Optical neural network device

Abstract

The utility model discloses an optical neural network device which comprises a light-emitting part, a detection part and an optical path assembly, a Fourier transform part and an inverse Fourier transform part of the optical path assembly are located on the same side of an optical signal transmission area, and the light-emitting part, a mask part and the detection part are located on the other side of the optical signal transmission area. According to the optical neural network device, the optical path is folded, so that the occupied space volume of the optical neural network device is greatly reduced compared with an existing linear optical neural network, and the distance from optical signal emission to optical signal receiving is shorter correspondingly, so that the electric signal transmission speed is higher, the loss is less, and the transmission efficiency is improved. Therefore, the optical neural network has a better application prospect, and the progress of the optical neural network technology is promoted.

Description

TECHNICAL FIELD

[0001] The utility model relates to the field of light calculation, especially a light neural network device. BACKGROUND

[0002] The light neural network built by optical elements can partially replace the electric signal processing unit to carry out convolution processing and extract the key features of the target, and has faster operation speed and higher energy efficiency ratio compared with the processing of electric signals, but the existing light neural network still has some problems in this scene, the most important of which is that the volume of the optical element is too large compared with the electronic element, and the long space distance needs to be occupied between the emission and reception of the light signal due to the limitation of the focal length of the light path, which further leads to more electric signal loss in the time consumption of the transmission of the electric signal at a farther distance, so the existing light neural network built by optical elements cannot well meet the use requirement and restricts the development of the light neural network. SUMMARY

[0003] To solve at least one of the problems of the volume of the light neural network built by optical elements being too large, the time consumption and loss of the electric signal transmission, the purpose of the present application is to provide a light neural network device with smaller volume and higher integration 。

[0004] To achieve the above-mentioned utility model purpose, an embodiment of the utility model provides a light neural network device, comprising:

[0005] A light emitting part for emitting light signals;

[0006] A detection part for receiving light signals;

[0007] A light path assembly comprising a Fourier transform part, an inverse Fourier transform part and a mask part, the Fourier transform part, the inverse Fourier transform part and the mask part are all set as reflective optical elements, the light signals emitted by the light emitting part pass through the Fourier transform part, the mask part and the inverse Fourier transform part in sequence and then reach the detection part, the Fourier transform part and the inverse Fourier transform part are located on the same side of the light signal transmission area, and the light emitting part, the mask part and the detection part are located on the other side of the light signal transmission area.

[0008] As a further improvement of the utility model, the light neural network device further comprises an integrated chip, and the light emitting part, the mask part and the detection part are all integrated on the same integrated chip.

[0009] As a further improvement of the utility model, the object plane corresponding to the light emitting part, the reflecting surface of the mask part and the image plane corresponding to the detection part are located on the same plane.

[0010] As a further improvement of the utility model, at least one of the light emitting part, the mask part and the detection part is provided with an extinction light trap on the side facing the optical signal transmission area, which is used to eliminate other stray light on the target light path from the light emitting part, the Fourier transform part, the mask part, the inverse Fourier transform part to the detection part.

[0011] As a further improvement of the utility model, the reflecting surface of the mask part and the object surface corresponding to the light emitting part are not in the same plane.

[0012] And / or,

[0013] The reflecting surface of the mask part and the image surface corresponding to the detection part are not in the same plane.

[0014] As a further improvement of the utility model, the plane perpendicular to the center of the mask part and located between the incident light and the reflected light of the mask part is a symmetry plane, the Fourier transform part and the inverse Fourier transform part are symmetrically arranged on both sides of the symmetry plane, and the light emitting part and the detection part are symmetrically arranged on both sides of the symmetry plane.

[0015] As a further improvement of the utility model, the Fourier transform part and the inverse Fourier transform part are respectively arranged as a concave reflecting type optical element or a super surface reflecting type optical element, and the focal lengths of the Fourier transform part and the inverse Fourier transform part are the same.

[0016] As a further improvement of the utility model, the Fourier transform part converts the spatial domain optical signal into a frequency domain optical signal, the mask part is arranged as a super surface reflecting type optical element, a plurality of base elements of nano structure are arranged on the surface of the mask part, the amplitude of the complex amplitude of the frequency domain optical signal is adjusted by adjusting the size and spatial distribution of the base elements.

[0017] As a further improvement of the utility model, the optical neural network device further comprises a processor, the processor outputs the light emitting signal of the light emitting part and receives the optical signal of the detection part, the mask part is a silicon-based liquid crystal element, the optical neural network device comprises a regulating component, the processor outputs a voltage signal to the regulating component according to the received optical signal, the regulating component applies a controllable voltage to the silicon-based liquid crystal element according to the voltage signal, and the controllable voltage is used to control the spatial distribution of the liquid crystal in the silicon-based liquid crystal element to change the modulation result of the mask part on the optical signal.

[0018] As a further improvement of the utility model, the light emitting part is arranged as a high frequency light source, the detection part is arranged as a high detection frequency detector, and the switching frequency of the controllable voltage applied by the regulating component is consistent with the frequency of the light signal emitted by the light emitting part.

[0019] Compared with the prior art, the optical neural network device has the following beneficial effects: the optical neural network device is folded by arranging the light emitting part, the Fourier transform part, the mask part, the inverse Fourier transform part and the detection part on both sides of the optical signal transmission area, so that the space volume occupied by the optical neural network device is greatly reduced, the distance required from the emission of the optical signal to the reception of the optical signal is shorter, the speed of the electrical signal transmission is faster, and the loss is smaller, so that the optical neural network has better application prospect, and the progress of the optical neural network technology is promoted. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a structural schematic view of one embodiment of the optical neural network device of the first embodiment of the utility model;

[0021] Figure 2 is a structural schematic view of another embodiment of the optical neural network device of the first embodiment of the utility model;

[0022] Figure 3 is a structural schematic view of still another embodiment of the optical neural network device of the first embodiment of the utility model;

[0023] Figure 4 is a structural schematic view of still another embodiment of the optical neural network device of the first embodiment of the utility model;

[0024] Figure 5 is a schematic view of the mask part changing the spatial distribution of liquid crystals under the control of the regulating assembly of the second embodiment of the utility model;

[0025] Wherein, 10, light emitting part;20, detection part;30, optical path assembly;31, Fourier transform part;32, mask part;33, inverse Fourier transform part;40, extinction optical trap;50, step part;91, super surface reflection type optical element;92, concave reflection type optical element. DETAILED DESCRIPTION

[0026] The utility model will be described in detail below in combination with the specific embodiments shown in the drawings. However, these embodiments do not limit the utility model, and the changes in structure, method or function made by those skilled in the art according to these embodiments are all included in the protection scope of the utility model.

[0027] It should be understood that the terms for indicating spatial relative positions used herein, such as "upper", "above", "lower", "below", etc. are used for the purpose of convenient description to describe the relationship of one unit or feature relative to another unit or feature as shown in the drawings. The terms for indicating spatial relative positions can be intended to include different positions of the device in use or work other than the positions shown in the drawings.

[0028] The utility model discloses an embodiment provides a kind of optical neural network device with smaller volume, higher integration.

[0029] As Figures 1-4 As shown in the drawing, the optical neural network device includes a light emitting part 10, a detection part 20, and a light path assembly 30. The light emitting part 10 is used to emit optical signals, the detection part 20 is used to receive optical signals, and the light path assembly 30 can modulate optical signals according to processing tasks, extract feature parameters of input images, realize parallel analog computation, and change the propagation direction of optical signals. In application, the optical neural network device can be used as an optical convolution layer of a hybrid neural network to perform convolution operations on input images and complete image recognition, image classification, and other tasks.

[0030] The light path assembly 30 includes a Fourier transform part 31, a mask part 32, and an inverse Fourier transform part 33. The optical signals emitted by the light emitting part 10 pass through the Fourier transform part 31, the mask part 32, and the inverse Fourier transform part 33 in sequence and then reach the detection part 20. The Fourier transform part 31 and the inverse Fourier transform part 33 can focus light in a three-dimensional space to realize Fourier transform of incident optical signals. More specifically, the Fourier transform part 31 can convert optical image signals in the spatial domain into spatial frequency signals in the frequency domain, the mask part 32 can adjust the complex amplitude distribution of the spatial frequency signals in the frequency domain, and the inverse Fourier transform part 33 can convert the spatial frequency signals in the frequency domain back into optical image signals in the spatial domain.

[0031] In this embodiment, the Fourier transform part 31, the inverse Fourier transform part 33, and / or the mask part 32 are super lenses, metasurface devices, diffractive optical elements (DOEs), or curved lenses. Super lenses can realize miniaturization and planarization of optical elements while retaining their optical functions. By specially designing the geometry and arrangement of nanoscale microstructure units in the super lens, the amplitude and phase of input optical signals can be simultaneously controlled.

[0032] For the mask part 32, the geometry of nanoscale structures at different positions on the super lens, such as height, period, arrangement, radius, and geometric shape, can be flexibly controlled to simultaneously control the amplitude and phase of the complex amplitude signals in the Fourier plane. The required convolution kernel, i.e., the complex amplitude distribution in the frequency domain after Fourier transform, can be calculated to determine the size and spatial distribution of the nanoscale structures of the mask part 32.

[0033] The following three embodiments further illustrate various implementations of the optical neural network device. Embodiment 1 is an embodiment of various structures of the optical path component 30, Embodiment 2 is one embodiment of the mask part 32, and Embodiment 3 is one embodiment of the light-emitting part 10 and the detector part 20. More specifically, the main differences between the embodiments are as follows: Embodiment 1 shows that the Fourier transform part 31 and the inverse Fourier transform part 33 adopt a surface reflection type optical element structure, a concave reflection type optical element 92 structure, add an extinction light trap 40, and change the plane on which the mask part 32 is located; Embodiment 2 shows a mask part 32 that can be used in multiple task scenarios, which is equivalent to a variable convolution kernel in a convolution calculation optical system; Embodiment 3 shows a high-speed light-emitting part 10 and a high-speed detector part 20, which can be combined with Embodiment 2.

[0034] In addition, to clearly express the positions and directions described in this embodiment, the following embodiments define references. Figures 1-4 In the diagram, the upper part refers to "up" as discussed below, the lower part refers to "down," the left side refers to "left," and the right side refers to "right." For example, in the diagram, the Fourier transform unit 31 and the inverse Fourier transform unit 33 are located on the upper side of the optical signal transmission area, the light-emitting unit 10, the mask unit 32, and the detector unit 20 are located on the lower side of the optical signal transmission area, and the Fourier transform unit 31 is located to the left of the inverse Fourier transform unit 33. The opposite direction is the right side. Here, "up," "down," "left," and "right" are merely noun definitions and are not limited to the physical meaning of "up" and "down." For example, when the optical path component 30 is placed horizontally on a horizontal plane, the up and down directions of the light-emitting unit 10, the detector unit 20, and the optical path component 30 can be located on the same plane. Furthermore, "up" and "down" can be interchanged, and "left" and "right" can be interchanged. For example, when the optical path component 30 is placed horizontally on a horizontal plane, the up and down directions of the light-emitting unit 10, the detector unit 20, and the optical path component 30 can be located on the same plane. Figure 1 Reversing, mirroring horizontally or vertically, or rotating at any angle will not change its specific function.

[0035] Example 1

[0036] This implementation example Figure 1 As shown, the Fourier transform unit 31, the inverse Fourier transform unit 33, and the mask unit 32 are all configured as reflective optical elements. The Fourier transform unit 31 and the inverse Fourier transform unit 33 are located on the upper side of the optical signal transmission area, while the light-emitting unit 10, the mask unit 32, and the detector unit 20 are located on the lower side of the optical signal transmission area.

[0037] The light path is folded multiple times between the light emitting part 10, the Fourier transform part 31, the mask part 32, the inverse Fourier transform part 33 and the detection part 20 during the light signal is emitted from the light emitting part 10 to the detection part 20, and the propagation direction of the light path is changed three times, so the distance of the whole structure in the left-right direction or the distance in the up-down direction is less than the distance between the light emitting part 10 and the detection part 20 when the light propagates along a straight line, so compared with the prior art, the further folding of the light path is smaller in length and higher in integration to achieve the same light path transmission distance.

[0038] Figure 1 In the embodiment, the light path of the light emitting part 10, the Fourier transform part 31, the mask part 32, the inverse Fourier transform part 33 and the detection part 20 in sequence changes direction in the up-down direction all the time, but always transmits to the right, that is, the light path originally transmitted to the oblique upper direction is folded between the light elements, and finally forms a shape similar to the letter M, which is high in folding efficiency and greatly reduces the overall volume.

[0039] In addition, since the light path is folded, the distance between the light emitting part 10 and the detection part 20 is shorter than that of the prior art, and the distance required from the emission of the light signal to the reception of the light signal is shorter, so the speed of the electrical signal transmission is faster and the loss is less.

[0040] Further, the light emitting part 10, the detection part 20 and the mask part 32 of the embodiment are all located on an integrated chip, which not only further improves the integration and effectively reduces the overall size of the system, but also increases the processing speed of the optical computing system.

[0041] Further, the object plane corresponding to the light emitting part 10, the reflecting surface of the mask part 32 and the image plane corresponding to the detection part 20 are located on the same plane.

[0042] Such arrangement can further reduce the size of the optical neural network device by using the planar structure, so the integration of the whole system is very high, and the thickness of the whole system can be controlled to be very thin. In addition, the Fourier transform part 31, the mask part 32 and the inverse Fourier transform part 33 can all be arranged as super lenses, and the thickness of the whole system can be controlled to be in the order of millimeters.

[0043] and continue to refer to Figure 1 As shown, the plane perpendicular to the center of the mask part 32 and located between the incident light and the reflected light of the mask part 32 is defined as the symmetry plane S1, the Fourier transform part 31 and the inverse Fourier transform part 33 are symmetrically arranged on both sides of the symmetry plane, and the light emitting part 10 and the detection part 20 are symmetrically arranged on both sides of the symmetry plane, and correspondingly, the light path from the light emitting part 10 to the Fourier transform part 31 and then to the mask part 32 is symmetrically arranged on both sides of the symmetry plane S1 with the light path from the mask part 32 to the inverse Fourier transform part 33 and then to the detection part 20.

[0044] Further, as shown in Figures 1-4 The Fourier transform unit 31 and the inverse Fourier transform unit 33 are respectively arranged as a concave reflective optical element 92 or a metasurface reflective optical element 91, and the focal lengths of the Fourier transform unit 31 and the inverse Fourier transform unit 33 are the same. Figure 1 In the embodiment shown in FIG. 2, the Fourier transform unit 31 and the inverse Fourier transform unit 33 are both arranged as metasurface reflective optical elements 91, Figure 2 In the embodiment shown in FIG. 3, the Fourier transform unit 31 and the inverse Fourier transform unit 33 are both arranged as concave reflective optical elements 92, Figure 3 In the embodiment shown in FIG. 4, the Fourier transform unit 31 is arranged as a metasurface reflective optical element 91, and the inverse Fourier transform unit 33 is arranged as a concave reflective optical element 92, Figure 4 In the embodiment shown in FIG. 5, the Fourier transform unit 31 is arranged as a concave reflective optical element 92, and the inverse Fourier transform unit 33 is arranged as a metasurface reflective optical element 91. In addition, the mask unit can be a metasurface optical element.

[0045] Figures 1-4 The light emitting unit 10, the Fourier transform unit 31, the mask unit 32, the inverse Fourier transform unit 33, and the detection unit 20 are all shown in the same plane (the plane in which the up, down, left, and right directions are located, which is perpendicular to the front and back directions). When at least part of these optical elements uses a metasurface optical element, off-axis focusing can also be achieved, that is, these optical elements can also not all be in the same plane (the plane in which the up, down, left, and right directions are located), that is, Figures 1-4 Only a special case is shown in which all these elements are in the same plane, or the coordinates of each optical element in the front and back directions are not all the same, and the projection is shown as falling in the same plane.

[0046] The concave reflective optical element 92 is simple in structure, and only needs to have a focal length that meets the specific design requirements. The metasurface reflective optical element 91 is complex in structure and needs to reach the expected focusing position.

[0047] Referring again to Figure 3 As shown in FIG. 6, at least one of the light emitting unit 10, the mask unit 32, and the detection unit 20 is provided with an extinction light trap 40 on the side facing the optical signal transmission area, and the extinction light trap 40 is used to eliminate other stray light on the target light path from the light emitting unit 10, the Fourier transform unit 31, the mask unit 32, the inverse Fourier transform unit 33, to the detection unit 20.

[0048] The extinction light trap 40 is used to avoid crosstalk caused by stray light signals. In the embodiment, as shown in Figure 3 The extinction light trap 40 is arranged above the light emitting unit 10, the mask unit 32, and the detection unit 20, and only allows Figure 3 the light of the target to pass through, reducing the interference of some other reflected stray light on the optical signal.

[0049] Figure 4 Also for avoiding stray light from interfering with the light signal, and Figure 3 The difference lies in that: Figure 4 The reflecting surface of the mask part 32 does not lie in the same plane as the object surface corresponding to the light emitting part 10.

[0050] And / or,

[0051] The reflecting surface of the mask part 32 does not lie in the same plane as the image surface corresponding to the detection part 20.

[0052] As shown in Figure 4 A step part 50 is arranged below the mask part 32, which raises the mask part 32, so that the reflecting surface of the mask part 32 is closer to the Fourier transform part 31 and the inverse Fourier transform part 33 than the object surface corresponding to the light emitting part 10 and the image surface corresponding to the detection part 20.

[0053] In other embodiments, the step part 50 can also make the mask part 32 lower, that is, the reflecting surface of the mask part 32 is farther away from the Fourier transform part 31 and the inverse Fourier transform part 33 than the object surface corresponding to the light emitting part 10 and the image surface corresponding to the detection part 20.

[0054] Since the reflecting surface of the mask part 32 does not lie in the same plane as the object surface corresponding to the light emitting part 10 and the image surface corresponding to the detection part 20, the mask part 32 is not easily affected by stray light from the light emitting part 10 and the detection part 20, reducing the interference of other reflected stray light with the light signal.

[0055] Embodiment 2

[0056] The difference between this embodiment 2 and embodiment 1 is that the mask part in embodiment 1 can be an existing metasurface or a mask plate made of an SLM (Spatial Light Modulator). Embodiment 1 can only be specially designed for a single task in a certain scene and can only process a single task. Embodiment 2 can be used for mask part 32 in various task scenarios, which is equivalent to a variable convolution kernel in a convolutional optical system.

[0057] Specifically, the optical neural network device further comprises a processor, which outputs a light emitting signal of the light emitting part 10 and receives a light signal of the detection part 20, and the mask part 32 is a silicon-based liquid crystal component. The optical neural network device comprises a regulation component, and the processor outputs a voltage signal to the regulation component according to the received light signal. The regulation component applies a controllable voltage to the silicon-based liquid crystal component according to the voltage signal, and the controllable voltage is used to control the spatial distribution of liquid crystals in the silicon-based liquid crystal component to change the modulation result of the mask part 32 on the light signal.

[0058] The mask part 32 is manufactured by using a liquid crystal on silicon (LCoS) technology, can realize the modulation of the spatial distribution of the amplitude and phase of the light signal, and controls the spatial distribution of the liquid crystal in the liquid crystal layer by addressing the voltage applied to the mask plate, thereby realizing different regulation effects of the spatial distribution of the amplitude and phase of the light signal, that is, realizing the programmable complex amplitude regulation capability, corresponding to the variable convolution kernel in the entire convolution calculation optical neural network. The mask part 32 can also use other materials such as electro-optic polymer, chalcogenide glass and other materials with adjustable refractive index to regulate the light signal. The following mask part 32 takes the material using the liquid crystal on silicon technology as an example for description.

[0059] More specifically, the regulation assembly loads a voltage on the liquid crystal layer of the mask part 32 at each pixel, so that the liquid crystal molecules in the corresponding position of the liquid crystal layer are rotated in space. After the rotation of the liquid crystal molecules, the polarization state of the incident light can be changed by using the birefringence effect of the liquid crystal molecules to realize the amplitude regulation of the outgoing light. When the liquid crystal molecules are rotated in space, the equivalent refractive index of the liquid crystal layer changes, thereby realizing the phase adjustment of the incident light.

[0060] Therefore, by adjusting the voltage applied to the liquid crystal layer of the mask part 32, the phase and amplitude of the incident light can be simultaneously regulated, and by changing the size of the added voltage, different regulation effects can be realized, that is, the programmable complex amplitude regulation capability of the incident light is realized, so that the convolution kernel can be iterated continuously by programming to construct the mask part 32 meeting the specific task requirements.

[0061] Since in the process of running the entire system, especially in the embodiment 2, the mask part 32 needs to be adjusted according to the task requirements, the input image needs to be adjusted and the mask part 32 needs to be adjusted, the image processing effect depends on the comparison of the input image and the output image, the mask part 32 convolution kernel effect is judged according to the processing effect, and then it is adjusted, the above steps are repeatedly iterated, multiple convolution calculations and optimal convolution kernel selection are realized. The iteration speed of this feedback process depends largely on the comparison speed between the input image and the output image. Therefore, the spatial distance between the light emitting part 10 and the detection part 20 is shortened as much as possible, which can most directly improve the comparison speed between the input image and the output image, the signal response speed is fast, and then the feedback adjustment process of programming the complex amplitude adjustment mask plate according to the comparison result of the input and output images is accelerated, and finally the mask part 32 quickly iterates to the target convolution kernel according to the current task.

[0062] Embodiment 3

[0063] The difference between this embodiment 3 and embodiments 1-2 is that the light emitting part 10 of embodiments 1-2 can be a common light source, and the detection part 20 can be a common CMOS, for example, the detection frequency of embodiments 1-2 is in the order of kilohertz. While the light emitting part 10 of this embodiment 3 is set as a high-frequency light source, and the detection part 20 is set as a high-detection-frequency detector, the frequency corresponding to the high frequency is in the order of megahertz.

[0064] Specifically, the light emitting part 10 can be composed of parallel light sources, such as laser light sources combined with waveguide arrays, or laser light sources combined with digital microscopes, or laser light sources combined with liquid crystal spatial light modulators, to generate image signals and emit parallel light signals to the Fourier transform part 31. Among them, the laser light source combined with the waveguide array outputs a high-frequency signal, and the detection part 20 uses a photodiode array as a detector to realize high-frequency detection of the signal.

[0065] And because embodiments 1-2 set the light emitting part 10 and the detection part 20 relatively close to each other, the sum of the distances from the integrated chip to the light emitting part 10 and the detection part 20 is relatively smaller than that of the prior art, so the delay of the electrical signal is relatively smaller, and then when using a high-frequency light source and a high-detection-frequency detector, the response speed is faster, that is, on the basis of the light emitting part 10 and the detection part 20 of embodiments 1-2 being set relatively close to each other, the embodiment 3 has greater realizability.

[0066] In particular, when this embodiment 3 is combined with embodiment 2, it will also produce the following technical effects: embodiment 2 has higher efficiency when performing batch processing of convolution operations of multiple groups of different input image signals and different convolution kernels at the same time, and can switch the complex amplitudes of the image light signals of the input convolution calculation optical system at high speed.

[0067] At this time, the switching frequency of the controllable voltage applied by the control assembly is consistent with the frequency of the light signal emitted by the light emitting part 10, and the convolution operation of the input image signal and the different convolution kernels can be realized at a higher frequency, reducing the time required for iterative convolution kernels.

[0068] Compared with the common technology, the embodiment has the following beneficial effects:

[0069] The optical neural network device designs the structure of the optical path assembly 30, sets the light emitting part 10, the Fourier transform part 31, the mask part 32, the inverse Fourier transform part 33, and the detection part 20 on both sides of the light signal transmission area, which is equivalent to folding the optical path, so that the space volume occupied by the optical neural network device relative to the existing linear optical neural network is greatly reduced, and the distance required from the emission of the light signal to the reception of the light signal is shorter, so the speed of the electrical signal transmission is faster and the loss is less. Therefore, the optical neural network has better application prospects and promotes the progress of optical neural network technology.

[0070] It should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand.

[0071] The above series of detailed descriptions are only specific descriptions of the feasible embodiments of the present application, and are not intended to limit the protection scope of the present application. Any equivalent embodiments or changes made without departing from the spirit of the present application should be included in the protection scope of the present application.

Claims

1. An optical neural network device, comprising: The light-emitting part is used for emitting light signals. The light path assembly includes a Fourier transform part, an inverse Fourier transform part and a mask part, all of which are set as reflective optical elements, the light signals emitted by the light-emitting part pass through the Fourier transform part, the mask part and the inverse Fourier transform part in sequence to reach the detection part, the Fourier transform part and the inverse Fourier transform part are located on the same side of the light signal transmission area, and the light-emitting part, the mask part and the detection part are located on the other side of the light signal transmission area. The light-emitting part, the mask part and the detection part are integrated on the same integrated chip. The object plane corresponding to the light-emitting part, the reflecting surface of the mask part and the image plane corresponding to the detection part are located on the same plane.

2. The optical neural network device of claim 1, wherein, At least one of the light-emitting part, the mask part and the detection part is provided with an extinction optical trap on the side of the light signal transmission area, and the extinction optical trap is used to eliminate other stray light on the target light path passing through the light-emitting part, the Fourier transform part, the mask part, the inverse Fourier transform part and the detection part.

3. The optical neural network device of claim 2, wherein, The reflecting surface of the mask part and the object plane corresponding to the light-emitting part are not on the same plane.

4. The optical neural network device of claim 3, wherein, And / or, 5. The optical neural network device of claim 2, wherein, The reflecting surface of the mask part and the image plane corresponding to the detection part are not on the same plane. The plane perpendicular to the center of the mask part and located between the incident light and the reflected light of the mask part is a symmetry plane, the Fourier transform part and the inverse Fourier transform part are symmetrically arranged on both sides of the symmetry plane, and the light-emitting part and the detection part are symmetrically arranged on both sides of the symmetry plane. The Fourier transform part and the inverse Fourier transform part are respectively set as concave reflective optical elements or super surface reflective optical elements, and the focal lengths of the Fourier transform part and the inverse Fourier transform part are the same.

6. The optical neural network device of claim 1, wherein, The Fourier transform part converts the spatial domain light signal into a frequency domain light signal, the mask part is set as a super surface reflective optical element, the mask part surface is provided with a plurality of base elements of nano structure, the amplitude of the complex amplitude of the frequency domain light signal is adjusted by adjusting the size and spatial distribution of the base elements.

7. The optical neural network device of claim 6, wherein, The light neural network device further includes a processor, the processor outputs light-emitting signals to the light-emitting part and receives light signals from the detection part, the mask part is a silicon-based liquid crystal element, the light neural network device includes a regulating assembly, the processor outputs a voltage signal to the regulating assembly according to the received light signals, the regulating assembly applies a controllable voltage to the silicon-based liquid crystal element according to the voltage signal, and the controllable voltage is used to control the spatial distribution of the liquid crystal in the silicon-based liquid crystal element to change the modulation result of the mask part on the light signal.

8. The optical neural network device of claim 1, wherein, The light-emitting part is set as a high-frequency light source, the detection part is set as a high-detection-frequency detector, and the switching frequency of the controllable voltage applied by the regulating assembly is consistent with the frequency of the light signals emitted by the light-emitting part.

9. The optical neural network apparatus according to claim 1 or 8, wherein, ​ 10. The optical neural network device of claim 9, wherein, ​

Images