WO2021227767A1

WO2021227767A1 - All-optical diffractive neural network and system implemented on optical waveguide and/or optical chip

Info

Publication number: WO2021227767A1
Application number: PCT/CN2021/087526
Authority: WO
Inventors: 赵卫; 程东; 杜炳政; 布兰特•埃弗雷特•李特尔; 谢小平; 罗伊•戴维森; 王翔; 李伟恒; 姚宏鹏; 范修宏
Original assignee: 中国科学院西安光学精密机械研究所; 西安奇芯光电科技有限公司
Priority date: 2020-05-13
Filing date: 2021-04-15
Publication date: 2021-11-18
Also published as: CN111683304B; CN111683304A

Abstract

In order to solve the technical problems that the computation rate and loss of a photoelectric hybrid system are still limited by the rate and ohmic loss of an electric clock, the existing all-optical neural network system which is implemented by using discrete optical elements has high cost, large size and limited expansibility, and an error is amplified during cascading, the present invention provides an all-optical diffractive neural network and system implemented on an optical waveguide and/or an optical chip. In the present invention, an all-optical connection is realized by means of a diffraction free propagation region on the optical waveguide and/or the chip, and the connection between more waveguide neurons can be realized with the same size, thereby effectively solving the problems of a small number of neurons connected and poor system expansion; in addition, an error-tolerant rate is better as more neurons are connected, and therefore the present invention has a high recognition accuracy.

Description

在光波导和/或光芯片上实现的全光衍射神经网络及***All-optical diffraction neural network and system realized on optical waveguide and/or optical chip

背景技术Background technique

随着大数据、云计算、物联网的发展，具有感知、学习与决策的人工智能计算机的开发与推进，未来数据的体量将会爆发式增长。基于冯诺依曼结构的现代计算机在计算、感知、通信、学习以及决策等方面具有广泛应用，然而相比于其对应的生物结构——中枢神经***来说功耗大、算力弱。未来在大数据、云计算、物联网、人工智能领域，低功耗、快速而又有效的分析数据成为人们越来越关注的重点。例如在特定领域的图像实时处理、从手机上传的语音信息在云端在线识别响应等方面的应用会越来越多。着力于冯诺依曼计算结构在人工神经网络方面存在的问题，一部分学者将更多的工作主要聚焦在开发定制化的电学结构芯片(例如ASIC和FPGA)来解决人工神经网络的计算速度、高功耗问题。另一部分学者将目光转向了光学结构进行研究。这主要是因为光路结构与电路结构相比其功耗小，计算速度快。然而光电混合***通常需要频繁的光电转换，使计算速率与功耗仍然受限于电时钟的速率和欧姆损耗。With the development of big data, cloud computing, and the Internet of Things, as well as the development and advancement of artificial intelligence computers with perception, learning and decision-making, the volume of data will explode in the future. Modern computers based on the von Neumann structure have a wide range of applications in computing, perception, communication, learning, and decision-making. However, compared to its corresponding biological structure, the central nervous system, the power consumption is large and the computing power is weak. In the future, in the fields of big data, cloud computing, Internet of Things, and artificial intelligence, low power consumption, fast and effective data analysis will become the focus of more and more attention. For example, there will be more and more applications in the real-time processing of images in specific fields, and the online recognition and response of voice information uploaded from mobile phones in the cloud. Focusing on the problems of von Neumann’s computational structure in artificial neural networks, some scholars will focus more on the development of customized electrical structure chips (such as ASIC and FPGA) to solve the computational speed and high speed of artificial neural networks. Power consumption issues. Another part of scholars turned their attention to the optical structure for research. This is mainly because the optical circuit structure has low power consumption and fast calculation speed compared with the circuit structure. However, the photoelectric hybrid system usually requires frequent photoelectric conversion, so that the calculation rate and power consumption are still limited by the rate of the electrical clock and the ohmic loss.

为解决光电混合***存在的问题，在计算领域的一部分学者将目光转向了全光神经网络。目前来看，全光神经网络有潜在替代基于微电-光电混合方案的神经网络。全光神经网络优势在于其线性转换以及一定程度的非线性转换可以以光速实现，且光网络功耗很小，探测速度可以高至100GHz。众所周知，通过光学棱镜实现的傅里叶变换以及部分矩阵的光学变换不需要任何的能耗。应用棱镜的光学傅里叶变换进行计算滤波常见在光学成像方面。成像***需要诸多的分立光学元件来实现相应的矩阵变换。通过该类***完成的计算、变换、滤波速度快，几乎没有能量损耗。然而该类***有一定缺陷，例如成本高，尺寸大，稳定性差等；MIT在文章“Deep Learning with Coherent Nanophotonic Circuits”中提出了一种全光神经网络方案，该方案优势在于光网络集成在光芯片内，尺寸小。然而该方案神经元连接数少，***扩展性有限，级联时误差会被放大。In order to solve the problems of optoelectronic hybrid systems, some scholars in the field of computing have turned their attention to all-optical neural networks. At present, the all-optical neural network has the potential to replace the neural network based on the micro-electricity-photoelectric hybrid scheme. The advantage of the all-optical neural network is that its linear conversion and a certain degree of non-linear conversion can be realized at the speed of light, and the power consumption of the optical network is very small, and the detection speed can be as high as 100 GHz. As we all know, the Fourier transform realized by optical prisms and the optical transform of partial matrices do not require any energy consumption. The application of the optical Fourier transform of a prism for computational filtering is commonly used in optical imaging. The imaging system needs many discrete optical components to realize the corresponding matrix transformation. The calculation, transformation, and filtering done by this type of system are fast, and there is almost no energy loss. However, this type of system has certain defects, such as high cost, large size, poor stability, etc.; MIT proposed an all-optical neural network solution in the article "Deep Learning with Coherent Nanophotonic Circuits". The advantage of this solution is that the optical network is integrated in the optical network. Within the chip, the size is small. However, the number of neuron connections in this scheme is small, the system scalability is limited, and the error will be magnified when cascading.

发明内容Summary of the invention

为解决光电混合***计算速率与损耗仍然受限于电时钟的速率和欧姆损耗，以及现有利用分立光学元件实现全光神经网络***成本高、尺寸大、扩展性有限、级联时误差会被放大的技术问题，本发明提供了一种在光波导和/或光芯片上实现的全光衍射神经网络及***。In order to solve the calculation rate and loss of the photoelectric hybrid system is still limited by the rate and ohmic loss of the electrical clock, and the existing use of discrete optical components to achieve an all-optical neural network system has high cost, large size, limited scalability, and errors when cascading For the technical problem of amplification, the present invention provides an all-optical diffraction neural network and system implemented on an optical waveguide and/or an optical chip.

本发明的技术方案是：The technical scheme of the present invention is:

在光波导和/或光芯片上实现的全光衍射神经网络，其特殊之处在于：包括输入层、网络层和输出层；The special feature of the all-optical diffraction neural network implemented on the optical waveguide and/or optical chip is that it includes an input layer, a network layer and an output layer;

输入层用于将未载有信息的光分束、将待识别信息由电信号转换成光信号并对该光信号编码后加载至网络层；The input layer is used to split the light that does not carry information, convert the information to be identified from an electrical signal into an optical signal, encode the optical signal, and load it to the network layer;

网络层用于对接收到的光信号进行光信息全连接交换和权重调整；所述光信息全连接交换通过光衍射自由传输区实现；The network layer is used to perform optical information full connection exchange and weight adjustment on the received optical signals; the optical information full connection exchange is realized through the optical diffraction free transmission zone;

输出层用于将网络层识别后的光信号进行合并和分发。The output layer is used to combine and distribute the optical signals identified by the network layer.

进一步地，所述输入层包括依次连接的输入层输入波导、光分路器、输入层阵列波导、输入层阵列幅相调制器和输入层阵列输出波导；所述输入层输入波导为单个输入波导或者阵列输入波导。Further, the input layer includes an input layer input waveguide, an optical splitter, an input layer array waveguide, an input layer array amplitude and phase modulator, and an input layer array output waveguide connected in sequence; the input layer input waveguide is a single input waveguide Or array input waveguide.

进一步地，所述光分路器有多个，且并行设置。Further, there are multiple optical splitters, and they are arranged in parallel.

进一步地，所述光分路器为多级级联的1×2MMI、Y分路器、1×2定向耦合器、N×M MMI和光衍射自由传输区中的一种或多种结构级联构成。Further, the optical splitter is a multi-stage cascaded 1×2MMI, Y splitter, 1×2 directional coupler, N×M MMI, and optical diffraction free transmission zone. constitute.

