CN112101514A - Diffraction neural network adopting pyramid structure diffraction layer for light supplement and implementation method - Google Patents
Diffraction neural network adopting pyramid structure diffraction layer for light supplement and implementation method Download PDFInfo
- Publication number
- CN112101514A CN112101514A CN202010729442.5A CN202010729442A CN112101514A CN 112101514 A CN112101514 A CN 112101514A CN 202010729442 A CN202010729442 A CN 202010729442A CN 112101514 A CN112101514 A CN 112101514A
- Authority
- CN
- China
- Prior art keywords
- diffraction
- layer
- light
- neural network
- area
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a diffraction neural network adopting a pyramid-structured diffraction layer for light supplement and an implementation method, wherein the diffraction neural network divides the diffraction layers in the network into two groups, namely a light supplement area diffraction layer and a diffraction area diffraction layer; the size of the light supplementing area diffraction layer is changed to enable the light supplementing area diffraction layer to be in a pyramid structure, so that part of input light can bypass a plurality of light supplementing area diffraction layers; the diffraction zone is between the light supplement zone and the detection layer, and the diffraction layers of the diffraction zone are all standard in size. The method firstly designs a basic scheme of the diffraction neural network, confirms the relation between the network performance and the layer number, and then designs a light supplement scheme according to task requirements and network characteristics. According to the invention, the idea of network performance is improved by building more layers of networks by improving the brightness, the selection space of diffraction layer materials and detectors is increased, and the environmental adaptability and the adjustment space of the networks in practical application are enhanced; no extra light source and device are introduced, the size is small, the energy consumption is low, and the installation and the application are convenient.
Description
Technical Field
The invention relates to a diffraction neural network adopting a pyramid structure diffraction layer for supplementing light and an implementation method thereof, which improve the receiving brightness of a detection layer by adjusting the structure of an optical diffraction layer so as to improve the depth of the network, belong to the field of combination of deep learning and physical optics, and mainly relate to the physical implementation of light diffraction and a neural network. The method has wide application prospect in various optical diffraction device-based neural network application systems and schemes.
Background
Traditional computers are designed based on the von neumann scheme and are built up from logic cells consisting of transistors. Integrated circuits have been facing a technological bottleneck, following the rate of development of moore's law over the last few decades. Photons do not generate heat, and can not be influenced by electromagnetic induction and capacitance effect to cause signal degradation, and the optical neural network has the characteristics of large time bandwidth area, high interconnection capacity and inherent parallel processing in tasks such as optical signal transmission and the like, can greatly improve the performance of the neural network, and can be produced accordingly. The research on the optical neural network is different from the realization method by the side focus, and can be divided into three categories, namely, the construction of photoelectric devices, particularly networks taking a photoelectric modulator as a core, networks taking a programmable optical chip as a carrier, and networks for task-oriented customized devices. A method of constructing a neural network by combining a multilayer BP network with an optical diffraction phenomenon (see literature: Linxing et al, all-optical machine learning using a diffractive neural network, Science, 2018: eat 8084- (Xing L, Yair R, Yardimci N T, et al, all-optical machine learning using differential street networks [ J ]. Science, 2018: eat 8084-)), using coherent light as input light; placing a task target to be processed on an input layer, transmitting light rays into a trained diffraction layer group after the light rays are diffracted by the input layer, representing a neuron by each point on the diffraction layer, wherein neuron parameters comprise an amplitude modulation coefficient and a phase modulation coefficient, and finally obtaining a network result by a detector on a detection layer; random gradient descent was chosen for training, with MSE as the loss function. The process can be summarized as a three-step framework: training completely on a computer; manufacturing diffraction layer elements according to the trained parameters; the building element is applied. The performance of the Diffractive Neural Network can be further improved by adding a 4-f system consisting of a nonlinear layer and a double lens (reference document: Dougi et al, Fourier space Diffractive Neural Network, Physical review report, 2019,123 (2)) (Yan Tao, Wu Jianmin, Zhou Tiankuang, Xie Hao, Xu Feng, Fan Juntao, Fan Lu, Lin Xig, Dai Qiongha.Fourier-space differential Deep Neural Network [ J ]. Physical review letters,2019,123 (2)). The 4-f system can convert an input signal into a frequency domain, so that the problems which cannot be solved by a basic network such as significance detection are realized.
