CN117764131A

CN117764131A - Optical neural network device

Info

Publication number: CN117764131A
Application number: CN202311804143.3A
Authority: CN
Inventors: 孙磊; 韩旭; 史坦; 范鸿杰; 刘思扬
Original assignee: Suzhou Shanhe Photoelectric Technology Co ltd
Current assignee: Suzhou Shanhe Photoelectric Technology Co ltd
Priority date: 2023-12-26
Filing date: 2023-12-26
Publication date: 2024-03-26

Abstract

The invention discloses an optical neural network device, which comprises a light emitting part, a detecting part and an optical path component, wherein a Fourier transform part, an inverse Fourier transform part and a mask part of the optical path component are all arranged as reflection type optical elements, the Fourier transform part and the inverse Fourier transform part are positioned on the upper side of an optical signal transmission area, the light emitting part, the mask part and the detecting part are positioned on the lower side of the optical signal transmission area, the inverse Fourier transform part is used for emitting an optical signal to the detecting part positioned on the lower left side of the inverse Fourier transform part, and the light emitting part and the detecting part are both arranged on the left side of the mask part. The luminous part and the detection part of the optical neural network device can be arranged as close as possible, so that the distance from the chip to the chip outputting the luminous signal to the chip receiving the optical signal is shorter, the time interval is smaller, and the space volume is greatly reduced, therefore, the optical neural network device has faster processing speed while reducing the volume, and has better application prospect.

Description

Optical neural network device

Technical Field

The invention relates to the field of optical computing, in particular to an optical neural network device.

Background

The optical neural network built by the optical element can partially replace an electric signal processing unit to carry out convolution processing, the key characteristics of a target are extracted, compared with the processing of electric signals, the optical neural network has higher operation speed and higher energy efficiency ratio, but the application of the existing optical neural network in the scene has some problems, wherein the optical neural network built by the existing optical element cannot well meet the use requirement due to the fact that compared with an electronic element, the volume of the optical element is overlarge, the limitation of the focal length of an optical path is combined, a longer space distance is required to be occupied between sending and receiving of the optical signals, the longer distance means that a longer time interval exists between outputting of the luminous signals from a chip and receiving of the optical signals from the chip, obstacles exist when high-frequency information is processed, and more electric signal loss exists when the electric signals are transmitted at a longer distance, so that the development of the optical neural network built by the existing optical neural network is limited.

Disclosure of Invention

In order to solve the problems that an optical neural network built by optical elements in the prior art is overlarge in size and has a longer time interval between outputting a luminous signal from a chip and receiving the optical signal from the chip, the invention aims to provide an optical neural network device with smaller size and higher signal processing speed of the chip.

To achieve the above object, an embodiment of the present invention provides an optical neural network device, including:

a light emitting section for emitting an optical signal;

a detection section for receiving an optical signal;

the optical path component comprises a Fourier transform part, an inverse Fourier transform part and a mask part, wherein the Fourier transform part, the inverse Fourier transform part and the mask part are all arranged as reflection type optical elements, an optical signal emitted by the light emitting part sequentially passes through the Fourier transform part, the mask part and the inverse Fourier transform part and then reaches the detection part, the Fourier transform part and the inverse Fourier transform part are positioned on the upper side of an optical signal transmission area, the light emitting part, the mask part and the detection part are positioned on the lower side of the optical signal transmission area, the inverse Fourier transform part is used for emitting the optical signal to the detection part positioned on the lower left side of the inverse Fourier transform part, and the light emitting part and the detection part are both arranged on the left side of the mask part.

As a further improvement of the present invention, the optical neural network device further includes an integrated chip, and the light emitting portion and the detecting portion are both integrated on the same integrated chip.

As a further improvement of the present invention, the inverse fourier transform section is provided as a super-surface reflection type optical element, and the surface of the inverse fourier transform section is provided with a plurality of elements of a nanostructure, and the distribution of the elements conforms to a design formulaWherein (X, Y) is a position coordinate on the inverse Fourier transform unit (X) _f ,Y _f ) Is the coordinates of the focus after focusing.

As a further improvement of the present invention, the incident light entering the mask portion and the reflected light reflected by the mask portion are both located on the left side of the mask portion.

As a further improvement of the invention, the design formula of the beam deflection corresponding to each position in the X direction on the mask part is phi _mask (X)＝-2π/λ[X×(sinθ _I -sinθ _i )]Wherein X is the position on the X axis of the mask portion, θ _I Angle of reflected light, θ _i Is the angle of the incident light.

As a further improvement of the present invention, the object plane corresponding to the light emitting portion, the reflection plane of the mask portion, and the image plane corresponding to the detecting portion are located on the same plane.

As a further improvement of the present invention, an extinction optical trap is provided on a side of at least one of the light emitting section, the mask section, and the detecting section toward the optical signal transmission region, the extinction optical trap being for eliminating other stray light on a target optical path to the detecting section via the light emitting section, the fourier transform section, the mask section, and the inverse fourier transform section.

As a further improvement of the present invention, the reflecting surface of the mask portion and the object surface corresponding to the light emitting portion are not in the same plane;

and/or the number of the groups of groups,

the reflection surface of the mask part and the image surface corresponding to the detection part are not in the same plane.

As a further improvement of the present invention, the fourier transform section is provided as a concave reflection type optical element or a super surface reflection type optical element.

As a further improvement of the invention, the fourier transform part converts the spatial domain optical signal into the frequency domain optical signal, the mask part is arranged as a super-surface reflection type optical element, the surface of the mask part is provided with a plurality of nano-structured elements, and the amplitude of the complex amplitude of the frequency domain optical signal is regulated and controlled by adjusting the size and the spatial distribution of the elements.

As a further improvement of the present invention, the optical neural network device further includes a processor, the processor outputs a light emitting signal to the light emitting part and receives an optical signal of the detecting part, the mask part is a liquid crystal on silicon part, the mask part is also integrated on the integrated chip, the optical neural network device includes 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 liquid crystal on silicon part according to the voltage signal, and the controllable voltage is used for controlling a spatial distribution of liquid crystal in the liquid crystal on silicon part so as to change a modulation result of the optical signal by the mask part.

As a further improvement of the present invention, the light emitting part is configured as a high-frequency light source, the detecting part is configured as a high-detection frequency detector, and the switching frequency of the controllable voltage applied by the regulating and controlling component is consistent with the frequency of the light signal emitted by the light emitting part.

Compared with the prior art, the invention has the following beneficial effects: according to the optical neural network device, through the design of the optical path component structure and the adjustment of the positions and the light reflecting directions of the Fourier transform part, the Fourier transform part and the mask part, other parts can be omitted between the light emitting part and the detection part, namely, the light emitting part and the detection part can be arranged as close as possible, further, the distance from the chip to the chip for outputting the light emitting signal to the chip for receiving the light signal is shorter, the time interval is smaller, and the optical path is folded, so that the occupied space volume of the optical neural network device is greatly reduced relative to the existing linear optical neural network, and the optical neural network device has higher processing speed and better application prospect when the volume is reduced.

Drawings

Fig. 1 is a schematic structural view of an optical neural network device according to a first embodiment of the present invention;

fig. 2 is a schematic structural diagram of an optical neural network device according to a second embodiment of the present invention;

fig. 3 is a schematic structural diagram of an optical neural network device according to a third embodiment of the present invention;

fig. 4 is a schematic structural diagram of an optical neural network device according to a fourth embodiment of the present invention;

fig. 5 is a schematic structural diagram of an optical neural network device according to a fifth embodiment of the present invention;

FIG. 6 is a schematic view showing a mask portion changing the spatial distribution of liquid crystal under the control of a regulating member according to a sixth embodiment of the present invention;

wherein, 10, the light-emitting part; 20. a detection unit; 30. an optical path component; 31. a Fourier transform unit; 32. a mask portion; 33. an inverse fourier transform unit; 40. extinction optical traps; 50. a step portion; 91. a super surface reflective optical element; 92. concave reflective optical element.

Detailed Description

The present invention will be described in detail below with reference to specific embodiments shown in the drawings. These embodiments are not intended to limit the invention and structural, methodological, or functional modifications of these embodiments that may be made by one of ordinary skill in the art are included within the scope of the invention.

It will be appreciated that terms such as "upper," "above," "lower," "below," and the like, as used herein, refer to spatially relative positions and are used for ease of description to describe one element or feature's relationship to another element or feature as illustrated in the figures. The term spatially relative position may be intended to encompass different orientations of the device in use or operation in addition to the orientation depicted in the figures.

An embodiment of the invention provides an optical neural network device with smaller volume and higher signal processing speed of a chip.

As shown in fig. 1 to 5, the optical neural network device includes a light emitting portion 10, a detecting portion 20 and an optical path component 30, wherein the light emitting portion 10 is used for emitting an optical signal, the detecting portion 20 is used for receiving the optical signal, the optical path component 30 can modulate the optical signal according to a processing task, extract characteristic parameters of an input image, implement parallel analog calculation, and change a propagation direction of the optical signal. In application, the optical neural network device can be used as an optical convolution layer of a hybrid neural network to carry out convolution operation on an input image so as to finish tasks such as image recognition, image classification and the like.

The optical path module 30 includes a fourier transform unit 31, a mask unit 32, and an inverse fourier transform unit 33, and the optical signal emitted from the light emitting unit 10 reaches the detecting unit 20 after passing through the fourier transform unit 31, the mask unit 32, and the inverse fourier transform unit 33 in this order. Wherein the fourier transform unit 31 and the inverse fourier transform unit 33 can realize focusing of light in a three-dimensional space and fourier transform of an incident light signal. More specifically, the fourier transform unit 31 may convert the optical image signal in the spatial domain into a spatial frequency signal in the frequency domain, the mask unit 32 may adjust the complex amplitude distribution of the spatial frequency signal in the frequency domain, and the inverse fourier transform unit 33 may convert the spatial frequency signal in the frequency domain back into the optical image signal in the spatial domain.

The fourier transform section 31, and/or the inverse fourier transform section 33, and/or the mask section 32 in this embodiment are provided as a superlens, a supersurface device, a diffractive optical element (Diffractive Optical Elements, DOE), or a curved lens, which can achieve miniaturization and planarization of the optical element while retaining its optical function, and can regulate and control the amplitude and phase of an input optical signal simultaneously by specially designing the geometry and arrangement of nanoscale microstructure elements in the superlens.

For the mask portion 32, the geometric dimensions of the nanostructures at different positions on the superlens, such as height, period, arrangement, radius, geometry, and other parameters, can be flexibly adjusted and controlled, so as to realize simultaneous adjustment and control of the amplitude and phase of the complex amplitude signal in the optical frequency domain in the fourier plane. The nanostructure size and spatial distribution of the mask portion 32 can be determined by calculating the required convolution kernel, i.e., the complex amplitude distribution in the frequency domain via fourier transform.

Various implementations of the optical neural network device will be further described below in terms of 7 embodiments, wherein embodiments 1 to 5 are embodiments of various structures of the optical path component 30, embodiment 6 is one of the mask portions 32, and embodiment 7 is one of the light emitting portion 10 and the detecting portion 20. More specifically, the main differences between the embodiments are: the mask section 32 of embodiment 1 can change the reflection angle of the reflected light while realizing the complex amplitude adjustment of the optical signal in the frequency domain, and the inverse fourier transform section 33 can control the position of the focused focal point; the mask unit 32 of embodiment 2 is implemented in such a manner that only complex amplitude adjustment of the optical signal in the frequency domain is required, and the position of the focused focal point can be controlled by the inverse fourier transform unit 33; example 3 adds an extinction optical trap 40; example 4 changed the plane of the mask portion 32; example 5 shows further embodiments of the fourier transform section 31; embodiment 6 shows a mask section 32 that can be used for a variety of task scenarios, equivalent to a variable convolution kernel in a convolution computing optical system; embodiment 7 shows a light emitting portion 10 that emits light at a high speed and a detecting portion 20 that detects light at a high speed, and can be incorporated in embodiment 6.

In order to clearly express the position and direction described in the present embodiment, in the following embodiments, the directions described with reference to fig. 1 to 5 are defined, the upper side in the drawing is the upper side discussed below, the lower side in the drawing is the lower side discussed below, the left side in the drawing is the left side discussed below, the right side in the drawing is the right side discussed below, for example, the fourier transform section 31 and the inverse fourier transform section 33 are located on the upper side of the optical signal transmission region, the light emitting section 10, the mask section 32 and the detecting section 20 are located on the lower side of the optical signal transmission region, the fourier transform section 31 is located on the left side of the inverse fourier transform section 33, the opposite direction is the right side, and in addition, the side perpendicular to the plane in which the upper, lower, left and right are located is the front, and the side is the rear. The vertical and horizontal directions are defined by only one name, and are not limited to "vertical" in a physical sense; for example, when the optical path unit 30 is horizontally placed on a horizontal plane, the light emitting unit 10, the detecting unit 20, and the optical path unit 30 may be vertically placed on the same plane in the physical state, and may be vertically interchanged, or horizontally interchanged, for example, when fig. 1 is inverted, horizontally or vertically mirrored, or rotated at any angle, without changing the specific functions to be realized.

Example 1

In this embodiment, as shown in fig. 1, the fourier transform unit 31, the inverse fourier transform unit 33, and the mask unit 32 are each provided as a reflective optical element, the fourier transform unit 31 and the inverse fourier transform unit 33 are located above the optical signal transmission region, the light emitting unit 10, the mask unit 32, and the detection unit 20 are located below the optical signal transmission region, the inverse fourier transform unit 33 is configured to emit the optical signal to the detection unit 20 located below and to the left of the inverse fourier transform unit 33, and the light emitting unit 10 and the detection unit 20 are each provided to the left of the mask unit 32.

The fourier transform unit 31, the inverse fourier transform unit 33, and the mask unit 32 of the present embodiment are each provided as a reflective optical element, that is, they are each capable of changing the propagation direction of light, and if the optical elements are not specially designed, the incident angle of light is generally equal to the reflection angle, the fourier transform unit 31 and the inverse fourier transform unit 33 of the present embodiment may be provided as a super-surface reflective optical element 91, and the surfaces thereof are provided with a plurality of elements of a nano structure, so that the reflection direction of reflected light can be accurately adjusted.

As shown in fig. 1, the path of light propagation is such that an optical signal emitted from the light emitting unit 10 is emitted to the fourier transform unit 31 in an inclined state, the optical signal is converted into an optical frequency domain signal by the fourier transform unit 31, the optical signal is reflected to the mask unit to the right downward, the mask unit 32 modulates the amplitude of the optical signal in complex amplitude, and then the optical signal is reflected to the inverse fourier transform unit 33, and the inverse fourier transform unit 33 can convert the spatial frequency signal in the frequency domain back into an optical image signal in the spatial domain again, and then the optical signal is emitted to the right downward direction, and the optical signal is emitted to the detection unit 20.

The fourier transform unit 31 in fig. 1 of the present embodiment is disposed at the upper right of the light emitting unit 10, so that the angles between the direction of the incident light and the direction of the reflected light of the fourier transform unit 31 and the vertical plane thereof may be equal or unequal, and if equal, the fourier transform unit is a general reflector; if the fourier transform section 31 is disposed in the upper left of the light emitting section 10 in other embodiments, the direction of the incident light and the direction of the reflected light of the fourier transform section 31 are necessarily not equal to the angle of the perpendicular thereto.

The directions of the incident light and the reflected light of the inverse fourier transform section 33 and the mask section are inevitably not equal to each other in terms of angle with respect to the respective median planes. For example, in fig. 1, the fourier transform unit 31 and the inverse fourier transform unit 33 are both located on the left side of the mask unit, that is, the reflection of light by the mask unit 32 is not a general specular reflection, but a specific reflection direction can be adjusted so that both the incident light entering the mask unit 32 and the reflected light reflected by the mask unit 32 are located on the left side of the mask unit 32.

In embodiment 1, the optical path is folded a plurality of times between 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 during the period from the light emitting unit 10 to the incidence detection unit 20, and the propagation direction of the optical path is changed three times, so that the distance in the horizontal direction or the distance in the vertical direction of the entire structure is smaller than the distance between the light emitting unit 10 and the detection unit 20 when the light propagates along a straight line, so that the same optical path transmission distance is realized in the present embodiment, the further folding of the optical path is made smaller, the length is made smaller, and the integration is made higher than in the prior art.

And, setting up like this makes between luminous portion 10 and the detection portion 20 can not set up other parts, and the two can set up as close as possible promptly, and then makes the chip output luminous signal to the chip receive the distance between the optical signal shorter, time interval is less, and then makes the processing speed faster, has better application prospect.

Further, the optical neural network device further includes an integrated chip, and the light emitting part 10 and the detecting part 20 are integrated on the same integrated chip. Due to the folding of the optical path, the light emitting part 10 can be closely adjacent to the detecting part 20, that is, the light emitting part 10 and the detecting part 20 which are relatively close to each other can be arranged on the same integrated chip, so that the difficulty of the assembly process can be reduced, and the cost of products can be reduced.

More importantly, unlike the prior art, which requires that two long wires extending from the integrated chip are respectively connected to the light emitting part and the detecting part at two ends of the light path, the sum of the distances from the processor of the integrated chip to the light emitting part 10 and the detecting part 20 can be made closer, that is, the data transmission distance is greatly reduced, and the processing speed of the data inside the integrated chip can be improved.

Of particular importance, in embodiment 6 below, features integrated on the same integrated chip incorporate a variable mask portion 32, and such a layout may further increase the speed of feedback, increase the speed of system processing, and refer to embodiment 6 in particular.

Further, the inverse fourier transform section 33 is provided as a super surface reflection type optical element 91, and a plurality of elements of the nanostructure are provided on the surface of the inverse fourier transform section 33.

Unlike the fourier transform unit 31, if the super-surface reflection type optical element 91 is also provided, the incident angle and the emission angle corresponding to the fourier transform unit 31 may be equal, and if the super-surface reflection type optical element 91 corresponding to the inverse fourier transform unit 33 is directly reflected and focused, the cross-talk of the signal may be caused by the reflection of the optical signal portion back to the mask unit 32 or the duty ratio of the stray light signal may be increased, so that the optical signal is prevented from being reflected back to the mask unit 32 again for better transmission effect, and the inverse fourier transform unit 33 of the present embodiment may preferably achieve the effect of focusing in any direction and any position.

The distribution of these elements on the inverse fourier transform section 33 conforms to the design formula:

wherein (X, Y) is the position coordinates on the inverse Fourier transform unit 33, (X) _f ,Y _f ) For focusing the coordinates of the back focus, here, the X direction is defined as the left direction or the right direction in fig. 1-5, and the Y direction is defined as the forward direction or the backward direction perpendicular to the plane shown in fig. 1-5. By designing the phase distribution corresponding to the inverse fourier transform unit 33 in this way, it is possible to achieve focusing of the incident optical signal in any direction.

Further, in embodiment 1, a beam deflection phase may be added to the mask portion 32, and a design formula of beam deflection corresponding to each position in the X direction on the mask portion 32 is as follows:

φ _mask (X)＝-2π/λ[X×(sinθ _I -sinθ _i )](equation II)

Where X is the position on the X-axis of the mask portion 32, θ _I Angle of reflected light, θ _i At this time, each light element is on the plane shown in fig. 1, and the coordinate of the Y axis corresponding to the illustrated plane is 0. Similarly, when the coordinates of each optical element in the Y-axis are not 0, the mask portion can also generate a phase gradient in the Y-direction, and the phase phi in the Y-direction on the mask portion can be similarly confirmed by the formula II _mask (Y) the phase design formula at each position of the mask portion is phi _mask (X)+φ _mask (Y)。

In addition, when at least some of the light elements of the light emitting section 10, the fourier transform section 31, the mask section 32, the inverse fourier transform section 33, and the detecting section 20 use super-surface light elements, off-axis focusing may be achieved, that is, the light elements may not be all in the same plane (the plane is vertical to the front-rear direction), that is, fig. 1 to 5 show only a specific example in which all the light elements are in one plane, or the coordinates of the light elements in the Y direction are not the same, and are illustrated as projections that all fall on the same plane, and the phase design formula at each position of the mask section is adjusted correspondingly according to the positional relationship of the light elements in space.

Thus, by adjusting the phase of the mask portion 32, not only complex amplitude adjustment of the optical signal in the frequency domain plane can be achieved, but also the reflection angle of the reflected light can be changed, deflection of the reflected optical signal to the incident light side as shown in fig. 1 can be achieved, and small angle reflection in the case of large angle incidence can be achieved, and the mask portion 32 of embodiment 1 can shorten the lateral dimension of the entire optical neural network device relative to the other types of mask portions 32 in embodiment 2.

Further, the object plane corresponding to the light emitting portion 10, the reflection plane of the mask portion 32, and the image plane corresponding to the detecting portion 20 are located on the same plane. The optical neural network device can be further reduced in size by using the planarized structure, so that the overall system is highly integrated and the thickness of the overall system can be controlled to be thin. In addition, the fourier transform section 31, the mask section 32, and the inverse fourier transform section 33 may be provided as a superlens, and the thickness of the entire system may be controlled to be in the order of millimeters.

Further, the mask portion 32 is configured as a super-surface reflection type optical element 91, and a plurality of nano-structured elements are arranged on the surface of the mask portion 32, and the amplitude of the complex amplitude of the frequency domain optical signal is regulated by adjusting the size and the spatial distribution of the elements. The different transmittances of light at different positions on the mask portion 32 can realize the regulation of the amplitude of the optical frequency domain complex amplitude signal. The optical signal is converted into an optical frequency domain signal by the fourier transform unit 31, modulated by the mask unit 32, and then enters the inverse fourier transform unit 33 to be transformed into a spatial domain optical signal, whereby the convolution calculation is performed on the input image corresponding to the optical signal based on the convolution kernel using the mask unit 32.

In addition, since the light emitting part 10 and the detecting part 20 are not in a straight line, the incident light signal area and the light collecting light signal area are easily separated by adding the light shielding plate, so that the effect of eliminating stray light is better achieved, and the crosstalk of stray light signals is avoided.

And, the focal lengths of the fourier transform section 31 and the inverse fourier transform section 33 may be different, so that an effect of enlargement or reduction can be achieved.

Example 2

This embodiment 2 differs from embodiment 1 in that: in embodiment 1, both the incident light entering the mask portion 32 and the reflected light reflected by the mask portion 32 are located on the left side of the mask portion 32, that is, the mask portion 32 must adjust the emitting direction of the optical signal so that the incident light and the reflected light are located on the same side. In embodiment 2, as shown in fig. 2, the incident light entering the mask portion 32 is located on the left side of the mask portion 32, and the reflected light exiting the mask portion 32 is located on the right side of the mask portion 32, and if the angle between the reflected light and the incident light is equal to the angle between the incident light and the vertical plane, the reflection of the light by the mask portion is a normal mirror, that is, the adjustment of the direction of the light emission is not required.

That is, if the angles between the reflected light and the incident light are equal to the vertical plane, the mask portion of the present embodiment does not need to be adjusted according to the above formula two, and only the inverse fourier transform portion 33 needs to be provided according to the above formula one.

The present embodiment provides relatively fewer modifications to the mask portion 32 as compared to embodiment 1, and the length of the optical neural network device in the lateral direction is relatively longer in fig. 2, so that it can be selected in particular in conjunction with the engineering scene.

Example 3

This embodiment 3 differs from embodiment 1 in that: at least one of the light emitting section 10, the mask section 32, and the detecting section 20 is provided with an extinction optical trap 40 toward the side of the optical signal transmission region, and the extinction optical trap 40 is used to eliminate other stray light on the target optical path to the detecting section 20 through the light emitting section 10, the fourier transform section 31, the mask section 32, and the inverse fourier transform section 33.

In this embodiment, as shown in fig. 3, the extinction optical traps 40 are disposed above the light emitting portion 10, the mask portion 32 and the detecting portion 20, so that only the light of the object in fig. 3 can pass through, and interference of other reflected stray light on the light signal is reduced.

Example 4

Embodiment 4 is also for avoiding interference of stray light with an optical signal, and differs from embodiment 3 in that: the reflection surface of the mask portion 32 of embodiment 4 is not on the same plane as the object surface of the light emitting portion 10;

and/or the number of the groups of groups,

the reflection surface of the mask portion 32 is not on the same plane as the image surface corresponding to the detection portion 20.

As shown in fig. 4, a step 50 is provided below the mask 32, and the step 50 steps up the mask 32 so that the reflection surface of the mask 32 is closer to the fourier transform unit 31 and the inverse fourier transform unit 33 than the object surface of the light emitting unit 10 and the image surface of the detecting unit 20.

In other embodiments, the step portion 50 may also make the mask portion 32 lower, i.e., the reflection surface of the mask portion 32 is farther from the fourier transform portion 31 and the inverse fourier transform portion 33 than the object surface corresponding to the light emitting portion 10 and the image surface corresponding to the detecting portion 20.

Since the reflection surface of the mask portion 32 is not on the same plane as the object surface corresponding to the light emitting portion 10 and the image surface corresponding to the detecting portion 20, the mask portion 32 is not easily affected by stray light from the light emitting portion 10 and the detecting portion 20, and interference of other reflected stray light with an optical signal is reduced.

Example 5

This embodiment 5 differs from embodiment 1 in that: this example illustrates various embodiments of the fourier transform section 31, for example, the fourier transform section 31 may be provided as a concave reflective optical element 92 or a super-surface reflective optical element 91. As shown in fig. 5, the fourier transform unit 31 is configured as a concave reflection type optical element 92, and the concave reflection type optical element 92 has a simple structure, and only needs to make its focal length meet specific design requirements.

Example 6

This embodiment 6 differs from embodiments 1 and 2 in that: the mask portions in embodiments 1 and 2 may be an existing super surface or a mask made with SLM (Spatial Light Modulator ), and embodiments 1 and 2 can be specifically designed only for a single task in a specific scene, and only can process a single task. While example 6 can be used for the mask portion 32 of various task scenes, it corresponds to a variable convolution kernel in a convolution computing optical system.

Specifically, the optical neural network device further includes a processor, the processor outputs a light emitting signal to the light emitting portion 10 and receives the light signal of the detecting portion 20, the mask portion 32 is a liquid crystal on silicon device, the optical neural network device includes a regulating component, the processor outputs a voltage signal to the regulating component according to the received light signal, the regulating component applies a controllable voltage to the liquid crystal on silicon device according to the voltage signal, and the controllable voltage is used for controlling the spatial distribution of the liquid crystal in the liquid crystal on silicon device so as to change the modulation result of the light signal by the mask portion 32.

The mask portion 32 is manufactured by adopting a liquid crystal on silicon technology (LCoS, liquid Crystal on Silicon), and can realize simultaneous modulation of the spatial distribution of the amplitude and the phase of the optical signal, and control the spatial distribution of the liquid crystal in the liquid crystal layer by addressing and controlling the voltage applied to the mask plate, so as to realize different modulation effects of the spatial distribution of the amplitude and the phase of the optical signal, namely realize programmable complex amplitude modulation capability, and correspond to the variable convolution kernel in the whole convolution calculation optical neural network. The mask portion 32 may be made of other refractive index adjustable materials such as electro-optic polymers and chalcogenide glass, and the mask portion 32 will be described below by taking a material using liquid crystal on silicon technology as an example.

More specifically, the modulating component applies a voltage to the liquid crystal layer of the mask portion 32 on each pixel, so that the liquid crystal molecules in the liquid crystal layer at the corresponding position spatially rotate, and after the liquid crystal molecules rotate, the polarization state of the incident light can be changed to realize the amplitude modulation of the emergent light by using the birefringence effect of the liquid crystal molecules, and when the liquid crystal molecules spatially rotate, the equivalent refractive index of the liquid crystal layer changes, so that the phase modulation of the incident light is realized.

Therefore, the phase and amplitude of the incident light can be regulated and controlled simultaneously by adjusting the voltages applied to the two ends of the liquid crystal layer of the mask portion 32, and different regulating and controlling effects can be realized by changing the magnitude of the added voltage, namely, the programmable regulating and controlling capability of the incident light complex amplitude can be realized, so that the continuous iteration of the convolution kernel can be realized through programming the convolution kernel to construct the mask portion 32 meeting the requirement of a specific task.

Since the adjustment of the input image and the adjustment of the adjustment function of the mask portion 32 are required to be continuously performed during the operation of the whole system, especially when the mask portion 32 is required to be adjusted according to the task requirement in embodiment 6, the effect of image processing depends on the comparison of the input image and the output image, the convolution kernel effect of the mask portion 32 is determined according to the processing effect, and then the adjustment is performed, and the above steps are repeatedly iterated repeatedly, so as to implement multiple convolution calculations and optimal convolution kernel selection. The iteration speed of this feedback process depends largely on the comparison speed between the input image and the output image, so that the spatial distance between the light emitting part 10 and the detecting part 20 is shortened as much as possible, the comparison speed between the input image and the output image can be improved most directly, the signal response speed is high, and further the feedback adjustment process of programming the complex amplitude adjustment mask plate according to the comparison result of the input image and the output image is accelerated, and finally the mask part 32 is quickly iterated to the convolution kernel of the target according to the current task.

The mask portion 32 of the present embodiment may also be integrated on an integrated chip, so that the light emitting portion 10, the detecting portion 20 and the mask portion 32 are all located on the integrated chip, which not only can further improve the integration level, effectively reduce the overall size of the system, but also can increase the processing speed of the optical computing system.

Example 7

This example 7 differs from examples 1 to 6 in that: the light emitting section 10 of embodiments 1 to 6 may be a normal light source, and the detecting section 20 may be a normal CMOS, for example, the detection frequency of embodiments 1 to 6 is on the order of kilohertz. Whereas the light emitting section 10 of embodiment 7 is provided as a high-frequency light source, the detecting section 20 is provided as a high-detection frequency detector, and the frequency corresponding to the high frequency is in the order of megahertz.

Specifically, the light emitting unit 10 may be configured by a parallel light source, for example, a laser light source combined with a waveguide array, a laser light source combined with a digital microscopic device, or a laser light source combined with a liquid crystal spatial light modulator, and generates an image signal to emit an optical signal of parallel light to the fourier transform unit 31. Wherein the laser light source outputs a high frequency signal in combination with the waveguide array, and accordingly, the detection section 20 uses the photodiode array as a detector to realize high frequency detection of the signal.

And since embodiments 1 to 6 set the light emitting section 10 and the detecting section 20 relatively close, the sum of the distances from the integrated chip to the light emitting section 10 and the detecting section 20 is relatively small, the delay of the electric signal is relatively small, and then when a high-frequency light source and a high-detection frequency detector are used, the response speed is faster, that is, the embodiment 7 is made to have greater realism on the basis that the light emitting section 10 and the detecting section 20 of embodiments 1 to 6 are set relatively close.

In particular, when this embodiment 7 is combined with embodiment 6, the following technical effects will also be produced: embodiment 7 has higher efficiency in performing batch processing of convolution operations of a plurality of sets of different input image signals and different convolution kernels at the same time, and can perform high-speed switching of complex amplitudes of image light signals input to a convolution computing optical system.

At this time, the switching frequency of the controllable voltage applied by the control component is consistent with the frequency of the light signal emitted by the light emitting part 10, so that the convolution operation of the input image signal and different convolution kernels can be realized at a higher frequency, and the time required for iterating the convolution kernels is reduced.

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

(1) According to the optical neural network device, through the design of the structure of the optical path component 30 and the adjustment of the positions and the light reflecting directions of the Fourier transform part 31, the inverse Fourier transform part 33 and the mask part 32, other parts can be omitted between the light emitting part 10 and the detection part 20, namely, the light emitting part and the detection part can be arranged as close as possible, so that the distance from the chip to the chip for outputting the light emitting signal to the chip for receiving the light signal is shorter, the time interval is shorter, and the optical path is folded, so that the occupied space volume of the optical neural network device is greatly reduced compared with the existing linear optical neural network, the processing speed of the optical neural network device is faster while the volume is reduced, and the optical neural network device has better application prospect.

(2) The light emitting part 10 and the detecting part 20 can be integrated on the same integrated chip, so that data is transmitted inside the integrated chip, signal loss is reduced, and operation speed is greatly improved.

(3) Various optical neural network device embodiments, some can avoid receiving the interference that reflected stray light caused to the detection signal, some can adjust the convolution kernel that mask portion 32 corresponds, some combine high-speed luminous portion 10 and high-speed detection portion 20, make optical neural network device realize quick convolution kernel iteration, and the running speed is faster while the performance is stronger.

It should be understood that although the present disclosure describes embodiments, not every embodiment is provided with a separate embodiment, and that this description is for clarity only, and that the skilled artisan should recognize that the embodiments may be combined as appropriate to form other embodiments that will be understood by those skilled in the art.

The above list of detailed descriptions is only specific to practical embodiments of the present invention, and they are not intended to limit the scope of the present invention, and all equivalent embodiments or modifications that do not depart from the spirit of the present invention should be included in the scope of the present invention.

Claims

1. An optical neural network device, comprising:

a light emitting section for emitting an optical signal;

a detection section for receiving an optical signal;

2. The optical neural network device of claim 1, further comprising an integrated chip, wherein the light emitting portion and the detecting portion are both integrated on the same integrated chip.

3. The optical neural network device according to claim 2, wherein the inverse fourier transform is configured as a super surface reflection type optical element, and the surface of the inverse fourier transform is configured with a plurality of elements of a nanostructure, and the distribution of the elements conforms to a design formulaWherein (X, Y) is a position coordinate on the inverse Fourier transform unit (X) _f ,Y _f ) Is the coordinates of the focus after focusing.

4. The optical neural network device according to claim 2, wherein the incident light entering the mask portion and the reflected light reflected by the mask portion are both located on the left side of the mask portion;

the fourier transform unit is disposed at the upper right side of the light emitting unit.

5. The optical neural network device of claim 4, wherein the design formula of beam deflection corresponding to each position in the X direction on the mask portion is # _mask (X)＝-2π/λ[X×(sinθ _I -sinθ _i )]Wherein X is the position on the X axis of the mask portion, θ _I Angle of reflected light, θ _i Is the angle of the incident light.

6. The optical neural network device according to claim 2, wherein the object plane corresponding to the light emitting portion, the reflection plane of the mask portion, and the image plane corresponding to the detection portion are located on the same plane.

7. The optical neural network device according to claim 6, wherein at least one of the light emitting section, the mask section, and the detecting section is provided with an extinction optical trap for eliminating other stray light on a target optical path to the detecting section via the light emitting section, the fourier transform section, the mask section, and the inverse fourier transform section, toward a side of an optical signal transmission region.

8. The optical neural network device according to claim 2, wherein the reflection surface of the mask portion and the object surface corresponding to the light emitting portion are not in the same plane;

and/or the number of the groups of groups,

9. The optical neural network device according to claim 2, wherein the fourier transform section is provided as a concave reflective optical element or a super surface reflective optical element.

10. The optical neural network device according to claim 2, wherein the fourier transform section converts the spatial domain optical signal into the frequency domain optical signal, the mask section is configured as a super-surface reflection type optical element, the mask section surface is configured with a plurality of nano-structured elements, and the amplitude of the complex amplitude of the frequency domain optical signal is regulated by adjusting the size and spatial distribution of the elements.

11. The optical neural network device of claim 10, further comprising a processor that outputs a light emission signal to the light emitting portion and receives the light signal from the detecting portion, the mask portion being a liquid crystal on silicon device, the mask portion also being integrated on the integrated chip, the optical neural network device comprising a regulation component that outputs a voltage signal to the regulation component based on the received light signal, the regulation component applying a controllable voltage to the liquid crystal on silicon device based on the voltage signal, the controllable voltage being used to control a spatial distribution of liquid crystals in the liquid crystal on silicon device to change a modulation result of the light signal by the mask portion.

12. The optical neural network device of claim 11, wherein the light emitting portion is configured as a high frequency light source, the detection portion is configured as a high detection frequency detector, and the switching frequency of the controllable voltage applied by the control component is consistent with the frequency of the light signal emitted from the light emitting portion.