进一步地，所述网络层由单个子网络层构成，或者由N个子网络层级联构成，前一个子网络层的输出连接下一个子网络层的输入；N≥2；Further, the network layer is composed of a single sub-network layer, or a cascade of N sub-network layers, the output of the previous sub-network layer is connected to the input of the next sub-network layer; N≥2;

每个子网络层包括通过阵列波导相连接的网络层光衍射自由传输区和网络层阵列幅相调制器；Each sub-network layer includes a network layer optical diffraction free transmission area and a network layer array amplitude and phase modulator connected by an array waveguide;

网络层光衍射自由传输区用于实现全光网络信息交换，包括依次连接的衍射区阵列输入波导、自由传输区域和衍射区输出阵列波导，阵列输入波导与自由传输区域的连接处形成阵列波导光栅结构；The optical diffraction free transmission area of the network layer is used to realize the all-optical network information exchange, including the diffractive area array input waveguide, the free transmission area and the diffraction area output array waveguide connected in sequence, and the arrayed waveguide grating is formed at the junction of the array input waveguide and the free transmission area structure;

网络层阵列幅相调制器用于实现信息交换后光信息的权重调整。The network layer array amplitude and phase modulator is used to adjust the weight of optical information after information exchange.

进一步地，至少一个子网络层中的网络层光衍射自由传输区由多个子自由传输区构成，这多个子自由传输区通过阵列波导实现串行级联、并行设置或者串并混合设置。Further, the optical diffraction free transmission area of the network layer in at least one sub-network layer is composed of a plurality of sub-free transmission areas, and the plurality of sub-free transmission areas are serially cascaded, arranged in parallel, or arranged in a mixed series and parallel through an arrayed waveguide.

进一步地，第n个子网络层中的网络层光衍射自由传输区的部分输出端连接第n+m个子网络层中的网络层光衍射自由传输区的部分输入端；1≤n≤N；1≤m≤N-n。Further, part of the output end of the optical diffraction free transmission area of the network layer in the nth sub-network layer is connected to part of the input end of the optical diffraction free transmission area of the network layer in the n+mth sub-network layer; 1≤n≤N; 1 ≤m≤Nn.

进一步地，输出层包括依次连接的输出层阵列输入波导、输出层光衍射自由传输区和输出层阵列输出波导。Further, the output layer includes an output layer array input waveguide, an output layer light diffraction free transmission area, and an output layer array output waveguide connected in sequence.

进一步地，输出层光衍射自由传输区有多个，且并行设置。Further, there are multiple light diffraction free transmission areas of the output layer, and they are arranged in parallel.

进一步地，所述输入层阵列幅相调制器通过有源调制器件或无源结构实现；所述网络层阵列幅相调制器通过有源调制器件或无源结构实现。Further, the input layer array amplitude and phase modulator is realized by an active modulation device or a passive structure; the network layer array amplitude and phase modulator is realized by an active modulation device or a passive structure.

进一步地，所述有源调制器件为电光调制器、热光调制器、磁光调制器和/或电吸收调制器；所述无源结构在实现幅相调制时，实现方式有改变波导结构的长度、宽度、厚度和/或材料折射率，在实现幅值调制时，实现方式有偏移波导和/或截断波导。Further, the active modulation device is an electro-optical modulator, a thermo-optical modulator, a magneto-optical modulator, and/or an electro-absorption modulator; when the passive structure realizes amplitude and phase modulation, the realization method may change the waveguide structure The length, width, thickness and/or refractive index of the material, when realizing amplitude modulation, are realized by offset waveguide and/or truncated waveguide.

进一步地，输入层与网络层之间的阵列波导中相邻两根波导之间具有长度差，各子网络层之间的阵列波导中相邻两根波导之间具有长度差；网络层与输出层之间的阵列波导中相邻两根波导之间具有长度差；所述长度差大于等于0，具体根据光衍射自由传输区的矩阵表达式并结合计算迭代收敛确定。Further, the arrayed waveguide between the input layer and the network layer has a length difference between two adjacent waveguides, and the arrayed waveguides between each sub-network layer have a length difference between two adjacent waveguides; the network layer and the output In the arrayed waveguides between the layers, there is a length difference between two adjacent waveguides; the length difference is greater than or equal to 0, which is specifically determined according to the matrix expression of the light diffraction free transmission zone combined with the iterative convergence of calculation.

本发明同时提供了一种在光波导和/或光芯片上实现的全光衍射神经网络***，其特殊之处在于：由上述的全光衍射神经网络任意组合构成。The present invention also provides an all-optical diffraction neural network system implemented on an optical waveguide and/or an optical chip, which is special in that it is composed of any combination of the above-mentioned all-optical diffraction neural network.

本发明的优点：Advantages of the present invention:

1.现有的全光神经网络，由于基于MZI等结构的器件神经网络连接数相对较小，所以扩展性有限；并且现有的全光神经网络需要频繁的光电转换，误差积累会被放大。而本发明提出一种在光芯片上实现的全光衍射神经网络的方案，通过光波导和/或芯片上的衍射自由传输区实现全光连接，在相同尺寸下，可实现更多的波导神经元连接，因而可有效解决神经元连接数少，***扩展弱的问题；而更多的神经元连接容错率会更好，所以本发明识别精度更高。1. The existing all-optical neural network, due to the relatively small number of device neural network connections based on MZI and other structures, has limited scalability; and the existing all-optical neural network requires frequent photoelectric conversion, and error accumulation will be amplified. The present invention proposes an all-optical diffraction neural network solution implemented on an optical chip. The all-optical connection is realized through the optical waveguide and/or the diffraction free transmission area on the chip. Under the same size, more waveguide nerves can be realized. Cell connection can effectively solve the problem of fewer neuron connections and weak system expansion; and more neuron connections will have better fault tolerance, so the invention has higher recognition accuracy.

下表1是以图11所示的全光衍射神经网络为例，通过梯度下降法计算得出不同神经元数(对应本发明的阵列波导数)对识别率的影响。从表1可以得知，提高神经元数目是提高识别率与容错率的关键因素，本发明中的全光衍射神经网络是提高神经元数的有效方法，本发明能够有效提高识别率和容错率。The following Table 1 takes the all-optical diffraction neural network shown in FIG. 11 as an example, and calculates the influence of the number of different neurons (corresponding to the number of arrayed waveguides of the present invention) on the recognition rate by the gradient descent method. It can be seen from Table 1 that increasing the number of neurons is a key factor in improving the recognition rate and fault tolerance. The all-optical diffraction neural network in the present invention is an effective method to increase the number of neurons. The present invention can effectively improve the recognition rate and fault tolerance. .

表1Table 1

2.在应用方面，基于本发明的无源光芯片，其光学神经元结构扩展性强，易于做大规模的光神经网络结构，且在器件能耗和计算速度方面更有优势；而基于本发明的有源芯片，其神经网络结构具有可编程功能，能实现自适应神经元纠错与识别目标自定义功能。2. In terms of application, the passive optical chip based on the present invention has strong scalability of optical neuron structure, easy to build a large-scale optical neural network structure, and has more advantages in terms of device energy consumption and calculation speed; and based on this The invented active chip, whose neural network structure has a programmable function, can realize the self-defined function of self-adaptive neuron error correction and target recognition.

3.本发明中输入层与网络层之间的阵列波导、各子网络层之间的阵列波导，以及网络层与输出层之间阵列波导中相邻两波导具有长度差，使得光衍射传输具有方向性，从而可以控制输出通道端口，提高了传输的可控性。3. In the present invention, the arrayed waveguides between the input layer and the network layer, the arrayed waveguides between the sub-network layers, and the arrayed waveguides between the network layer and the output layer have length differences between adjacent two waveguides, so that the optical diffraction transmission has Directivity, which can control the output channel port and improve the controllability of transmission.

附图说明Description of the drawings

图1是全光衍射神经网络结构示意图一。Figure 1 is a schematic diagram of the structure of the all-optical diffraction neural network.

图2是全光衍射神经网络输入层的结构示意图。Figure 2 is a schematic diagram of the structure of the input layer of the all-optical diffraction neural network.

图3A方框内的光分路结构可以是MMI分路结构，或定向耦合器结构，或Y分支器结构；The optical branch structure in the box in Figure 3A can be an MMI branch structure, or a directional coupler structure, or a Y-splitter structure;

图3B方框内的光分路结构可以是MMI分路结构或自由传输区结构；The optical branch structure in the box of Fig. 3B may be an MMI branch structure or a free transmission area structure;

图3C是Y分路器结构示意图。Figure 3C is a schematic diagram of the Y splitter structure.

图3D是定向耦合器结构示意图。Figure 3D is a schematic diagram of the structure of the directional coupler.

图3E是多模自映像干涉器结构示意图。Figure 3E is a schematic diagram of the structure of a multi-mode self-reflection interferometer.

图4是全光衍射神经网络网络层结构示意图。Figure 4 is a schematic diagram of the network layer structure of the all-optical diffraction neural network.

图5A是光神经网络的光衍射自由传输区结构示意图。FIG. 5A is a schematic diagram of the structure of the optical diffraction free transmission zone of the optical neural network.

图5B是光神经网络光衍射自由传输区网络结构示意图。Figure 5B is a schematic diagram of the optical neural network optical diffraction free transmission zone network structure.

图5C是光神经网络光衍射自由传输区的矩阵表达式；Fig. 5C is the matrix expression of the free transmission area of optical diffraction in the optical neural network;

图5D是光神经网络子隐藏层的矩阵表达式。Figure 5D is the matrix expression of the sub-hidden layer of the optical neural network.

图6是全光衍射神经网络输出层的结构示意图。Figure 6 is a schematic diagram of the output layer of the all-optical diffraction neural network.

图7是全光衍射神经网络的结构示意图二。Figure 7 is the second structural diagram of the all-optical diffraction neural network.

图8是全光衍射神经网络的结构示意图三。Figure 8 is the third structural diagram of the all-optical diffraction neural network.

图9是全光衍射神经网络的结构示意图四。Figure 9 is the fourth structural diagram of the all-optical diffraction neural network.

图10是全光衍射神经网络的结构示意图五。Fig. 10 is a schematic diagram of the structure of the all-optical diffraction neural network.

图11是光神经网络结构一。Figure 11 is an optical neural network structure one.

图12是光神经网络结构二。Figure 12 is the second structure of the optical neural network.

附图标记：Reference signs:

1-输入层；11-输入层阵列输入波导；12-光分路器；13-输入层阵列波导；14-输入层阵列幅相调制器；15-输入层阵列输出波导；1-input layer; 11-input layer array input waveguide; 12-optical splitter; 13-input layer array waveguide; 14-input layer array amplitude and phase modulator; 15-input layer array output waveguide;

2-网络层；21-子网络层；211-网络层光衍射自由传输区；212-网络层阵列幅相调制器；213-网络层阵列输入波导；214-网络层阵列输出波导；215-衍射区输出阵列波导；216-自由传输区域；217-衍射区阵列输入波导；218-阵列波导光栅结构；2-network layer; 21-sub-network layer; 211-network layer optical diffraction free transmission zone; 212-network layer array amplitude and phase modulator; 213-network layer array input waveguide; 214-network layer array output waveguide; 215-diffraction Zone output array waveguide; 216-free transmission zone; 217-diffraction zone array input waveguide; 218-array waveguide grating structure;

3-输出层；31-输出层阵列输入波导；32-输出层光衍射自由传输区；33-输出层阵列输出波导。3-output layer; 31-output layer array input waveguide; 32-output layer light diffraction free transmission zone; 33-output layer array output waveguide.

具体实施方式Detailed ways

以下结合附图对本发明作进一步说明。The present invention will be further described below in conjunction with the drawings.

本发明的全光衍射神经网络在光波导和/或光芯片上实现，图1是本发明全光衍射神经网络其中一种基础网络结构，图1中示出了其关键组成部分，包括输入层1、网络层2(在有些光计算结构里也称为隐藏层)和输出层3。The all-optical diffraction neural network of the present invention is implemented on optical waveguides and/or optical chips. Figure 1 is one of the basic network structures of the all-optical diffraction neural network of the present invention. Figure 1 shows its key components, including the input layer. 1. Network layer 2 (also called hidden layer in some optical computing structures) and output layer 3.

如图1、2所示，输入层1包括依次连接的输入层阵列输入波导11或单个输入波导、光分路器12、输入层阵列波导13、输入层阵列幅相调制器14和输入层阵列输出波导15；多路光从输入层1的输入层阵列输入波导11端进入，经过光分路器12分光后将未载有信息的光耦合到输入层阵列波导13上，接着通过输入层阵列波导13将光传输到输入层阵列幅相调制器14，进入待调制状态，最后通过输入层阵列幅相调制器14将待识别信息由电信号加载(或称转换)成光信号，输出至输入层阵列输出波导15端。其中，输入层阵列幅相调制器14可通过调幅来加载光信号，也可通过调相来加载光信号，或同时进行幅相调制。图1中输入层阵列幅相调制器14间的省略号代表重复并行排列的幅相调制器。As shown in Figures 1 and 2, the input layer 1 includes an input layer array input waveguide 11 or a single input waveguide, an optical splitter 12, an input layer array waveguide 13, an input layer array amplitude and phase modulator 14 and an input layer array connected in sequence Output waveguide 15; multiple light enters from the input layer array input waveguide 11 end of the input layer 1, after splitting the light by the optical splitter 12, the light that does not carry information is coupled to the input layer array waveguide 13, and then passes through the input layer array The waveguide 13 transmits the light to the input layer array amplitude and phase modulator 14 and enters the state to be modulated. Finally, the input layer array amplitude and phase modulator 14 loads (or converts) the information to be identified by the electrical signal into an optical signal, and outputs it to the input Layer array output waveguide 15 end. Among them, the input layer array amplitude and phase modulator 14 can load the optical signal by amplitude modulation, or load the optical signal by phase modulation, or perform amplitude and phase modulation at the same time. The ellipsis between the input layer array amplitude-phase modulators 14 in FIG. 1 represents the amplitude-phase modulators arranged repeatedly in parallel.

如图1、4所示，输入层1的输入层阵列输出波导15连通网络层2(或称隐藏层)的输入端。网络层2由单个子网络层21构成，或由N个子网络层21级联而成，N≥2；每个子网络层21包括通过阵列波导相连接的网络层光衍射自由传输区211和网络层阵列幅相调制器212，前一层子网络层21的输出连接下一层子网络层21的输入。光信号在N个级联的子网络层21依次通过网络层光衍射自由传输区211和网络层阵列幅相调制器212，分别完成光信息交换和光信号的权重调整。在网络层2的信息光完成最后一层子网络层21的识别后，识别信号会从最后一层子网络层21的网络层阵列输出波导214端输出。图1中网络层阵列幅相调制器212间的省略号代表重复并行排列的幅相调制器，子网络层输出波导与输入波导间的省略号代表重复串行排列的子网络层结构。As shown in Figures 1 and 4, the input layer array output waveguide 15 of the input layer 1 is connected to the input end of the network layer 2 (or hidden layer). The network layer 2 consists of a single sub-network layer 21, or a cascade of N sub-network layers 21, N≥2; each sub-network layer 21 includes a network layer optical diffraction free transmission area 211 and a network layer connected by an array waveguide Array amplitude and phase modulator 212, the output of the sub-network layer 21 of the previous layer is connected to the input of the sub-network layer 21 of the next layer. The optical signal passes through the network layer optical diffraction free transmission area 211 and the network layer array amplitude-phase modulator 212 in the N cascaded sub-network layers 21 in order to complete optical information exchange and optical signal weight adjustment respectively. After the information light of the network layer 2 completes the identification of the last sub-network layer 21, the identification signal will be output from the network layer array output waveguide 214 end of the last sub-network layer 21. The ellipsis between the network layer array amplitude and phase modulators 212 in FIG. 1 represents the amplitude and phase modulators arranged repeatedly in parallel, and the ellipsis between the output waveguide and the input waveguide of the sub-network layer represents the structure of the sub-network layer repeatedly arranged in series.

如图1、6所示，光从网络层2的网络层阵列输出波导214端输出后传输进入到输出层3的输入端。网络层2的网络层阵列输出波导214端的每根波导会一一对应的连接输出层3的输出层阵列输入波导31。光从输出层3的输入端进入后，传输至输出层光衍射自由传输区32的输入端口并发生衍射完成最后一次信息交换，接着交换的光信息通过输出层光衍射自由传输区32的阵列输出端口收集光信号，最后传输至输出层3的输出层阵列输出波导33端，接着利用光探测器(是本发明之外的结构，非本发明的一部分)收集光信号并完成光电转换。As shown in FIGS. 1 and 6, the light is output from the network layer array output waveguide 214 end of the network layer 2 and then transmitted to the input end of the output layer 3. Each waveguide at the end of the network layer array output waveguide 214 of the network layer 2 is connected to the output layer array input waveguide 31 of the output layer 3 in a one-to-one correspondence. After the light enters from the input end of the output layer 3, it is transmitted to the input port of the light diffraction free transmission area 32 of the output layer and diffraction occurs to complete the last information exchange, and then the exchanged optical information is output through the array of the light diffraction free transmission area 32 of the output layer The port collects the optical signal, and finally transmits it to the output layer array output waveguide 33 end of the output layer 3, and then uses a photodetector (a structure outside of the present invention, not part of the present invention) to collect the optical signal and complete photoelectric conversion.

通过上述描述可以看出，在整个全光衍射神经网络进行信息计算识别时是不发生光电转换的，整个识别过程只在初始加载信号和最终输出信号会进行光电转换，这样***将避免频繁的光电转换，大大提升计算效率。It can be seen from the above description that photoelectric conversion does not occur when the entire all-optical diffraction neural network performs information calculation and recognition. The entire recognition process only performs photoelectric conversion on the initial loading signal and the final output signal, so that the system will avoid frequent photoelectric conversion. Conversion greatly improves calculation efficiency.

下面依次详细描述输入层1、网络层2和输出层3的功能。The functions of the input layer 1, the network layer 2 and the output layer 3 are described in detail below in order.

如图2所示，输入层1的一个功能是将输入层阵列输入波导11的光信号进行分发传输到输入层阵列幅相调制器14上。有很多种类型的波导单元可实现光分路器12的功能。例如通过阵列1×2分路器的串行级联可完成1分N路的输出，如图3A所示。同样的，通过2×2的分路器可将N路输入层阵列输入波导11的N路信号分成M路输出。另外可能的结构，例如通过级联的MMI(多模自映像干涉器)结构将N路输入光转为M路输出光，以及通过光衍射自由传输区将N路输入光分为M路输出光，如图3B所示。输入层1的第二个功能是将光分路器12输出的光通过输入层阵列幅相调制器14完成待识别光信号编码。编码后，在输入层1输出端的待识别光信号将会在网络层2中进行识别处理。As shown in FIG. 2, one function of the input layer 1 is to distribute and transmit the optical signals of the input layer array input waveguide 11 to the input layer array amplitude-phase modulator 14. There are many types of waveguide units that can implement the function of the optical splitter 12. For example, through the serial cascade of the array 1×2 splitter, 1 minute N output can be completed, as shown in Fig. 3A. Similarly, the N-channel signals of the N-channel input layer array input waveguide 11 can be divided into M-channel output through the 2×2 splitter. Another possible structure, for example, through the cascaded MMI (Multimode Self-Mapping Interferometer) structure, the N input light is converted into M output light, and the N input light is divided into M output light through the optical diffraction free transmission zone. , As shown in Figure 3B. The second function of the input layer 1 is to pass the light output from the optical splitter 12 through the input layer array amplitude and phase modulator 14 to complete the optical signal encoding to be identified. After encoding, the optical signal to be identified at the output of the input layer 1 will be identified in the network layer 2.

如图4所示，当网络层2由N个子网络层21级联构成时，这N个子网络层21依次记为第1子网络层、…、第N子网络层；单个子网络层21包括通过阵列波导相连的网络层光衍射自由传输区211和网络层阵列幅相调制器212。每一个子网络层21会通过网络层光衍射自由传输区211完成光信号的全连接和交换。网络层光衍射自由传输区211的光学结构示意图如图5A中所示，包括依次连接的衍射区阵列输入波导217、自由传输区域216和衍射区输出阵列波导215，衍射区阵列输入波导217与自由传输区域216的连接处形成阵列波导光栅结构218，光从衍射区阵列输入波导217进入并传输到阵列波导光栅结构218时会发生衍射，衍射光在自由传输区域216传输后全连接输出到衍射区输出阵列波导215，这样便完成了光信息的全连接交换；输入层1最终输出的信号传输至网络层2的输入端口后，变成了网络层2的输入部分。网络层的输入波导我们称为网络层阵列输入波导213。在网络层2，光输入网络层阵列输入波导213耦合至网络层光衍射自由传输区211的衍射区阵列输入波导217，而在网络层光衍射自由传输区211的输出部分由一组衍射区输出阵列波导215来接收输入端衍射过来的光。网络层光衍射自由传输区211输入端的光信号通过衍射传输并耦合到自由衍射区输出部分的每个衍射区输出阵列波导215。光从网络层光衍射自由传输区211的输入端通过衍射全部连接输出到网络层光衍射自由传输区211的输出端口，我们称为光信号在网络层光衍射自由传输区211的全连接和交换。网络层光衍射自由传输区211的全连接和交换的网络示意图如图5B所示，阵列波导光栅结构和自由传输区域为光信号从每个输入端口到每个输出端口提供一个全光连接和交换，每个波导输入都会通过衍射完成一个全光连接，将信号输出至全部的波导输出端。网络层光衍射自由传输区211的全光连接和交换可通过图5C的矩阵矢量运算 (矢量矩阵相乘)来表示。图5C中的式子中，A _n表示衍射区阵列输入波导217中第N个输入波导端口的复数值，该值包含有第N个输入波导端口的幅值和相位信息；B _nm表示从衍射区阵列输入波导217中第N个输入波导端口衍射传输到衍射区输出阵列波导215中第M个输出波导端口的复数值，该值包含第N个输入波导端口衍射传输到衍射区第M个输出波导端口的幅值和相位信息；C _n表示第衍射区输出阵列波导215中第N个输出波导端的光信号复数值，该复数值包含有第N个输出波导的幅值信息和相位信息。在神经网络的网络层，神经网络的传输函数由网络层光衍射自由传输区211和输入端的子波干涉决定，而子波复数值是由前一层网络(即输入层1或前一层的子网络层21)的幅相调制器决定的；每一个子网络层21的另一个功能是通过网络层阵列幅相调制器212提供信号的权重。每个在网络层光衍射自由传输区211处理过的光信号接着被网络层阵列幅相调制器212进行调制。我们称网络层阵列幅相调制器212的信号处理为权重。最后，在网络层阵列幅相调制器212处理完信号后，最后一层子网络层的网络层阵列幅相调制器212输出端的网络层阵列输出波导214变为整个网络层2的输出端；单个子网络层的整体结构提供一个矩阵乘加和的功能，矩阵计算公式如图5D所示，M _n表示网络层阵列幅相调制器212中第N个子网络层的网络层幅相调制器的复数值，包含第N个子网络层的网络层幅相调制器的幅值和相位调制信息；A′ _n表示子网络层的输出，同样也是下一子网络层的输入。 As shown in Figure 4, when the network layer 2 is composed of N sub-network layers 21 cascaded, the N sub-network layers 21 are sequentially denoted as the first sub-network layer, ..., the N-th sub-network layer; a single sub-network layer 21 includes The network layer optical diffraction free transmission area 211 and the network layer array amplitude and phase modulator 212 are connected by the array waveguide. Each sub-network layer 21 completes the full connection and exchange of optical signals through the optical diffraction free transmission zone 211 of the network layer. The schematic diagram of the optical structure of the optical diffraction free transmission area 211 of the network layer is shown in FIG. An arrayed waveguide grating structure 218 is formed at the junction of the transmission area 216. When light enters from the diffractive area array input waveguide 217 and is transmitted to the arrayed waveguide grating structure 218, diffraction occurs. After the diffracted light is transmitted in the free transmission area 216, it is fully connected and output to the diffraction area. The output array waveguide 215 completes the full-connection exchange of optical information; after the final output signal of the input layer 1 is transmitted to the input port of the network layer 2, it becomes the input part of the network layer 2. The input waveguide of the network layer is called the network layer array input waveguide 213. In the network layer 2, the optical input network layer array input waveguide 213 is coupled to the diffraction zone array input waveguide 217 of the network layer optical diffraction free transmission zone 211, and the output part of the network layer optical diffraction free transmission zone 211 is output by a set of diffraction zones The arrayed waveguide 215 receives the light diffracted from the input end. The optical signal at the input end of the optical diffraction free transmission area 211 of the network layer is transmitted through diffraction and coupled to each diffraction area output array waveguide 215 of the output portion of the free diffraction area. The light from the input end of the optical diffraction free transmission area 211 of the network layer is all connected and output through diffraction to the output port of the optical diffraction free transmission area 211 of the network layer. We call it the full connection and exchange of the optical signal in the optical diffraction free transmission area 211 of the network layer. . The network diagram of the full connection and exchange of the optical diffraction free transmission area 211 at the network layer is shown in Fig. 5B. The arrayed waveguide grating structure and the free transmission area provide an all-optical connection and exchange for the optical signal from each input port to each output port. , Each waveguide input will complete an all-optical connection through diffraction, and output the signal to all waveguide output ends. The all-optical connection and exchange of the optical diffraction free transmission zone 211 of the network layer can be represented by the matrix vector operation (vector matrix multiplication) of FIG. 5C. In the formula in Fig. 5C, A _n represents the complex value of the N-th input waveguide port in the diffractive zone array input waveguide 217, and this value contains the amplitude and phase information of the N-th input waveguide port; B _nm represents the diffracted The complex value of the Nth input waveguide port in the array input waveguide 217 diffracted and transmitted to the M output waveguide port in the diffracted output array waveguide 215. This value includes the Nth input waveguide port diffracted and transmitted to the M-th output in the diffractive area The amplitude and phase information of the waveguide port; C _n represents the complex value of the optical signal at the end of the Nth output waveguide in the output array waveguide 215 in the diffraction zone, and the complex value contains the amplitude information and phase information of the Nth output waveguide. In the network layer of the neural network, the transfer function of the neural network is determined by the light diffraction free transmission area 211 of the network layer and the wavelet interference at the input end, and the wavelet complex value is determined by the previous layer of the network (that is, the input layer 1 or the previous layer The amplitude and phase modulator of the sub-network layer 21) is determined; another function of each sub-network layer 21 is to provide the signal weight through the network layer array amplitude and phase modulator 212. Each optical signal processed in the optical diffraction free transmission area 211 of the network layer is then modulated by the network layer array amplitude and phase modulator 212. We call the signal processing of the network layer array amplitude and phase modulator 212 as the weight. Finally, after the network layer array amplitude and phase modulator 212 processes the signal, the network layer array output waveguide 214 at the output end of the network layer array amplitude and phase modulator 212 of the last sub-network layer becomes the output end of the entire network layer 2; the overall configuration of a sub-network layer and matrix multiply-add function, as shown in FIG matrix calculation formula, M _n represents the complex amplitude and phase modulator of the network layer of the network layer of amplitude and phase modulator array 212 of N sub-network layer 5D Value, including the amplitude and phase modulation information of the network layer amplitude and phase modulator of the Nth sub-network layer; _A'n represents the output of the sub-network layer, which is also the input of the next sub-network layer.

如图6所示，输出层3提供一个在网络层2识别信号的合并和分发功能。有多种波导结构单元可以实现这一功能。例如并行排列的阵列2×1MMI结构可将N个输入波导的N路信号分配成N/2路信号。其他可能的分布结构，例如将N路阵列波导的N路信号分配成M路信号，或者通过MMI耦合器，亦或者通过并行的光衍射自由传输区。在输出层3的输出波导的信号通过上述结构分配完信号后，随即传输到芯片外并通过光电探测器把光信号转换为电信号。As shown in Figure 6, the output layer 3 provides a function of combining and distributing identification signals at the network layer 2. There are a variety of waveguide structure units that can achieve this function. For example, a parallel array of 2×1 MMI structure can distribute N signals of N input waveguides into N/2 signals. Other possible distribution structures, such as distributing N signals of N array waveguides into M signals, or through MMI couplers, or through parallel optical diffraction free transmission areas. After the signal of the output waveguide of the output layer 3 has been distributed through the above structure, it is then transmitted to the outside of the chip and the optical signal is converted into an electrical signal by a photodetector.

对于输入层1中的光分路器12，有四种结构或者这四种结构的任意组合可以实现，这四种结构分别为Y分路器、定向耦合器、MMI结构、以及光衍射自由传输区，相应的示意图如图3C、3D、3E、5A所示。在这些结构中，Y分路器，1×2MMI结构，1×2定向耦合器可将在1路波导上的1路输入信号分为2路输出，产生一个一分二的分路功能。一组串行级联的1×2分路器可将1个输入波导的1路光信号分到N路输出波导上。对于光衍射自由传输区结构的光分路器，在N路输入波导上的N路信号可以通过光衍射自由传输区的输入端衍射并在光衍射自由传输区传输扩散，接着在光衍射自由传输区的输出端的M个阵列波导上进行接收，这一传输过程提供一个N×M的光信号分配。For the optical splitter 12 in the input layer 1, there are four structures or any combination of these four structures that can be realized. The four structures are Y splitter, directional coupler, MMI structure, and free optical diffraction transmission. The corresponding schematic diagrams are shown in Figures 3C, 3D, 3E, and 5A. In these structures, the Y splitter, 1×2MMI structure, and 1×2 directional coupler can divide one input signal on one waveguide into two outputs, resulting in a one-to-two splitting function. A group of serially cascaded 1×2 splitters can split 1 optical signal from 1 input waveguide to N output waveguides. For the optical splitter with the structure of the optical diffraction free transmission zone, the N signals on the N input waveguide can be diffracted by the input end of the optical diffraction free transmission zone and transmitted and diffused in the optical diffraction free transmission zone, and then freely transmitted in the optical diffraction Receive on the M arrayed waveguides at the output end of the zone. This transmission process provides an N×M optical signal distribution.

对于输入层1中的输入层阵列幅相调制器14和网络层2中的网络层阵列幅相调制器212结构相同，但是作用不同；输入层阵列幅相调制器14通过调幅调相表达待识别信息，网络层阵列幅相调制器212通过调幅调相表达光信号权重调整。输入层阵列幅相调制器14和网络层阵列幅相调制器212，均可通过有源或无源的结构实现。在无源结构的幅相调制上，相位的调制可通过改变阵列波导的宽度、厚度、长度和/或材料折射率来完成，幅值的调制可通过波导的偏移、截断等方式完成，或者上述任意方法的组合可完成幅相调制。例如，网络层2的一部分网络层阵列输出波导214可以根据计算参数将波导的宽度和长度按照某种程度进行变化，这样便可完成无源的相位调制，实现权重的一个表达。同样的，在一部分输入层1的输入层阵列输出波导15上，选择性的截断波导可以实现一组输入图像的二元编码功能。在有源调制器件方面，幅相信息的调制可以通过电光效应、热光效应、磁光效应或电吸收的方式实现。例如，在波导周围分布的热调制器可以同时提供幅值和相位的调制。The input layer array amplitude and phase modulator 14 in the input layer 1 and the network layer array amplitude and phase modulator 212 in the network layer 2 have the same structure, but have different functions; the input layer array amplitude and phase modulator 14 expresses to be identified through amplitude and phase modulation Information, the network layer array amplitude and phase modulator 212 expresses the weight adjustment of the optical signal through amplitude and phase modulation. Both the input layer array amplitude and phase modulator 14 and the network layer array amplitude and phase modulator 212 can be realized by an active or passive structure. In the amplitude and phase modulation of the passive structure, the phase modulation can be completed by changing the width, thickness, length and/or refractive index of the material of the arrayed waveguide, and the amplitude modulation can be completed by the offset and truncation of the waveguide, or A combination of any of the above methods can complete amplitude and phase modulation. For example, a part of the network layer array output waveguide 214 of the network layer 2 can change the width and length of the waveguide to a certain extent according to the calculated parameters, so that passive phase modulation can be completed and an expression of weight can be realized. Similarly, on a part of the input layer array output waveguide 15 of the input layer 1, the selective truncation of the waveguide can realize the binary encoding function of a group of input images. In terms of active modulation devices, the modulation of amplitude and phase information can be achieved by means of electro-optical effect, thermo-optical effect, magneto-optical effect or electrical absorption. For example, thermal modulators distributed around the waveguide can provide both amplitude and phase modulation.

对于网络层2中的网络层光衍射自由传输区211和输出层3中的输出层光衍射自由传输区32，其结构可以为矩形或弧形。光衍射自由传输区的输入和输出两边端口以及衍射区域提供一个全光网络的信息传输和交换功能。输入输出端口的宽度，以及衍射区域的尺寸和形貌决定了网络传输的网络矩阵参数。For the network layer light diffraction free transmission area 211 in the network layer 2 and the output layer light diffraction free transmission area 32 in the output layer 3, the structure may be rectangular or arc-shaped. The input and output ports of the optical diffraction free transmission zone and the diffraction zone provide the information transmission and exchange function of an all-optical network. The width of the input and output ports, as well as the size and shape of the diffraction area determine the network matrix parameters of network transmission.

图1中，输入层1、网络层2和输出层3是串行级联的，输入层1只有一个级联的光分路区域，每个单独的子网络层21只有一个光衍射区域。图1的网络结构可以理解为，待分类识别信息从进入光芯片到识别完成后输出至光芯片外的整个过程，是通过待识别信息的输入(输入层1)，计算与识别(网络层2)、输出结果的分类(输出层3)逐层计算处理完成的。输入层1待识别信息的输入，网络层2的计算识别方法和输出层3的结果输出是没有进行分类输入、计算和输出的。In Figure 1, the input layer 1, the network layer 2 and the output layer 3 are cascaded in series, the input layer 1 has only one cascaded optical branching area, and each individual sub-network layer 21 has only one light diffraction area. The network structure of Figure 1 can be understood as the whole process of the identification information to be classified from entering the optical chip to outputting to the outside of the optical chip after the identification is completed, through the input of the information to be identified (input layer 1), calculation and identification (network layer 2) ), the classification of the output result (output layer 3) is calculated and processed layer by layer. The input of the information to be identified in the input layer 1, the calculation and identification method of the network layer 2 and the result output of the output layer 3 are not classified input, calculation and output.

图7为本发明全光衍射神经网络的第二种基础结构，该结构中网络层2和输出层3的网络结构与图1是一致的。也就是说针对识别信号，网络层2的计算识别方法和输出层3的结果输出是不分类的。与图1相比，图7的区别在于，输入层1的光分路器12是多层并行分布的，也可以看成是输入层1由多个并行设置的子输入层构成。图7中并行的光分路器可以理解为，不同类型或不同种类的待识别信息同时从输入层的输入端口进入，并从输出端口输出进入到网络层2处理。FIG. 7 is the second basic structure of the all-optical diffraction neural network of the present invention. In this structure, the network structure of the network layer 2 and the output layer 3 is the same as that of FIG. 1. That is to say, for the recognition signal, the calculation recognition method of the network layer 2 and the result output of the output layer 3 are not classified. Compared with FIG. 1, the difference of FIG. 7 is that the optical splitter 12 of the input layer 1 is distributed in multiple layers in parallel, and it can also be seen that the input layer 1 is composed of multiple sub-input layers arranged in parallel. The parallel optical splitter in FIG. 7 can be understood as that different types or types of information to be identified enter at the same time from the input port of the input layer, and output from the output port to the network layer 2 for processing.

图8为本发明全光衍射神经网络的第三种基础结构，该结构中输入层1和网络层2的网络结构与图1相比是相同的。也就是说，在网络结构中输入层1的待识别信息输入与网络层2的计算识别方法是没有进行分类输入和分类计算的。与图1相比，图8的区别在于，输出层3的输出层光衍射自由传输区32是一个多层并联结构，输出波导数可根据需求进行相应的增加、减少或不变。这一并联结构可以理解为识别结果是多种类型的，不同种类或不同类型的输出结果同时从全光神经衍射网络芯片的输出端口输出，并进行光电转换。FIG. 8 is the third basic structure of the all-optical diffraction neural network of the present invention. In this structure, the network structure of the input layer 1 and the network layer 2 is the same as that of FIG. 1. That is to say, in the network structure, the input of the information to be identified in the input layer 1 and the calculation and recognition method of the network layer 2 do not perform classification input and classification calculation. Compared with FIG. 1, the difference of FIG. 8 is that the output layer light diffraction free transmission area 32 of the output layer 3 is a multi-layer parallel structure, and the number of output waveguides can be increased, decreased or unchanged according to requirements. This parallel structure can be understood as the recognition results are of multiple types, and different types or types of output results are simultaneously output from the output port of the all-optical neural diffraction network chip and undergo photoelectric conversion.

图9为本发明全光衍射神经网络的第四种基础结构，该结构中输入层1 和输出层3的网络结构与图1相比是相同的。也就是说，输入层1的信息输入与输出层3的识别结果输出是没有进行分类的。与图1相比，图9的区别在于，网络层2的网络层光衍射自由传输区211是多层并行结构。并行的网络层光衍射自由传输区可以理解为单层网络层通过子衍射结构具有了多种信息处理的功能，或可以理解为单层识别结果为多信息识别或多处理方式融合的结果。FIG. 9 is the fourth basic structure of the all-optical diffraction neural network of the present invention. In this structure, the network structure of the input layer 1 and the output layer 3 is the same as that of FIG. 1. In other words, the information input of input layer 1 and the recognition result output of output layer 3 are not classified. Compared with FIG. 1, the difference of FIG. 9 is that the network layer light diffraction free transmission area 211 of the network layer 2 is a multi-layer parallel structure. The parallel network layer optical diffraction free transmission zone can be understood as a single-layer network layer with multiple information processing functions through the sub-diffraction structure, or can be understood as a single-layer recognition result as a result of multi-information recognition or multi-processing fusion.

图10为本发明全光衍射神经网络的第五种基础结构，该结构中输入层1和输出层3的网络结构与图9相比是相同的，区别在于，图10中网络层2的第n个子网络层中的网络层光衍射自由传输区的部分输出端连接第n+m个子网络层中的网络层光衍射自由传输区的部分输入端；1≤n≤N；1≤m≤N-n；N为网络层2中子网络层的总层数。这可以理解为一部分信号不需要在某一个或者某几个子网络层中进行具体的计算，而直接进入到下一子网络层或者下下一个子网络层甚至最后一个子网络层进行处理。Fig. 10 is the fifth basic structure of the all-optical diffraction neural network of the present invention. The network structure of the input layer 1 and the output layer 3 in this structure is the same as that of Fig. 9, the difference is that the network layer 2 in Fig. 10 Part of the output end of the optical diffraction free transmission area of the network layer in the n sub-network layers is connected to part of the input end of the optical diffraction free transmission area of the network layer in the n+m sub-network layer; 1≤n≤N; 1≤m≤Nn ; N is the total number of sub-network layers in network layer 2. This can be understood as a part of the signal that does not require specific calculations in one or several sub-network layers, but directly enters the next sub-network layer or the next sub-network layer or even the last sub-network layer for processing.

除了上述图1、图7-10所示的五个基础网络结构，更为复杂功能的网络可以通过上述五种基础网络结构的任意组合实现。In addition to the five basic network structures shown in Figures 1 and 7-10, networks with more complex functions can be implemented through any combination of the five basic network structures.

为了更好的理解本发明，以下给出两个示例。In order to better understand the present invention, two examples are given below.

示例1：Example 1:

如图11所示，本示例所提供的全光衍射神经网络具有图像分类识别的功能，网络结构包括一层输入层1、由两层子网络层21构成的网络层2和一层输出层3。这一光网络结构是基于无源集成光波导实现的。本示例的输入层1是由一个输入层输入波导11、一个光分路器12(本示例由光衍射自由传输区实现)、一组输入层阵列幅相调制器14(本示例为无源幅值调制器，通过截断波导实现)和一组输入层阵列输出波导15组成的。相干光从输入层1的输入端口进入到全光衍射神经网络中，并且通过光衍射自由传输区将1路光分为128路光输出至光衍射自由传输区的输出端。紧接着根据自定义的编码方式将128路光通过无源幅值调制器将相应的幅值信息通过光载入到输入层阵列输出波导15内。紧接着，输入层1部分的输出端成为网络层2部分的输入端。在本示例中，需要注意的是在不考虑无源幅值调制器带来的相位误差的情况下，网络层2与输入层1之间连接的阵列波导的相位误差为0，也就是说在相同工艺条件下，连接输入层1与网络层2的阵列波导长度、宽度和厚度是相同的；本示例的网络层2中，每个子网络层21都由一组网络层阵列输入波导213、一个网络层光衍射自由传输区211、一组网络层阵列幅相调制器212和一组网络层阵列输出波导214组成。在网络层2，载有图像信息的光通过网络层阵列输入波导213端输入后，信息的交换和识别分别通过网络层光衍射自由传输区211和网络层阵列幅相调制器212完成。在本示例中，128通道的信息可以瞬间在网络层光衍射自由传输区211交换完成，接着在网络层阵列幅相调制器212部分通过改变阵列波导的宽度和长度产生相位差达到相位调制的要求。在完成网络层2的识别后，载有图像识别结果的光在输出层3的输出层光衍射自由传输区内32完成最后一次信息交换与分配后将从输出层3的输出层阵列输出波导33输出。本示例的无源全光衍射神经网络示例的优势是光计算网络的能量消耗几乎为0，并且大规模光神经元的光神经网络易于通过本方案实现。As shown in Figure 11, the all-optical diffraction neural network provided by this example has the function of image classification and recognition. The network structure includes an input layer 1, a network layer 2 composed of two sub-network layers 21, and an output layer 3. . This optical network structure is based on passive integrated optical waveguides. The input layer 1 of this example is composed of an input layer input waveguide 11, an optical splitter 12 (this example is realized by the optical diffraction free transmission zone), a set of input layer array amplitude and phase modulators 14 (this example is a passive amplitude The value modulator is realized by truncating the waveguide) and a set of input layer array output waveguide 15. The coherent light enters the all-optical diffraction neural network from the input port of the input layer 1, and the light is divided into 128 channels through the optical diffraction free transmission area and output to the output end of the optical diffraction free transmission area. Then, according to the custom encoding method, 128 channels of light are passed through the passive amplitude modulator, and the corresponding amplitude information is loaded into the input layer array output waveguide 15 through the light. Immediately afterwards, the output terminal of the input layer 1 part becomes the input terminal of the network layer 2 part. In this example, it should be noted that without considering the phase error caused by the passive amplitude modulator, the phase error of the arrayed waveguide connected between the network layer 2 and the input layer 1 is 0, which means that Under the same process conditions, the length, width and thickness of the arrayed waveguides connecting the input layer 1 and the network layer 2 are the same; in the network layer 2 of this example, each sub-network layer 21 consists of a set of network layer array input waveguides 213, one The network layer optical diffraction free transmission area 211, a group of network layer array amplitude and phase modulators 212 and a group of network layer array output waveguides 214 are composed. In the network layer 2, after the light carrying image information is input through the network layer array input waveguide 213, the information exchange and recognition are respectively completed through the network layer light diffraction free transmission area 211 and the network layer array amplitude and phase modulator 212. In this example, 128 channels of information can be exchanged instantly in the optical diffraction free transmission zone 211 of the network layer, and then in the network layer array amplitude and phase modulator 212 part, the phase difference is generated by changing the width and length of the array waveguide to meet the requirements of phase modulation. . After the recognition of the network layer 2 is completed, the light carrying the result of the image recognition will be output from the output layer array of the output layer 3 to the output layer array after the completion of the last information exchange and distribution in the output layer light diffraction free transmission zone 32 of the output layer 3 Output. The advantage of the passive all-optical diffraction neural network example of this example is that the energy consumption of the optical computing network is almost zero, and the optical neural network of large-scale optical neurons is easy to implement through this solution.

示例2：Example 2:

如图12所示，本示例所提供的全光衍射神经网络具有元音分类识别的功能，网络结构包括一层输入层1、由四层子网络层级联构成的网络层2和一层输出层3。这一光网络是基于有源集成光波导实现的。在本示例中，连接输入层1与网络层2、各子网络层的阵列波导的长度差有两种，为0μm和11.2μm。其中，输入层1与第1子网络层之间的阵列波导中相邻两根波导的长度差、第1子网络层和第2子网络层之间的阵列波导中相邻两根波导的长度差、第3子网络层和第4子网络层之间的阵列波导中相邻两根波导的长度差、第4子网络层和输出层3之间的阵列波导中相邻两根波导之间的长度差均为11.2μm。第2子网络层和第3子网络层之间的阵列波导中相邻两根波导之间的长度差为0μm；波导之间的长度差可以通过图5C的矩阵表达式并结合计算迭代收敛即可确定。该网络结构的输入层1是由输入波导、6级串行级联的光分路器12(MMI结构)、1组64路的输入层阵列幅相调制器14(本示例的输入层阵列幅相调制器14是通过热光调制器阵列实现的)、以及1组64通道的输入层阵列输出波导15组成。1路光从6级串行级联的光分路器12经过后，分成了等幅值的64路光，接着光从光分路器12输出连接到一组64路的输入层阵列幅相调制器14上，电元音信号是通过这一组64路输入层阵列幅相调制器14将电信号转为载有元音信息的光信号，并输出至64通道的输入层阵列输出波导15。从输入层1输出的光紧接着进入到网络层2的输入端。每一单个子网络层21由一组64路网络层阵列输入波导213、一个网络层光衍射自由传输区211，一组64路的网络层阵列幅相调制器212和一组64路的网络层阵列输出波导214组成。携带有元音信息的光通过四层子网络层的网络层光衍射自由传输区211和网络层阵列幅相调制器212分别完成了信息交换和信息权重计算，在第4子网络层21的输出端输出后，识别后的光信息进入到了输出层3。光信息在输出层3的输出层光衍射自由传输区内32完成最后一次信息交换与分配后输出至输出层3的输出层阵列输出波导33。最后，识别后的元音光信息通过光电转换器转换成电信号。相比于示例1，本示例的优势在于上位机控制器可以实时的基于在线结果进行分析反馈。上位机控制器可以根据测试结果，调整网络层的网络层阵列幅相调制器212来提高识别率。除此之外，上位机控制器可以根据待识别种类的需求，通过改变网络层阵列幅相调制器212的调制信息，使全光衍射神经网络芯片具有识别内容定制化的功能。As shown in Figure 12, the all-optical diffraction neural network provided by this example has the function of vowel classification and recognition. The network structure includes an input layer 1, a network layer 2 composed of a cascade of four sub-network layers, and an output layer 3. This optical network is based on active integrated optical waveguides. In this example, there are two types of length differences between the arrayed waveguides connecting the input layer 1 and the network layer 2 to each sub-network layer, which are 0 μm and 11.2 μm. Among them, the length difference between two adjacent waveguides in the arrayed waveguide between the input layer 1 and the first sub-network layer, and the length of two adjacent waveguides in the arrayed waveguide between the first sub-network layer and the second sub-network layer Difference, the length difference between two adjacent waveguides in the arrayed waveguide between the third sub-network layer and the fourth sub-network layer, between the two adjacent waveguides in the arrayed waveguide between the fourth sub-network layer and the output layer 3 The difference in length is 11.2μm. The length difference between two adjacent waveguides in the arrayed waveguide between the second sub-network layer and the third sub-network layer is 0 μm; the length difference between the waveguides can be calculated through the matrix expression of Fig. 5C and combined with the iterative convergence of calculation. Can be determined. The input layer 1 of the network structure is composed of an input waveguide, a 6-stage serial cascaded optical splitter 12 (MMI structure), a group of 64 channels of input layer array amplitude and phase modulator 14 (the input layer array amplitude of this example) The phase modulator 14 is realized by a thermo-optic modulator array), and a group of 64-channel input layer array output waveguide 15 is composed. After 1 path of light passes through the 6-stage serial cascaded optical splitter 12, it is divided into 64 paths of equal amplitude, and then the light output from the optical splitter 12 is connected to a set of 64 paths of input layer array amplitude. On the modulator 14, the electrical vowel signal is converted into an optical signal carrying vowel information through the 64-channel input layer array amplitude and phase modulator 14, and output to the 64-channel input layer array output waveguide 15 . The light output from the input layer 1 then enters the input end of the network layer 2. Each single sub-network layer 21 consists of a set of 64-channel network layer array input waveguide 213, a network-layer optical diffraction free transmission area 211, a group of 64-channel network layer array amplitude and phase modulators 212 and a group of 64-channel network layer Array output waveguide 214 is composed. The light carrying vowel information passes through the network layer optical diffraction free transmission area 211 of the four sub-network layers and the network layer array amplitude and phase modulator 212 respectively completes information exchange and information weight calculation, and the output of the fourth sub-network layer 21 After the terminal is output, the identified optical information enters the output layer 3. The optical information is output to the output layer array output waveguide 33 of the output layer 3 after completing the last information exchange and distribution in the optical diffraction free transmission zone 32 of the output layer 3. Finally, the identified vowel and light information is converted into electrical signals by a photoelectric converter. Compared with example 1, the advantage of this example is that the upper computer controller can perform analysis and feedback based on online results in real time. The host computer controller can adjust the network layer array amplitude and phase modulator 212 of the network layer according to the test result to improve the recognition rate. In addition, the host computer controller can change the modulation information of the network layer array amplitude and phase modulator 212 according to the needs of the type to be identified, so that the all-optical diffraction neural network chip has the function of customizing the identification content.

Claims

在光波导和/或光芯片上实现的全光衍射神经网络，其特征在于：包括输入层(1)、网络层(2)和输出层(3)；The all-optical diffraction neural network implemented on the optical waveguide and/or optical chip is characterized in that it includes an input layer (1), a network layer (2) and an output layer (3);

输入层(1)用于将未载有信息的光分束、将待识别信息由电信号转换成光信号并对该光信号编码后加载至网络层(2)；The input layer (1) is used to split the light that does not carry information, convert the information to be identified from an electrical signal into an optical signal, encode the optical signal, and load it to the network layer (2);

网络层(2)用于对接收到的光信号进行光信息全连接交换和权重调整；所述光信息全连接交换通过光衍射自由传输区实现；The network layer (2) is used to perform optical information full connection exchange and weight adjustment on the received optical signals; the optical information full connection exchange is realized through the optical diffraction free transmission zone;

输出层(3)用于将网络层(2)识别后的光信号进行合并和分发。The output layer (3) is used to combine and distribute the optical signals identified by the network layer (2).
根据权利要求1所述的在光波导和/或光芯片上实现的全光衍射神经网络，其特征在于：所述输入层(1)包括依次连接的输入层输入波导、光分路器(12)、输入层阵列波导(13)、输入层阵列幅相调制器(14)和输入层阵列输出波导(15)；所述输入层输入波导为单个输入波导或者阵列输入波导。The all-optical diffraction neural network implemented on optical waveguides and/or optical chips according to claim 1, characterized in that: the input layer (1) comprises input layer input waveguides and optical splitters (12 ), the input layer array waveguide (13), the input layer array amplitude and phase modulator (14) and the input layer array output waveguide (15); the input layer input waveguide is a single input waveguide or an array input waveguide.
根据权利要求2所述的在光波导和/或光芯片上实现的全光衍射神经网络，其特征在于：所述光分路器(12)有多个，且并行设置。The all-optical diffractive neural network implemented on an optical waveguide and/or an optical chip according to claim 2, characterized in that there are multiple optical splitters (12), and they are arranged in parallel.
根据权利要求2所述的在光波导和/或光芯片上实现的全光衍射神经网络，其特征在于：所述光分路器(12)为多级级联的1×2MMI、Y分路器、1×2定向耦合器、N×M MMI和光衍射自由传输区中的一种或多种结构级联构成。The all-optical diffraction neural network implemented on the optical waveguide and/or optical chip according to claim 2, characterized in that: the optical splitter (12) is a multi-stage cascaded 1×2 MMI, Y branch One or more of the structure cascade is composed of a filter, a 1×2 directional coupler, a N×M MMI, and a light diffraction free transmission zone.
根据权利要求1所述的在光波导和/或光芯片上实现的全光衍射神经网络，其特征在于：所述网络层(2)由单个子网络层(21)构成，或者由N个子网络层(21)级联构成，前一个子网络层(21)的输出连接下一个子网络层(21)的输入；N≥2；The all-optical diffraction neural network implemented on the optical waveguide and/or optical chip according to claim 1, characterized in that: the network layer (2) is composed of a single sub-network layer (21), or is composed of N sub-networks The layers (21) are cascaded, the output of the previous sub-network layer (21) is connected to the input of the next sub-network layer (21); N≥2;

每个子网络层(21)包括通过阵列波导相连接的网络层光衍射自由传输区(211)和网络层阵列幅相调制器(212)；Each sub-network layer (21) includes a network layer optical diffraction free transmission area (211) and a network layer array amplitude and phase modulator (212) connected by an array waveguide;

网络层光衍射自由传输区(211)用于实现全光网络信息交换，包括依次连接的衍射区阵列输入波导(217)、自由传输区域(216)和衍射区输出阵列波导(215)，阵列输入波导(217)与自由传输区域(216)的连接处形成阵列波导光栅结构(218)；The optical diffraction free transmission area (211) of the network layer is used to realize the all-optical network information exchange, including the diffraction area array input waveguide (217), the free transmission area (216) and the diffraction area output array waveguide (215), which are connected in sequence, and the array input An arrayed waveguide grating structure (218) is formed at the junction of the waveguide (217) and the free transmission area (216);

网络层阵列幅相调制器(212)用于实现信息交换后光信息的权重调整。The network layer array amplitude and phase modulator (212) is used to realize the weight adjustment of the optical information after the information exchange.
根据权利要求5所述的在光波导和/或光芯片上实现的全光衍射神经网络，其特征在于：至少一个子网络层(21)中的网络层光衍射自由传输区(211)由多个子自由传输区构成，这多个子自由传输区通过阵列波导实现串行级联、并行设置或者串并混合设置。The all-optical diffraction neural network implemented on the optical waveguide and/or optical chip according to claim 5, characterized in that: the optical diffraction free transmission area (211) of the network layer in at least one sub-network layer (21) is composed of more It is composed of multiple free transmission areas, and the multiple free transmission areas are serially cascaded, arranged in parallel, or arranged in a mixed series and parallel through the arrayed waveguide.
根据权利要求6所述的在光波导和/或光芯片上实现的全光衍射神经网络，其特征在于：第n个子网络层中的网络层光衍射自由传输区的部分输出端连接第n+m个子网络层中的网络层光衍射自由传输区的部分输入端；1≤n≤N；1≤m≤N-n。The all-optical diffraction neural network implemented on the optical waveguide and/or optical chip according to claim 6, characterized in that: part of the output end of the optical diffraction free transmission zone of the network layer in the nth sub-network layer is connected to the n+th sub-network layer. Part of the input end of the optical diffraction free transmission area of the network layer in the m sub-network layers; 1≤n≤N; 1≤m≤Nn.
根据权利要求1、2、3、4、5、6或7所述的全光衍射神经网络，其特征在于：输出层(3)包括依次连接的输出层阵列输入波导(31)、输出层光衍射自由传输区(32)和输出层阵列输出波导(33)。The all-optical diffraction neural network according to claim 1, 2, 3, 4, 5, 6 or 7, characterized in that: the output layer (3) comprises an output layer array input waveguide (31) and an output layer optical Diffraction free transmission zone (32) and output layer array output waveguide (33).
根据权利要求8所述的全光衍射神经网络，其特征在于：输出层光衍射自由传输区(32)有多个，且并行设置。The all-optical diffraction neural network according to claim 8, characterized in that there are multiple light diffraction free transmission areas (32) in the output layer, and they are arranged in parallel.
根据权利要求9所述的全光衍射神经网络，其特征在于：所述输入层阵列幅相调制器(14)通过有源调制器件或无源结构实现；所述网络层阵列幅相调制器(212)通过有源调制器件或无源结构实现。The all-optical diffraction neural network according to claim 9, characterized in that: the input layer array amplitude and phase modulator (14) is realized by an active modulation device or a passive structure; the network layer array amplitude and phase modulator (14) 212) Realized by an active modulation device or a passive structure.
根据权利要求10所述的全光衍射神经网络，其特征在于：所述有源调制器件为电光调制器、热光调制器、磁光调制器和/或电吸收调制器；所述无源结构在实现幅相调制时，实现方式有改变波导结构的长度、宽度、厚度和/或材料折射率，在实现幅值调制时，实现方式有偏移波导和/或截断波导。The all-optical diffraction neural network according to claim 10, wherein the active modulation device is an electro-optical modulator, a thermo-optical modulator, a magneto-optical modulator and/or an electro-absorption modulator; the passive structure When realizing amplitude and phase modulation, the realization method includes changing the length, width, thickness and/or material refractive index of the waveguide structure. When realizing amplitude modulation, the realization method includes offset waveguide and/or truncated waveguide.
根据权利要求5、6或7所述的全光衍射神经网络，其特征在于：输入层(1)与网络层(2)之间的阵列波导中相邻两根波导之间具有长度差，各子网络层之间的阵列波导中相邻两根波导之间具有长度差；网络层(2)与输出层(3)之间的阵列波导中相邻两根波导之间具有长度差；所述长度差大于等于0，具体根据光衍射自由传输区的矩阵表达式并结合计算迭代收敛确定。The all-optical diffraction neural network according to claim 5, 6 or 7, wherein the arrayed waveguide between the input layer (1) and the network layer (2) has a length difference between two adjacent waveguides, each The arrayed waveguides between the sub-network layers have a length difference between two adjacent waveguides; the arrayed waveguides between the network layer (2) and the output layer (3) have a length difference between two adjacent waveguides; The length difference is greater than or equal to 0, which is specifically determined according to the matrix expression of the light diffraction free transmission zone combined with the calculation iteration convergence.
在光波导和/或光芯片上实现的全光衍射神经网络***，其特征在于：由权利要求1-12任一所述的全光衍射神经网络任意组合构成。The all-optical diffraction neural network system implemented on the optical waveguide and/or the optical chip is characterized in that it is composed of any combination of the all-optical diffraction neural network according to any one of claims 1-12.