The loss of the diffraction layer to the optical signal energy is inevitable, and for example, the loss rate of the 3D printing material to 0.4THz microwave can reach 0.51. This results in two more ideas being used when building physical models: 1. microwave is selected as the input signal. This method has a problem of environmental suitability. Regardless of the optical neural network, when the purpose of image processing is to process the visible light information input by the lens, the final expectation is that the visible light information is directly processed without going through the step of converting the image in the computer into a signal and inputting the signal into the neural network, and therefore, it is a great trend to select the visible light as the input signal. 2. The number of layers of the network is reduced while visible light is selected as input, preprocessing is performed to a certain degree only through one layer of neural network, and finally the photoelectric combined neural network is set up in cooperation with a traditional computer to achieve the problem. The invention provides a light supplement scheme aiming at a diffraction neural network by taking optical loss as an access point.
The energy loss caused by the diffraction layer is an important problem of the diffraction neural network, and the number of achievable layers of the optical diffraction neural network is limited. In the process of many problems, the increase of the number of layers of the neural network can improve the performance of the neural network.
Disclosure of Invention
In order to solve the problem of limited network depth caused by optical loss and detector precision limitation, the invention provides a diffraction neural network adopting a pyramid structure diffraction layer for supplementing light and an implementation method thereof.
The invention relates to a diffraction neural network adopting a pyramid-structured diffraction layer for light supplement, wherein the diffraction layers in the network are divided into two groups, namely a light supplement area diffraction layer and a diffraction area diffraction layer. The whole light supplementing area further comprises an input layer, the size of the diffraction layer of the light supplementing area is changed, the light supplementing area is made to be pyramid-shaped, and therefore a part of input light can bypass the plurality of diffraction layers of the light supplementing area, and the purpose of reducing loss is achieved. The diffraction zone is between the light supplementing zone and the detection layer, and the diffraction layers of the diffraction zone are in standard sizes, so that the purpose of adding the added light into the network is achieved.
Furthermore, the invention adds a supplementary light intensity control area around each supplementary light area diffraction layer presenting pyramid structure, and selects an amplitude modulation material to control the intensity of the supplementary light. The amplitude modulation coefficient of the supplementary lighting intensity control area is a fixed constant which does not participate in network training.
The distances among the diffraction layers of the pyramid structure light supplement area are the same, the width of each layer is arranged according to an arithmetic progression, and the pyramid structure light supplement area has the same width w of the light supplement area.
Preferably, if only brightness improvement is aimed at, the width of the diffraction layer in the light supplement area in the previous layer is less than or equal to the width of the diffraction layer in the light supplement area in the next layer, so that the diffraction layer in the light supplement area has a stepped structure, and diffraction among the diffraction layers can be realized.
The invention further provides a method for realizing the diffraction neural network by adopting the pyramid-structured diffraction layer for light supplement. The method comprises the following specific steps:
the method comprises the following steps: designing an optical diffraction neural network according to a task target, and selecting elements such as diffraction layer materials, light sources, detectors and the like.
Step two: calculating the maximum number N of layers of the constructable diffraction neural network according to the selected materials and the detectorbAnd carrying out simulation confirmation; the number of layers of the simulation test network is more than NbThe neural network scheme of (3) to confirm the relationship between the network performance and the number of layers.
Step three: in the simulation result in the second step, a neural network scheme capable of improving the network performance by increasing the number of layers is selected, the diffraction layer of the neural network scheme is divided into a light supplement region diffraction layer and a diffraction region diffraction layer, and the size of the light supplement region diffraction layer is changed to enable the light supplement region diffraction layer to be in a pyramid structure (see fig. 2).
Preferably, the method further comprises the step four: and (3) adding a supplementary light intensity control area (see figure 3) around the diffraction layer of the pyramid-structured supplementary light area, and selecting an amplitude modulation material to control the intensity of supplementary light. The width w (see fig. 3 and 4) of the light supplement area and the amplitude modulation coefficient of the light supplement intensity control area are properly selected according to the task target and the network characteristics.
The invention relates to a diffraction neural network adopting a pyramid structure diffraction layer for light supplement and a realization method thereof, and has the advantages and effects that: the invention takes the optical loss as an access point, divides the diffraction layers into two groups, changes the size of the diffraction layers in a light compensation area close to the input, leads the diffraction layers to present a pyramid structure, and can lead a part of input light around to bypass a plurality of diffraction layers, thereby achieving the purpose of reducing the loss. And a plurality of diffraction layers with the same size are arranged in the modulation area behind the light supplementing area, so that the purpose of adding the added light into the network is achieved. Further, a light supplement intensity control area is added around the diffraction layer of the light supplement area of the pyramid-structured light supplement scheme. And the amplitude modulation material is selected to control the intensity of the supplementary light, so that the feasibility and the adjustment space of the supplementary light scheme are increased. The idea of building more layers of networks by improving the brightness is realized, so that the network performance is improved, the selection space of diffraction layer materials and detectors is increased, and the environmental adaptability and the adjustment space of the networks in practical application are enhanced. Meanwhile, no additional light source or device is introduced, the optical fiber has adaptability to various two-dimensional optical neural networks, has the characteristics of small volume, low energy consumption, convenience in installation and easiness in application, and has wide market prospect and application value.
Drawings
FIG. 1 is a diagram showing the basic structure of a diffractive neural network according to the present invention.
Fig. 2 is a network structure diagram of the light supplement scheme of the present invention.
FIG. 3 is a schematic diagram of a trainable portion of a patch diffraction layer of the present invention.
Fig. 4 is a detailed structural view of each of the light-compensating region diffraction layers in the present invention.
FIG. 5 shows the brightness of the detection layer under different amplitude modulation coefficients of the material according to the present invention.
Fig. 6a, b, c are comparisons of a four-layer network light supplement scheme with a two-layer non-light supplement scheme in the popular article classification problem of the present invention. Fig. 6a and 6b are the results of brightness simulation of the detection layer, and fig. 6c is the comparison of accuracy of the test set.
Fig. 7a, b and c are the comparison of the light supplement scheme of four-layer network and the light supplement scheme of three-layer network in the popular article classification problem. Fig. 7a and 7b are the results of brightness simulation of the detection layer, and fig. 7c is the comparison of accuracy of the test set.
Detailed Description
For better understanding of the technical solutions of the present invention, the following further describes embodiments of the present invention with reference to the accompanying drawings.
As shown in fig. 2, the diffraction layers in the diffraction network are divided into two groups, namely a light supplement area diffraction layer and a diffraction area diffraction layer. The whole light supplementing area further comprises an input layer, the size of the diffraction layer of the light supplementing area is changed, the light supplementing area is made to be pyramid-shaped, and therefore a part of input light can bypass the plurality of diffraction layers of the light supplementing area, and the purpose of reducing loss is achieved. The diffraction zone is between the light supplementing zone and the detection layer, and the diffraction layers of the diffraction zone are in standard sizes, so that the purpose of adding the added light into the network is achieved.
It should be noted that, in order to implement a fully-connected network, it is necessary to ensure that the output light of each neuron can reach each neuron of the next diffraction layer, and the maximum diffraction half-cone angle is calculated, which is expressed as follows:
fmax＝1/2df
wherein f ismaxRefers to the maximum spatial frequency, dfRepresenting the layer characteristic dimension. Sometimes, the diffraction layer is placed very tightly, and a network structure which only realizes partial linkage can also better solve some problems and even has better effect. In the invention, in order to better fuse the supplementary light and the input information, the interlayer distance capable of realizing full-connection calculation is selected.
The size of the neuron of the diffraction layer is matched with the wavelength of input light, the size of the diffraction layer is designed according to task requirements, and when nonlinear calculation is lacked, the neuron of the single-layer diffraction layer is not too much, so that overfitting is avoided. The light supplementing width of the diffraction layer of the light supplementing area is preferably 1/8-1/20 of the width of the diffraction layer, the undersized light supplementing width cannot reach the brightness target required by the number of the network layers, and the oversized light supplementing width can cause the information fusion unbalance of the supplementing light and the input light, so that the network performance is reduced.
Furthermore, a light supplement intensity control area is added around each light supplement area diffraction layer of the pyramid structure, and an amplitude modulation material is selected to control the intensity of supplement light; the amplitude modulation material matched with the light is specifically selected and adopted. The amplitude modulation coefficient of the supplementary lighting intensity control area is a fixed constant which does not participate in network training, and the supplementary lighting intensity control area can weaken the supplementary lighting to a certain extent, so that the energy balance between the supplementary lighting and the original lighting can be regulated and controlled more flexibly, and the feasibility and the selection space of a supplementary lighting scheme are increased. The idea of building more layers of networks by improving the brightness is realized, so that the network performance is improved, meanwhile, the selection space of the diffraction layer materials and the detector is increased, and the environmental adaptability and the adjustment space of the network in practical application are enhanced.
The embodiment is a scheme aiming at a two-dimensional plane neural network built based on an optical diffraction phenomenon, and a schematic block diagram is shown in fig. 1, and the specific implementation steps are as follows:
the method comprises the following steps: designing an optical diffraction neural network according to a task target, and selecting elements such as diffraction layer materials, light sources, detectors and the like. The light source is a coherent light source. The diffraction layer material can be selected from three materials of amplitude modulation, phase modulation and amplitude phase modulation, and the selection schemes are respectively selected according to different network training parameters:
in a diffractive neural network, each neuron on the diffractive layer can be represented by the following formula
Where t represents the influence of the neurons of the diffraction layer, called the transmission coefficient, i.e. the so-called weight,. phi.represents the change in phase, a represents the change in amplitude, and (x, y, z) is the spatial coordinate of the point. When only phase training is considered, a transparent material is selected as a diffraction layer, and the amplitude coefficient can be regarded as a constant. When only amplitude training is considered, the screen is regarded as an attenuation screen, only the intensity of light is affected, the phase coefficient can be regarded as a constant, and the liquid crystal screen is one of the most common attenuation screens. Considering the optimized cheapness of neural networks and the unique property of light, the scheme of selecting the phase modulation coefficients of neurons as training parameters is the most common. The material is mostly transparent material, can select 3D printing material, or the polyethylene that the loss rate is lower, polytetrafluoroethylene etc.. In other studies, glass containing impurities, phase masks (DOE), etc. may also be used to adjust the phase of light propagating in the network.
Selecting the phase as the modulation coefficient, firstly, the specific step of forward propagation is defined. Taking 3D printing material (VeroBlackPlus RGD875) as an example, a certain initial thickness is given to the diffraction layer, on the basis of which the influence of each neuron on the phase depends on the increased thickness difference, satisfying the following formula.
Δz＝λφ/2πΔn
n is a refractive index, z is a thickness of the diffraction layer, and different detector receiving design schemes are adopted in the detection layer aiming at different processing targets. For example, when the method is used for processing image classification problems of fewer categories, detectors with the number equal to the number of categories can be placed on the detection layer (see fig. 1), and a category number is assigned to each detector, so that the detector number with the brightest received signal is used as a network result. In the case of using visible light as an input signal, a device such as a camera may be used as a receiving element for an output signal.
Step two: and preliminarily confirming the maximum number of layers of the constructed diffraction neural network according to the selected materials and the detector. And performing simulation confirmation. And simultaneously, simulating and testing a neural network scheme with more network layers, and confirming the relation between the network performance and the layer number.
The diffraction neural network utilizes diffraction simulation full-link calculation of light, and the schematic block diagram is shown in FIG. 1. Coherent light is used as input light. A task target to be processed is placed on an input layer, light is transmitted into a trained diffraction layer group after being diffracted by the input layer, each point on the diffraction layer represents a neuron, and neuron parameters are composed of amplitude modulation coefficients and phase modulation coefficients. And finally, obtaining a network result by a detector on the detection layer. Random gradient descent was chosen for training, with MSE as the loss function.
Is provided with
The output of the ith neuron in the ith layer at the position with the coordinate (x, y, z) is influenced by the input, the transmission coefficient and the transmission parameter. The input of the previous layer, namely the input provided by the l-1 layer, adds the input of each neuron of the l-1 layer at the position of the ith neuron of the l layer to form the input of the neuron
The transmission parameter is related to the quantification of the distance relationship between two neurons in the two layers of neural networks and is expressed as
According to the Rayleigh-Sommerfeid diffraction equation, each neuron is regarded as a quadratic source of a wave, and the formula is as follows:
to summarize, the ith neuron in layer l has a mathematical relationship with the pth neuron in layer l +1 according to the following equation:
for the input layer, there is also a similar physical relationship:
the intensity of the detector area is then recorded at the output layer:
the selected training scheme is an optimization method of random gradient descent, and MSE is calculated as a loss function. Taking the phase coefficient as an example of the training parameter, the formula is as follows:
defining the maximum number N of layers of diffraction neural network which can be built under the selected scheme and deviceb. After the simulation is performed on the network with different layer numbers, if the performance can be further improved by increasing the layer number, the light supplement scheme is designed.
Step three: under the condition that the number of layers is increased, the network performance can be improved to a certain degree, the diffraction layer is divided into a light supplement area diffraction layer and a diffraction area diffraction layer, and the size of the light supplement area diffraction layer is changed to enable the light supplement area diffraction layer to be in a pyramid structure (see fig. 2).
Step four: and adding a light supplement intensity control area (see fig. 4) around the diffraction layer of the light supplement area in the pyramid structure light supplement scheme, and selecting an amplitude modulation material to control the intensity of the supplement light. The width of the supplementary lighting area and the amplitude modulation coefficient of the supplementary lighting intensity control area are properly selected according to the task target and the network characteristics.
In fact, such fill-in schemes have greater flexibility than just the standard pyramid architecture. In the pyramid framework, the distances between the diffraction layers of the light supplement region are the same, the width of each layer is arranged according to an arithmetic progression, and the light supplement region has the same width w. If only the aim of improving the brightness is to achieve the purpose, the width of the diffraction layer of the previous light supplement area is less than or equal to that of the diffraction layer of the next light supplement area, so that the diffraction layer of the light supplement area is in a stepped structure, and diffraction among diffraction layers can be achieved. However, it is found in experiments that the pyramid structure can improve more accuracy, and thus becomes a solution of choice.
Due to the complex value characteristic of light and the existence of diffraction, accurate supplementary light intensity derivation cannot be carried out through a formula, and only certain estimation can be carried out. Assuming that the total amount of light intensity of each diffraction layer is not affected by the diffraction and phase modulation of the diffraction layer, the fixation is Q, which is only proportionally attenuated by the diffraction layer material. Setting the width of each light supplement layer as W, the number of the light supplement layers as NA and the width of the standard diffraction layer as WB, the width W0 of the input diffraction layer as WB-2 xw, and setting the energy which can be provided by the light supplement layer per unit area for the next diffraction layer as ek。
Then, a layer network without light supplement has initial energy of Q0The energy loss rate, that is, the amplitude modulation coefficient is α, the amplitude modulation coefficient of the fill light intensity control area is 1, and the energy Ed that can be received by the detector layer is expressed as follows:
Ed＝Q0·αl
in the same network, when the number of supplementary lighting layers NA exists, the energy that the detector can receive is:
fs(i)＝(4w2+4w(WI+2w·i))·ek·αl-i
WI＝WB-2·w·NA
WI refers to the width of an input layer, an fs function is used for calculating the energy increase which can be provided by each light supplement area, in order to simplify the processing, an optical diffraction neural network with only two diffraction layers and one light supplement layer is considered, and then the formula is expressed as follows:
Ed＝Q0·α2
EdA＝Q0·α2+(4w2+4w·WI)·ek·α2
if the strength of the two layers of neural network detection layers after expected light supplement is similar to a single layer neural network, the formula needs to be satisfied:
EdA2＝Ed1
finally, the following steps are obtained:
QA＝(4w2+4w·WI)·ek
QAthe energy provided for single-layer supplementary lighting can be clearly seen, the required supplementary lighting energy increases with the increase of the loss rate, if the addition of more layers is to be realized through single-layer supplementary lighting, the last formula needs to be satisfied, and M is the network layer number approximate to the target.
When the multi-layer light supplement device is popularized to the multi-layer situation, similar simplification can be carried out, the effect of the light supplements of the first layers is ignored, only the light supplement of the last layer is considered, obviously, the problem similar to the single-layer light supplement is solved, and only the following formula needs to be satisfied:
therefore, the light supplement effect of the light supplement scheme is theoretically verified and guided by design. The light supplement scheme of the neural network can be preliminarily designed through the energy proportion approximately estimated by the formula, and the final light supplement scheme is determined by means of experimental verification and adjustment, so that network optimization is realized.
The amplitude modulation coefficient of the jth supplementary illumination intensity control area of the ith diffraction layer of the supplementary illumination area is
The increase of w and NA can increase light intensity, but the ratio of unbalanced light supplement to input information can damage network performance, and parameter allocation should be performed according to practical application. Selecting larger task from tasks with network performance obviously increased along with network layer number
w and NA in order to increase the number of layers tend to give better results. And selecting smaller tasks among tasks with network performance not increasing much along with the number of network layers
w and NA. Fig. 5 shows the brightness contrast of the network after light supplement and the detection layer of the non-supplemented optical network when materials with different amplitude modulation coefficients are used as the diffraction layer. It can be seen that the diffractive neural network scheme in which the material with a large amplitude modulation coefficient is used as the diffractive layer is easier to extend the depth of the network by compensating the light.
Taking the problem of popular article classification as an example, fig. 6a, b, and c show that when the amplitude modulation coefficient of the diffraction layer is 0.5, the brightness of the four-layer network detection layer is increased to the energy level of the three-layer network, the light supplement width is one tenth of the standard width, and the number of the light supplement layers is four. Fig. 6a and 6b are the results of the brightness simulation of the detection layer, and fig. 6c is the comparison of the accuracy of the test set.
Fig. 7a, b, and c show that when the amplitude modulation coefficient of the diffraction layer is 0.8, the brightness of the four-layer network detection layer is increased to the energy level of the two-layer network, the light supplement width is one tenth of the standard width, and the number of light supplement layers is four. Fig. 7a and 7b are the results of the brightness simulation of the detection layer, and fig. 7c is the comparison of the accuracy of the test set. Since the amplitude modulation coefficient of the diffraction layer does not affect the accuracy of the diffraction neural network without light compensation theoretically, compared with the data in fig. 6a, b and c, the accuracy of the network after light compensation is higher than that of a two-layer network, but slightly lower than that of a one-layer network. Therefore, for the tasks with less network performance increase with the number of network layers, a smaller one should be selected
w and NA.
The diffraction neural network scheme using the material with a large amplitude modulation coefficient (more than or equal to 0.7) as the diffraction layer and the task that the network performance is more obvious along with the increase of the number of the network layers are more suitable for the light supplement method provided by the invention.
Claims (7)
1. The utility model provides an adopt diffraction neural network of pyramid structure diffraction layer light filling which characterized in that: dividing diffraction layers in the network into two groups, namely a light supplementing area diffraction layer and a diffraction area diffraction layer; the whole light supplementing area also comprises an input layer, and the size of the diffraction layer of the light supplementing area is changed to enable the light supplementing area to be in a pyramid structure, so that part of input light can bypass the plurality of diffraction layers of the light supplementing area; the diffraction zone is arranged between the light supplementing zone and the detection layer, the diffraction layer of the diffraction zone is in a standard size, and added light is added into the network.
2. The diffractive neural network using the pyramid-structured diffractive layer for supplementing light according to claim 1, wherein: the pyramid framework is characterized in that distances among diffraction layers of the light supplement regions are the same, the width of each diffraction layer is arranged according to an arithmetic progression, and the light supplement regions have the same width w.
3. The diffractive neural network using the pyramid-structured diffractive layer for supplementing light according to claim 1, wherein: the neural network further comprises a supplementary lighting intensity control area which is added around each diffraction layer of the supplementary lighting area in the pyramid structure, and an amplitude modulation material is selected to control the intensity of supplementary lighting; the amplitude modulation coefficient of the supplementary lighting intensity control area is a fixed constant which does not participate in network training.
4. The diffractive neural network using the pyramid-structured diffractive layer for supplementing light according to claim 3, wherein: the pyramid framework is characterized in that distances among diffraction layers of the light supplement regions are the same, the width of each diffraction layer is arranged according to an arithmetic progression, and the light supplement regions have the same width w.
5. The diffractive neural network using the pyramid-structured diffractive layer for supplementing light according to claim 1, wherein: if the neural network only aims at improving the brightness, the pyramid framework is replaced by a stepped framework, namely, the width of the diffraction layer of the light supplement area at the previous layer is less than or equal to that of the diffraction layer of the light supplement area at the next layer, so that the diffraction layer of the light supplement area is in the stepped framework, and diffraction among diffraction layers can be realized.
6. A method for realizing a diffraction neural network by using a pyramid-structured diffraction layer for light supplement is characterized by comprising the following steps of: the method comprises the following specific steps:
the method comprises the following steps: designing an optical diffraction neural network according to a task target, and selecting a diffraction layer material, a light source and a detector;
step two: calculating the maximum number N of layers of the constructable diffraction neural network according to the selected materials and the detectorbAnd carrying out simulation confirmation; the number of layers of the simulation test network is more than NbThe neural network scheme of (1), confirming the relationship between the network performance and the number of layers;
step three: and in the simulation result in the second step, selecting a neural network scheme which can improve the network performance by increasing the number of layers, dividing the diffraction layer into a light supplementing region diffraction layer and a diffraction region diffraction layer, and changing the size of the light supplementing region diffraction layer to enable the light supplementing region diffraction layer to present a pyramid structure.
7. The method for implementing the diffractive neural network using the pyramid-structured diffraction layer for light supplement according to claim 6, wherein the method comprises: after the third step, the method further comprises the following steps: a light supplement intensity control area is added around the diffraction layer of the pyramid structure light supplement area, and an amplitude modulation material is selected to control the intensity of supplement light; the width w of the light supplement area and the amplitude modulation coefficient of the light supplement intensity control area are properly selected according to the task target and the network characteristics.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010729442.5A CN112101514B (en) | 2020-07-27 | 2020-07-27 | Diffraction neural network adopting pyramid structure diffraction layer for light supplement and implementation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010729442.5A CN112101514B (en) | 2020-07-27 | 2020-07-27 | Diffraction neural network adopting pyramid structure diffraction layer for light supplement and implementation method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112101514A true CN112101514A (en) | 2020-12-18 |
| CN112101514B CN112101514B (en) | 2022-06-07 |
Family
ID=73750483
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010729442.5A Active CN112101514B (en) | 2020-07-27 | 2020-07-27 | Diffraction neural network adopting pyramid structure diffraction layer for light supplement and implementation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112101514B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113822424A (en) * | 2021-07-27 | 2021-12-21 | 湖南大学 | All-optical diffraction neural network system based on super-structured surface |
| CN117521746A (en) * | 2024-01-04 | 2024-02-06 | 武汉大学 | Quantized optical diffraction neural network system and training method thereof |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110334804A (en) * | 2019-06-20 | 2019-10-15 | 清华大学 | Full optical depth diffraction nerve network system and method based on space segment coherent light |
| CN110929864A (en) * | 2019-12-05 | 2020-03-27 | 北京超放信息技术有限公司 | Optical diffraction neural network on-line training method and system |
| US20200104676A1 (en) * | 2018-09-28 | 2020-04-02 | Asml Netherlands B.V. | Providing a Trained Network and Determining a Characteristic of a Physical System |
| WO2020101863A2 (en) * | 2019-10-25 | 2020-05-22 | Futurewei Technologies, Inc. | Hybrid opto-electronic neural networks and methods for utilization thereof |
| WO2020102204A1 (en) * | 2018-11-12 | 2020-05-22 | Massachusetts Institute Of Technology | Large-scale artificial neural-network accelerators based on coherent detection and optical data fan-out |
| WO2020127422A1 (en) * | 2018-12-19 | 2020-06-25 | Lysia | Hyperspectral detection device |
-
2020
- 2020-07-27 CN CN202010729442.5A patent/CN112101514B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200104676A1 (en) * | 2018-09-28 | 2020-04-02 | Asml Netherlands B.V. | Providing a Trained Network and Determining a Characteristic of a Physical System |
| WO2020102204A1 (en) * | 2018-11-12 | 2020-05-22 | Massachusetts Institute Of Technology | Large-scale artificial neural-network accelerators based on coherent detection and optical data fan-out |
| WO2020127422A1 (en) * | 2018-12-19 | 2020-06-25 | Lysia | Hyperspectral detection device |
| CN110334804A (en) * | 2019-06-20 | 2019-10-15 | 清华大学 | Full optical depth diffraction nerve network system and method based on space segment coherent light |
| WO2020101863A2 (en) * | 2019-10-25 | 2020-05-22 | Futurewei Technologies, Inc. | Hybrid opto-electronic neural networks and methods for utilization thereof |
| CN110929864A (en) * | 2019-12-05 | 2020-03-27 | 北京超放信息技术有限公司 | Optical diffraction neural network on-line training method and system |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113822424A (en) * | 2021-07-27 | 2021-12-21 | 湖南大学 | All-optical diffraction neural network system based on super-structured surface |
| CN113822424B (en) * | 2021-07-27 | 2023-10-20 | 湖南大学 | All-optical diffraction neural network system based on super-structured surface |
| CN117521746A (en) * | 2024-01-04 | 2024-02-06 | 武汉大学 | Quantized optical diffraction neural network system and training method thereof |
| CN117521746B (en) * | 2024-01-04 | 2024-03-26 | 武汉大学 | Quantized optical diffraction neural network system and training method thereof |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112101514B (en) | 2022-06-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP7372009B2 (en) | Neural hardware accelerator for parallel distributed tensor computation | |
| Liu et al. | Tackling photonic inverse design with machine learning | |
| CN112101514B (en) | Diffraction neural network adopting pyramid structure diffraction layer for light supplement and implementation method | |
| CN110363294A (en) | Neural network is indicated using the path in network to improve the performance of neural network | |
| CN105873744A (en) | Volume based gradient index lens by additive manufacturing | |
| CN109472858A (en) | Differentiable rendering pipeline for reverse figure | |
| WO2020074989A1 (en) | Data representation for dynamic precision in neural network cores | |
| CN105023296A (en) | 3D-printing-oriented halftone projection and model generation method | |
| Jäger et al. | Designing optimized nano textures for thin-film silicon solar cells | |
| CN111582468B (en) | Photoelectric hybrid intelligent data generation and calculation system and method | |
| Fey et al. | Optical interconnects for neural and reconfigurable VLSI architectures | |
| CN114037070A (en) | Optical signal processing method, photonic neural network chip and design method thereof | |
| Idehenre et al. | Multi-directional beam steering using diffractive neural networks | |
| Taghvaee et al. | Radiation pattern prediction for metasurfaces: A neural network-based approach | |
| Huang et al. | Deep learning enabled nanophotonics | |
| Du et al. | Metagrating-based acoustic wavelength division multiplexing enabled by deterministic and probabilistic deep learning models | |
| Chen et al. | Analytic design of light extraction array for light guide plate based on extended sources | |
| Pan et al. | Lens of controllable optical field with thin film metallic glasses for UV-LEDs | |
| CN117233960B (en) | Optical system on-line design method and device based on intelligent light calculation | |
| Li et al. | Multiscale diffractive U-Net: a robust all-optical deep learning framework modeled with sampling and skip connections | |
| CN114519403B (en) | Optical diagram neural classification network and method based on-chip diffraction neural network | |
| CN113592084B (en) | Photon neural network on chip based on reverse optimization superstructure convolution kernel | |
| CN112384935A (en) | Hierarchical parallelism in a distributed neural network core network | |
| CN107861240A (en) | The preparation method and device of curved substrate multilayer diffraction optical element | |
| WO2021110299A1 (en) | Flexible precision neural inference processing units |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |