-19
Publication date: 2020-06-26

Abstract

The application provides an optical device for solving the problem that the light energy utilization rate is low in the prior art. The optical device includes: the system comprises a light source, N first spatial light modulation modules, N coupling modules, an optical medium, M detection modules and an operation control module; the light source emits at least one path of optical signals, wherein N output ports of the light source are connected with N first spatial light modulation modules; the N first spatial light modulation modules are connected with the N couplers; the N coupling modules couple the received modulated optical signals to N input positions in the optical medium; m output positions in the optical medium correspond to the M detection modules, and the waveguide written in the optical medium guides the incident light of the input position to reach the corresponding output positions; the M detection modules sample emergent light received from the output position to obtain sampling information; and the operation control module updates information carried by the optical signal emitted by the light source and updates the driving signals of the N first spatial light modulation modules based on the sampling information.

Description

Optical device

Technical Field

The present application relates to the field of optical technology, and more particularly, to an optical device.

Background

In recent years, artificial intelligence and machine learning have been promoted. Artificial Neural Network (ANN) technology has emerged as an important computing application in many fields such as face recognition, automatic driving, consumer terminals, and data centers, and has formed a great commercial value and social effect. At present, the implementation of ANN is based on traditional electronic computing elements, such as CPU, GPU, FPGA and dedicated acceleration chip. However, entering the post-molar (Moore) era, these traditional computing modules have a common power performance bottleneck. To address this difficulty, other computational methods, such as quantum computation, optical computation, DNA computation, and the like, have gained common attention in both academic and industrial circles.

In terms of optical neural networks (including optical mimicry neural networks), the approaches currently employed are: an optical vector-matrix multiplier.

The optical vector-matrix multiplier comprises a light source emission vector array and a detector vector array which are generally vertically arranged, and each pixel in the light source emission vector array and the detector vector array corresponds to one neuron; the related matrix mask plate is positioned between the light source emission vector array and the detector vector array, the transmittance of each pixel of the related matrix mask plate corresponds to the weight information of the output connection of the neuron, and the column number corresponds to the input neuron number and the row number corresponds to the output neuron number.

Even if the weights or connections corresponding to the neurons corresponding to the pixels do not exist, the light energy of each pixel in the light source emission vector array is evenly distributed to the pixels of the relevant matrix mask plate, and therefore the light energy utilization rate is low.

Disclosure of Invention

The application provides an optical device for solving the problem that the light energy utilization rate is low in the prior art.

In a first aspect, the present application provides an optical device comprising: the device comprises a light source, N first spatial light modulation modules, N coupling modules, an optical medium, M detection modules and an operation control module, wherein N and M are positive integers not less than 1; the light source comprises N output ports and is used for transmitting at least one path of optical signals, wherein the N output ports of the light source are respectively connected with the N first spatial light modulation modules; the N first spatial light modulation modules are respectively connected with the N couplers, and the first spatial light modulator is used for modulating a received optical signal; the N coupling modules are configured to couple the received modulated optical signal to N corresponding input positions in the optical medium; m output positions in the optical medium respectively correspond to the M detection modules, and the written waveguide in the optical medium is used for guiding incident light of an input position to reach the corresponding output positions; the M detection modules are used for receiving emergent light from corresponding output positions and sampling the emergent light to obtain sampling information; and the operation control module updates the information carried by the optical signal emitted by the light source and updates the driving signals of the N first spatial light modulation modules based on the sampling information obtained from the M detection modules.

According to the scheme, the light source port corresponding to the neuron without output does not emit the optical signal, so that the light energy utilization rate can be improved. In addition, weight control can be controlled by the spatial light modulator, the waveguide only provides an interconnection channel, a complex and sensitive waveguide transmission adjusting structure is not needed, and feasibility and reliability are improved. The waveguide can be structurally reconstructed, so that any path can be structurally realized, additional phase modulation devices such as Mach-Zehnder interferometers (MZIs) and the like are not needed, precise adjustment is not needed, the optical device is high in anti-interference capacity, wide-spectrum incoherent light transmission can be supported, and the waveguide is not limited to adjustment of waveguide transmission. In addition, the neurons can be arranged on the input spatial light modulation module according to requirements, and the flexibility is strong.

In one possible design, further comprising: and the reconstruction module is used for receiving the control information sent by the operation control module, and writing the control information into the waveguide and/or erasing part or all of the waveguide in the optical medium according to the control information.

Through the design, the module for reconstructing the waveguide is configured in the device, and the method is simple and easy to implement.

In one possible design, the reconstruction module includes a laser, a collimating lens, a second spatial light modulation module, and a focusing objective lens, which are sequentially arranged on the optical path transmission path; the laser is used for emitting writing laser or erasing laser according to the control information; the collimating lens is used for collimating the writing laser or the erasing laser; the second spatial light modulation module is used for adjusting the light intensity of the collimated laser based on the control information; and the focusing objective lens is used for focusing the laser adjusted by the second spatial light modulation module to generate light spots.

The above design provides a structure of a module for reconfiguring a waveguide, which is simple and easy to implement.

In one possible design, each optical signal output by the light source carries output information of one neuron in a neural network.

The design is simple and easy to realize by modulating the output information of the neuron by the light source.

In a possible design, the first spatial light modulation module is configured to modulate a received optical signal according to a first weight of at least one output connection of the neuron, where the output connection is a connection used by another neuron to which output information of one neuron is transmitted.

In the design, weight control can be controlled by the spatial light modulator, the waveguide only provides an interconnection channel, a complex and sensitive waveguide transmission adjusting structure is not needed, and feasibility and reliability are improved.

In one possible design, the waveguides present in the optical medium include an input waveguide, a connecting waveguide, an output waveguide;

the input waveguide is used for receiving the modulated optical signal and sending the modulated optical signal out through at least one connecting waveguide; the connecting waveguides are used for sending the received optical signals to the output waveguides, wherein each connecting waveguide corresponds to one input waveguide and one output waveguide; the output waveguide is used for receiving optical signals from at least one input waveguide and sending the optical signals out;

in one possible design, the number of connection waveguides to which the input waveguide is coupled is determined by the number of output connections of the neuron corresponding to the input waveguide and the second weight of each output connection of the corresponding neuron, and the number of connection waveguides to which the input waveguide is coupled to the output waveguide is determined by the second weight of the first output connection, wherein the first output connection is one of the output connections of the neuron corresponding to the input waveguide and is used for connecting the neuron corresponding to the output waveguide.

Through the design, the weight (second weight) of the output connection of the neuron can be further characterized by the number of the connection waveguides, and the neuron has the characteristic of discretization. The structure can be applied to the application of a specific artificial neural network, such as a pulse neural network. The structure of this embodiment can reduce the number of a part of the input interfaces with respect to the waveguide structure described above.

In one possible design, the waveguides involved in the embodiments of the present application may be three-dimensional waveguides. And no intersection exists among the three-dimensional waveguides, so that the interference can be avoided.

In one possible design, the optical medium is made of a photosensitive material, or the optical medium is made of a fused silica material, or the optical medium is made of a phase change material, or the optical medium is made of a magneto-optical material.

In a possible design, if the optical medium is made of a photosensitive material or a fused silica material, part or all of the three-dimensional waveguide included in the optical medium is written or erased under the action of an optical field; alternatively, the first and second electrodes may be,

if the optical medium is made of the phase-change material, part or all of the three-dimensional waveguides in the optical medium are written or erased under the action of an optical field, an electric field or a thermal field; alternatively, the first and second electrodes may be,

if the optical medium is made of magneto-optical material, part or all of the three-dimensional waveguide included in the optical medium is written or erased under the action of an optical field or a magnetic field.

In one possible design, the light source includes a light source array composed of N tunable lasers, one tunable laser corresponding to one output port; or the light source comprises a laser, a collimating lens and N third spatial light modulation modules which are sequentially arranged on the light path transmission path; the laser is used for emitting laser to the collimating lens; the collimating lens is used for collimating the laser and inputting collimated optical signals to the N third spatial light modulation modules;

and the third spatial light modulation module is used for modulating the collimated optical signal.

The above design illustrates two simple and feasible configurations of the light source.

In one possible design, the output information of the neuron is loaded in any one of the following ways:

intensity modulation, phase modulation, polarization direction modulation, pulse modulation, wavelength modulation, mode field modulation, polarization multiplexing, pulse multiplexing, wavelength multiplexing, or mode multiplexing.

In one possible design, the first weight of the output connection of the neuron is modulated in at least one of the following ways:

intensity modulation, phase modulation, polarization direction modulation, pulse modulation, wavelength modulation, or mode field modulation.

Drawings

Fig. 1 is a schematic structural diagram of an optical device according to an embodiment of the present disclosure;

FIG. 2 is a schematic structural diagram of another optical device according to an embodiment of the present disclosure;

FIG. 3 is a schematic structural diagram of another optical device according to an embodiment of the present disclosure;

fig. 4 is a schematic diagram illustrating a correspondence relationship between a light source and a control light modulator and a coupling module according to an embodiment of the present application;

fig. 5 is a schematic structural diagram of a light source according to an embodiment of the present disclosure;

FIG. 6A is a schematic diagram of a neuron connection structure according to an embodiment of the present disclosure;

fig. 6B is a schematic view of a waveguide structure corresponding to fig. 6A according to an embodiment of the present disclosure;

FIG. 7A is a schematic diagram of a neuron connection variation structure according to an embodiment of the present application;

fig. 7B is a schematic diagram of a waveguide structure variation corresponding to fig. 7A according to an embodiment of the present application;

FIG. 8 is a schematic view of another waveguide structure provided in embodiments of the present application;

fig. 9 is a schematic view of another waveguide structure provided in the embodiment of the present application.

Detailed Description

The embodiment of the application provides an optical device, which is used for solving the problem of low light energy utilization rate in the prior art. The method and the device are based on the same inventive concept, and because the principles of solving the problems of the method and the device are similar, the implementation of the device and the method can be mutually referred, and repeated parts are not repeated. The optical device provided by the embodiment of the application can realize the function of a neural network.

The optical device provided by the embodiment of the application can be used in various computing application fields, such as image recognition and analysis, complex system prediction and control, nonlinear optimization, heuristic search acceleration and the like.

The embodiments provided in the present application will be described in detail below with reference to the accompanying drawings. In the embodiments of the present application, a plurality means two or more. In addition, it is to be understood that the terms first, second, etc. in the description of the present application are used for distinguishing between the descriptions and not necessarily for describing a sequential or chronological order.

Referring to fig. 1, a schematic structural diagram of an optical device provided in an embodiment of the present application is shown.

The optical device includes: the device comprises a light source 11, N first spatial light modulation modules 12, N coupling modules 13, an optical medium 14, M detection modules 15 and an operation control module 16, wherein N and M are positive integers not less than 1.

The light source 11 includes N output ports for emitting at least one optical signal, where the N output ports of the light source are respectively connected to the N first spatial light modulation modules 12.

Optionally, the light source 11 is configured to generate light for loading output information of each neuron in the neural network, for example, each optical signal output by the light source carries output information of one neuron in the neural network. The output information of the neuron may be transmitted to the light source by the arithmetic control module 16. The neural network in the embodiment of the present application may include one or more layers of networks, and the involved neurons may be neurons in the same layer or neurons in different layers.

Illustratively, the loading of the output information of the neuron may be in any one of the following ways: intensity modulation, phase modulation, polarization direction modulation, pulse modulation, wavelength modulation, mode field modulation, polarization multiplexing, pulse multiplexing, wavelength multiplexing, or mode multiplexing. In other words, in the optical device shown in fig. 1, the light source 11 can generate at least one optical signal loaded with the output information of the neuron by any one of the loading methods.

Furthermore, in the present embodiment, the number of optical signals emitted by the light source is related to the number of output connections of the neurons comprised in the neural network.

In an alternative example, in the optical apparatus shown in fig. 1, the number of optical signals emitted by the light source 11 may be equal to the number of output connections of neurons included in the neural network, and one optical signal emitted by the light source 11 corresponds to one output connection of one neuron. Here, an output connection refers to a connection used by one neuron to which output information of another neuron is transmitted. If there are multiple output connections for a neuron, the different output connections of a neuron have the same modulation information on the optical signal emitted on the light source 11.

In another alternative example, the number of optical signals emitted by the light source 11 may be equal to the number of neurons for which output connections are present. When a plurality of output connections exist in one neuron, only one output information of the neuron is provided, so that the utilization rate of the optical resource can be improved.

In the optical device shown in fig. 1, the N first spatial light modulation modules 12 are respectively connected to the N coupling modules 13, and the first spatial light modulation modules 12 are configured to modulate a received optical signal.

Optionally, the first spatial light modulation module 12 is configured to modulate the received optical signal according to a first weight of at least one output connection of the neuron, where the output connection is a connection used by another neuron to which the output information of one neuron is transmitted.

There are various ways for implementing the first spatial light modulation module 12 to modulate the weight of the neuron output connection on the received optical signal, two of which are listed below.

A first possible way is that the first spatial light modulation module 12 modulates a first weight of an output connection of a neuron on a path of optical signal. A first possible way is to adapt to the way in which the number of optical signals emitted by the light source 11 is equal to the number of output connections of the neurons comprised in the neural network.

A second possible way is that the first spatial light modulation module 12 modulates the first weights of the plurality of output connections where one neuron exists on one optical signal. A second possibility applies to the way in which the number of optical signals emitted by the light source 11 is equal to the number of neurons comprised in the neural network in which there are output connections.

Illustratively, the modulation mode adopted by the first weight value of the output connection of the neuron is at least one of the following modes: intensity modulation, phase modulation, polarization direction modulation, pulse modulation, wavelength modulation, or mode field modulation.

In an alternative example, the N first spatial light modulation modules 12 may be implemented by one or more spatial light modulators. A spatial light modulator comprises a plurality of pixels, one pixel corresponding to each spatial light modulation module 12, i.e. one pixel corresponding to the output connections of one neuron. For example, the spatial light modulator may be of an amplitude type, and modulates digital information corresponding to the first weight value connected to the neuron output into the transmittance of the corresponding pixel.

The N coupling modules 13 are configured to couple the received modulated optical signal to corresponding N input positions in the optical medium 14.

Illustratively, the coupling module 13 may be a microlens, and an optical axis of the microlens overlaps an optical axis of the first spatial light modulation module 12. Other optical configurations for achieving coupling are of course possible.

In the optical device shown in fig. 1, M output positions in the optical medium 14 correspond to the M detection modules 15, respectively. In which certain interconnected waveguide structures may be written. A written waveguide in an optical medium may be used to guide incident light at an input location to a corresponding output location.

Illustratively, the waveguide written in the optical medium 14 in the embodiment of the present application may adopt a three-dimensional waveguide structure.

The M detection modules 15 are used for receiving emergent light from corresponding output positions, sampling the emergent light and obtaining sampling information. The detection module 15 mainly performs spatial sampling on the received emergent light, for example, converts light field distribution (e.g., light intensity, phase, etc.) into sampling information, where the sampling information may be digital information and may be input information of neurons. Illustratively, the M detection modules 15 may be implemented by one detector, such as a Complementary Metal Oxide Semiconductor (CMOS) area array detector, or by M detectors.

The operation control module 16 updates information carried by the optical signal emitted by the light source 11 and updates the driving signals of the N first spatial light modulation modules 12 based on the sampling information obtained from the M detection modules 15.

Illustratively, the information carried by the optical signal emitted by the light source 11 may be output information of a neuron. The drive signal may be a first weight of at least one output connection of the respective neuron.

The operation control module 16 may perform fusion processing on the sampling information to obtain output information of the neuron transmitted to the light source 11 and a first weight of at least one output connection of the neuron transmitted to the first spatial light modulation module 12.

Optionally, the arithmetic control module 16 may also have a function of implementing a digital neural network algorithm. In addition, the algorithm and the result of the arithmetic control module 16 for performing fusion processing on the sampling information can be cascaded or nested with other neural networks.

The calculation control module 16 may be implemented by one or more processing modules, one or more chips, or one or more processors. For example, a Central Processing Unit (CPU), an Application Specific Integrated Circuit (ASIC), a field-programmable gate array (FPGA), or a Complex Programmable Logic Device (CPLD).

In a possible example, the optical apparatus shown in fig. 1 may further include a storage module, where the storage module may be configured in the operation control module 16, or may be located outside the operation control module 16, and is configured to store program codes run by the operation control module 16, perform fusion processing on the sampling information to obtain output information of the neuron transmitted to the light source 11 and a first weight value of at least one output connection of the neuron transmitted to the first spatial light modulation module 12. But also for storing digital neural network algorithms, etc.

In the embodiment of the application, the weight control is controlled by the spatial light modulator, the waveguide only provides an interconnection channel, a complex and sensitive waveguide transmission adjusting structure is not needed, and the feasibility and the reliability are improved. The waveguide can be structurally reconstructed, so that any path can be structurally realized, additional phase modulation devices such as Mach-Zehnder interferometers (MZIs) and the like are not needed, precise adjustment is not needed, the optical device is high in anti-interference capacity, wide-spectrum incoherent light transmission can be supported, and the waveguide is not limited to adjustment of waveguide transmission. In addition, the neurons can be arranged on the input spatial light modulation module according to requirements. Therefore, the embodiment of the application has strong flexibility.

In the above embodiment, a hybrid calculation architecture of a digital neural network and an optical neural network may be further adopted, so that the advantages of digital calculation and optical calculation may be utilized to the maximum extent, and the optimal balance between calculation efficiency and energy consumption may be realized.

In the embodiments of the present application, the reconstruction of the waveguide in the optical medium may be realized by a structure other than the optical device, or a structure for realizing the reconstruction of the waveguide may be arranged in the optical device. Illustratively, referring to fig. 2, the optical device may further include a reconstruction module 17.

The reconstruction module 17 may be configured to receive the control information sent by the operation control module 16, and write a waveguide on the optical medium 14 and/or erase part or all of the waveguides in the optical medium 14 according to the control information.

The reconstruction module 17 can move under the control of the calculation control module 16, such as translating in three directions under a rectangular coordinate system. The reconstruction module 17 may write or erase the waveguide in the optical medium 14 point by point during movement. The reconstruction module 17 may also be controlled by an external device to write or erase the waveguide in the optical medium 14 point by point.

The reconstruction module 17 according to the embodiment of the present disclosure may be disposed inside the optical device, or may be disposed outside the optical device.

Illustratively, referring to fig. 3, the reconstruction module 17 includes a laser 171, a collimating lens 172, a second spatial light modulation module 173, and a focusing objective 174, which are arranged in sequence on the optical path transmission path.

Wherein, the laser 171 is used for emitting writing laser or erasing laser according to the control information. The laser 171 may be switched between the writing laser and the erasing laser according to control information.

Illustratively, in the embodiment of the present application, the writing laser and the erasing laser may be generated by different wavelengths or repetition frequencies.

The collimating lens 172 is configured to collimate the writing laser or the erasing laser.

The second spatial light modulation module 173 is configured to adjust the light intensity of the collimated laser light based on the control information.

The focusing objective 174 is configured to focus the laser light modulated by the second spatial light modulation module 173 to generate a light spot.

Illustratively, the function of the second spatial light modulation module 173 may be implemented by a spatial light modulator, and for convenience of distinction, the spatial light modulator implementing the function of the second spatial light modulation module 173 is referred to as a second spatial light modulator.

In one possible embodiment, the optical medium 14 may be made of a rewritable optically sensitive material.

In one example, optical media 14 is made of a photosensitive material.

If the optical medium 14 is made of a photosensitive material, a part or all of the three-dimensional waveguides included in the optical medium 14 may be written or erased under the action of an optical field. For example, the reconstruction module 17 may write or erase a part or all of the three-dimensional waveguide included in the optical medium 14 by an optical field effect.

In another example, the optical medium is made of fused silica material.

If the optical medium 14 is made of fused silica material, part or all of the three-dimensional waveguides included in the optical medium 14 may be written or erased under the action of the optical field. For example, the reconstruction module 17 may write or erase a part or all of the three-dimensional waveguide included in the optical medium 14 by an optical field effect.

In yet another example, the optical medium 14 is made of a phase change material.

If the optical medium 14 is made of a phase-change material, a part or all of the three-dimensional waveguides included in the optical medium 14 are written or erased under the action of an optical field, an electric field, or a thermal field. For example, the reconstruction module 17 may write or erase part or all of the three-dimensional waveguide included in the optical medium 14 by an optical field, an electric field, or a thermal field.

In yet another example, the optical medium is made of magneto-optical material.

If the optical medium is made of magneto-optical material, part or all of the three-dimensional waveguide included in the optical medium is written or erased under the action of an optical field or a magnetic field. For example, the reconstruction module 17 may write or erase part or all of the three-dimensional waveguide included in the optical medium 14 by an optical field or a magnetic field.

In a possible implementation manner, the light source 11 is a light source array composed of N tunable lasers, and one tunable laser corresponds to one output port. The light source array may be implemented by a coherent light source, such as a Vertical Cavity Surface Emitting Laser (VCSEL), or an incoherent light source, such as an organic light-emitting diode (OLED). The N adjustable lasers correspond to the N first spatial light modulation modules one by one.

For example, referring to fig. 4, the light source 11 included in the optical device is a light source array composed of m rows and N columns of tunable lasers, where m × N is N. In fig. 4, N first spatial light modulation modules 12 are implemented by one spatial light modulator, and for the sake of convenience of distinction, the spatial light modulator that implements the functions of the N first spatial light modulation modules 12 is referred to as a first spatial light modulator 12A. The first spatial light modulator 12A includes a two-dimensional pixel array of m rows and n columns, and the tunable lasers of m rows and n columns are in one-to-one correspondence with and aligned with the two-dimensional pixel array of m rows and n columns. Wherein a pixel may correspond to one or more output connections of a neuron. In fig. 4, taking the N coupling modules 13 as an example of the m rows and N columns of microlens arrays 13A, the m rows and N columns of microlens arrays 13A are in one-to-one correspondence with and aligned with the m rows and N columns of two-dimensional pixel arrays. Illustratively, the diameter size of the microlens may be the same as the pixel size. The microlenses 13A can focus the phase-shifted plane waves output by each pixel into converging plane waves that are coupled into a corresponding input position of the optical medium.

In another possible implementation manner, referring to fig. 5, the light source 11 includes a laser 111, a collimating lens 112, and N third spatial light modulators 113, which are sequentially arranged on the optical path transmission path.

The laser 111 is configured to emit laser light to the collimating lens 112. The laser light of the laser 111 may be output to the collimator lens 112 through a single-mode optical fiber. For example, the laser 111 may be a pigtailed semiconductor laser with an operating wavelength of 532 nm.

The collimating lens 112 is configured to collimate the laser light, and input a collimated optical signal to the N third spatial light modulation modules 113. The laser 111 emits a spherical wave (i.e., spherical light) and is then collimated into a plane wave (i.e., parallel light) by the collimator lens 112.

The third spatial light modulation module 113 is configured to modulate the collimated optical signal, that is, the third spatial light modulation module 113 modulates the collimated optical signal according to the output information of the neuron. The functions of the N third spatial light modulation modules 113 may be implemented by one or more spatial light modulators.

The waveguides present in optical medium 14 in the embodiments of the present application may take a variety of configurations.

In one example, the waveguides present in the optical medium 14 include an input waveguide and an output waveguide. One input waveguide corresponds to one output connection of one neuron and one output waveguide corresponds to one neuron. The optical signal modulated by the first spatial light modulation module 12 is coupled into an input waveguide through the coupling module 13. The input waveguide is configured to receive the modulated optical signal coupled by the coupling module 13, and send the modulated optical signal out through an output waveguide; and the output waveguide is used for receiving the optical signal from at least one input waveguide and sending the optical signal out. The number of input waveguides is equal to the number of output connections of the neurons comprised in the neural network and the number of output waveguides is equal to the number of neurons comprised in the neural network.

For example, a neural network includes two neurons,

neurons

601a and 601b, respectively, and as shown in fig. 6A, there are three output connections of neurons between

neurons

601a and 601b, output connection 602a from neuron 601a to neuron 601a, output connection 602b from neuron 601b to neuron 601b, and output connection 602c from neuron 601b to neuron 601 a.

The neural network of fig. 6A may be implemented by the structure shown in fig. 6B. Referring to fig. 6B, the waveguides written in the optical medium 34 include

input waveguides

603a, 603B, 603c and output waveguides 604a, 604B. Input waveguide 603a corresponds to output connection 602a, input waveguide 603b corresponds to output connection 602c, input waveguide 603c corresponds to output connection 602b, output waveguide 604a corresponds to neuron 601a, and output waveguide 604b corresponds to neuron 601 b. As can be seen in fig. 6A, there are two output connections,

output connections

602a and 602c respectively, to the neuron 601a and one output connection, output connection 601b to the neuron 601b, based on which it is determined that the input waveguide 603a and the input waveguide 603b are coupled to the output waveguide 604a and the input waveguide 603c is coupled to the output waveguide 604b, respectively. In addition, waveguides in the optical medium do not cross each other except for the waveguide where coupling is present.

In addition, the waveguides present in the optical medium 14 may be partially written and erased, i.e., only the waveguides that change are written and erased. The connection mode of the neural network changes, and as shown in fig. 7A, the connection mode of the neural network changes from state 1 to state 2. The changed neural network comprises two

neurons

701a and 701b, respectively, output connections of the two neurons exist between the

neurons

702a and 702b, the output connection 702a is output from the neuron 702a to reach the neuron 701b, and the output connection 702b is output from the neuron 701b to reach the neuron 701 a.

Referring to fig. 7B, the waveguides to which optical medium 34 is written in state 2 include input waveguides 704a, 704B and output waveguides 703a, 703B. Input waveguide 704a corresponds to output connection 702a, input waveguide 704b corresponds to output connection 702b, output waveguide 703a corresponds to neuron 701b, and output waveguide 703b corresponds to neuron 701 a.

The neural network changes from state 1 to state 2, i.e. the number of output connections is reduced from 3 to 2, and only one input waveguide needs to be erased on the basis of 1 in the original state, as shown in fig. 7B.

In addition, the number of light sources participating in the operation of the light source 11 is correspondingly reduced from 3 to 2, so that the input light energy is saved.

In the embodiment of the application, the optical neural network is realized by forming the interconnection through the waveguide in the optical medium. The waveguide may provide a physical connection for each neuron, providing a transmission channel for the corresponding optical signal. Further, by using a laser-erasable photosensitive material, the waveguides in the optical medium can be erased and rearranged, thereby realizing a reconfigurable optical neural network.

In the above example, the corresponding waveguides are formed only for the existing interconnects, reducing the number of waveguides, thereby improving the energy utilization of light and reducing crosstalk between interconnects. Taking the neuron structure in fig. 6A as an example, in the prior art, physical connections of 4 neurons need to be provided; in the present embodiment, only a physical connection of 3 neurons needs to be provided. Especially for the situation of large-scale neurons and sparse local interconnection, the embodiment of the application has remarkable improvement on the light energy utilization rate compared with the prior art. For example, for the case of 1000 neurons and 5000 neurons connected, the light energy utilization rate of the embodiment of the present application can be improved by 200 times compared with the prior art.

In another example, the waveguides present in the optical medium 14 may be divided into input waveguides, connecting waveguides, and output waveguides.

And the input waveguide is used for receiving the modulated optical signal and sending the modulated optical signal out through at least one connecting waveguide. The connecting waveguides are used for sending the received optical signals to the output waveguides, wherein each connecting waveguide corresponds to one input waveguide and one output waveguide. And the output waveguide is used for receiving the optical signal from at least one input waveguide and sending the optical signal out.

The number of the connecting waveguides coupled with the input waveguide is determined by the number of the output connections of the neuron corresponding to the input waveguide and the second weight of each output connection of the corresponding neuron, and the number of the connecting waveguides coupled between the input waveguide and the output waveguide is determined by the second weight of the first output connection, wherein the first output connection is one of the output connections of the neuron corresponding to the input waveguide and is used for connecting the neuron corresponding to the first output waveguide.

For example, referring to FIG. 8, another implementation of a waveguide structure written in optical medium 14 is shown. The different waveguides can be divided into three types, i.e., input waveguides, connecting waveguides, and output waveguides, according to their functions. One input waveguide corresponds to an output connection where one neuron exists, and one output waveguide corresponds to one neuron. In fig. 8, one input waveguide corresponds to one input position in the optical medium 14, and the input waveguides include 3 input waveguides, for example, 801a, 801b, and 801c, which correspond to the coupling modules 13 one to one and implement coupling; one output waveguide corresponds to one output position, for example, including 3 output waveguides, 802a, 802b, and 802, and corresponds to the detector modules 15 one to one, and implements coupling. The connecting waveguides corresponding to the input waveguides constitute 1 to many beam splitters and can evenly distribute the light energy.

In fig. 8, an input waveguide 801a is coupled to an output waveguide 802a by 3 connecting waveguides, 801c by 1 connecting waveguide; the input waveguide 801b is coupled to the output waveguide 802b by 2 connecting waveguides, 801a by one connecting waveguide; the input waveguide 801b is coupled to the output waveguide 802c by 1 connecting waveguide. For example, for the input waveguide 801a, the coupling is to 4 connecting waveguides, 1 of the 4 connecting waveguides is coupled to the output waveguide 802b, and 3 of the 4 connecting waveguides are coupled to the output waveguide 802a, that is, the optical energy output by the input waveguide 801a is distributed according to a ratio of 1:3, the second weight corresponding to the output connection of the neuron, that is, the second weight of the output connection of the neuron corresponding to the input waveguide 801a and the output waveguide 802b is 0.25, and the second weight of the output connection corresponding to the input waveguide 801a and the output waveguide 802a is 0.75. Wherein the waveguides are in three-dimensional space, do not cross each other, and are kept at proper distance from each other to reduce crosstalk.

In this embodiment, the weight (second weight) of the output connection of the neuron can be further represented by the number of the connection waveguides, and has a discretization characteristic. The structure can be applied to the application of a specific artificial neural network, such as a pulse neural network. The structure of this embodiment can reduce the number of a part of the input interfaces with respect to the waveguide structure described above.

In yet another example, the waveguides present in the optical medium 14 may be divided into input waveguides, a network of connecting waveguides, and output waveguides. One input waveguide corresponds to an output connection of a neuron, one output waveguide corresponds to a neuron, and the input waveguide is used for receiving the modulated optical signal and sending the modulated optical signal out through at least one connecting waveguide. The connecting waveguide network is used for sending the received optical signal to the output waveguide. The output waveguide is used for receiving the optical signal from the input waveguide and sending the optical signal out. The connecting waveguide network may be divided into a plurality of layers, and the number of waveguides between layers, the number of split beams of the beam splitter, the number of combined beams of the beam combiner, and the waveguide length may be predetermined in advance. Wherein the waveguides are in three-dimensional space, do not cross each other, and are kept at proper distance from each other to reduce crosstalk. In one aspect, the second weights corresponding to each of the connected waveguides are determined as written in optical medium 14. The first weight may thus be determined based on the second weight. On the other hand, the weights of the first spatial light modulation module may also be random. When the input light source 11 is a coherent light source and the output information of the neurons is encoded according to the intensity, the intensity and the phase of the pulse in each waveguide are random, and a corresponding neural network can be finally generated through a plurality of random division-superposition processes. In addition, other structural interconnections, such as micro-optical structures like scattering particles, micro-gratings, etc., may be used to achieve random path light transmission.

For example, referring to FIG. 9, there is shown yet another implementation of a waveguide structure written in optical medium 14. The different waveguides can be divided into three types, i.e., input waveguides, connection waveguide networks, and output waveguides according to their functions. In fig. 9, one input waveguide corresponds to one input position in the optical medium 14, and the input waveguides include 3, for example, 901a, 901b, and 901c, which correspond to the coupling modules 13 one to one and implement coupling; one output waveguide corresponds to one output position, for example, 3 output waveguides are included, 902a, 902b, and 902, and correspond to the detector modules 15 one to one, and implement coupling. The connecting waveguides corresponding to the input waveguides constitute 1 to many beam splitters and can evenly distribute the light energy. The waveguides in the connecting waveguide network are divided into a plurality of layers, the waveguides among the layers are randomly divided and gathered, and the lengths of the waveguides are configured into a random state, so that the random modulation of pulse phases is realized.

In this embodiment, on the one hand, the weight of the output connection of the neuron has the characteristics of random number and complex number. The input optical field can be randomly transformed with respect to the first two waveguide structure implementations. The structure is suitable for application of a specific artificial neural network, such as an echo state neural network. On the other hand, the layered structure reduces the number of local waveguides and simultaneously ensures the randomness of waveguide connection in the connected waveguide network.

As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

It will be apparent to those skilled in the art that various changes and modifications may be made in the present application without departing from the scope of the application. Thus, if such modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is intended to include such modifications and variations as well.

Claims

1. An optical device, comprising: the device comprises a light source, N first spatial light modulation modules, N coupling modules, an optical medium, M detection modules and an operation control module, wherein N and M are positive integers not less than 1;

the light source comprises N output ports and is used for transmitting at least one path of optical signals, wherein the N output ports of the light source are respectively connected with the N first spatial light modulation modules;

the N first spatial light modulation modules are respectively connected with the N couplers, and the first spatial light modulator is used for modulating a received optical signal;

the N coupling modules are configured to couple the received modulated optical signal to N corresponding input positions in the optical medium;

m output positions in the optical medium respectively correspond to the M detection modules, and the written waveguide in the optical medium is used for guiding incident light of an input position to reach the corresponding output positions;

the M detection modules are used for receiving emergent light from corresponding output positions and sampling the emergent light to obtain sampling information;

and the operation control module updates the information carried by the optical signal emitted by the light source and updates the driving signals of the N first spatial light modulation modules based on the sampling information obtained from the M detection modules.

2. The apparatus of claim 1, further comprising:

and the reconstruction module is used for receiving the control information sent by the operation control module, and writing the control information into the waveguide and/or erasing part or all of the waveguide in the optical medium according to the control information.

3. The apparatus of claim 2, wherein the reconstruction module comprises a laser, a collimating lens, a second spatial light modulation module, and a focusing objective lens arranged in sequence on the optical path transmission path;

the laser is used for emitting writing laser or erasing laser according to the control information;

the collimating lens is used for collimating the writing laser or the erasing laser;

the second spatial light modulation module is used for adjusting the light intensity of the collimated laser based on the control information;

and the focusing objective lens is used for focusing the laser adjusted by the second spatial light modulation module to generate light spots.

4. The apparatus of any one of claims 1-3, wherein each optical signal output by the optical source carries information about an output of a neuron in a neural network.

5. The apparatus of claim 4, wherein the first spatial light modulation module is configured to modulate a received optical signal according to a first weight of at least one output connection of the neuron, the output connection being a connection used by another neuron to which output information of one neuron is transmitted.

6. The apparatus of claim 4, wherein the waveguides present in the optical medium include an input waveguide, a connecting waveguide, an output waveguide;

the input waveguide is used for receiving the modulated optical signal and sending the modulated optical signal out through at least one connecting waveguide;

the connecting waveguides are used for sending the received optical signals to the output waveguides, wherein each connecting waveguide corresponds to one input waveguide and one output waveguide;

and the output waveguide is used for receiving the optical signal from at least one input waveguide and sending the optical signal out.

7. The apparatus of claim 6, wherein the number of connection waveguides to which the input waveguide is coupled is determined by a number of output connections of a neuron corresponding to the input waveguide and a second weight for each output connection of the corresponding neuron, the number of connection waveguides coupled between the input waveguide and the output waveguide being determined by the second weight for a first output connection, wherein the first output connection is one of the output connections of the neuron corresponding to the input waveguide and is used to connect the neuron corresponding to the output waveguide.

8. The apparatus of any of claims 1-3, wherein the waveguide is a three-dimensional waveguide.

9. The apparatus of claim 8, wherein the optical medium is made of a photosensitive material, or the optical medium is made of a fused silica material, or the optical medium is made of a phase change material, or the optical medium is made of a magneto-optical material.

10. The apparatus of claim 9, wherein if the optical medium is made of a photosensitive material or a fused silica material, part or all of the three-dimensional waveguide included in the optical medium is written or erased under the action of an optical field; alternatively, the first and second electrodes may be,

11. The apparatus of any of claims 1-3, wherein:

the light source comprises a light source array formed by N adjustable lasers, and one adjustable laser corresponds to one output port;

alternatively, the first and second electrodes may be,

the light source comprises a laser, a collimating lens and N third spatial light modulation modules which are sequentially arranged on a light path transmission path;

the laser is used for emitting laser to the collimating lens;

the collimating lens is used for collimating the laser and inputting collimated optical signals to the N third spatial light modulation modules;

12. The apparatus of claim 4, wherein the output information of the neuron is loaded in any one of the following ways:

13. The apparatus of claim 5, wherein the first weights of the output connections of the neurons are modulated in at least one of: