All-optical nonlinear layer, all-optical neural network system and calculation method

Abstract

The invention relates to an all-optical non-linear layer, an all-optical neural network system and a calculation method, the all-optical non-linear layer comprises a transparent carrier and at least one light limiting material covering the transparent carrier, and the all-optical non-linear layer is used for limiting the light intensity of an incident first light signal under the condition that the light intensity of the incident first light signal is larger than a preset threshold value. Performing nonlinear transformation on the first optical signal to transmit a second optical signal; wherein under the condition that the light intensity of the first light signal is greater than a preset threshold value, the absorptivity of the light limiting material to the first light signal is increased along with the increase of the light intensity of the first light signal. According to the all-optical nonlinear layer disclosed by the embodiment of the invention, the nonlinear response of light intensity can be realized under the condition that electric or magnetic assistance is not needed, and more flexibility and possibility are provided for the construction of an optical neural network.

Description

Technical Field

[0001] The present disclosure relates to the field of optical technology, and in particular to an all-optical nonlinear layer, an all-optical neural network system and a calculation method. Background Art

[0002] With the rapid development of data-driven artificial intelligence technology, the explosive growth of tasks has brought unprecedented demands on computing power and computing speed of computing systems. However, as electronic transistor integration technology is gradually disturbed by non-classical states, it has become difficult for electronic computing systems to maintain the development trend shown by Moore's Law, which means that the computing power of electronic computing systems will be difficult to continue to grow exponentially.

[0003] In order to break through this bottleneck, relevant technologies have proposed a computing system based on photons. Among them, photons are bosons, which are less susceptible to mutual interference than fermion electrons, and have the potential to achieve large-scale parallel computing. Therefore, all-optical computing can be used as a feasible solution for high-speed and high-computing computing systems, and all-optical artificial neural networks have also become one of the goals of artificial intelligence development.

[0004] However, the implementation principles and methods of all-optical nonlinear operations in artificial neural networks need to be clarified. The implementation methods of some nonlinear layers in related technologies often need to rely on electrical technology. For example, the calculated optical signal is output to the circuit, and the computer performs nonlinear processing and calculation, and then the electrical signal is converted into an optical signal to continue optical linear calculation. This "optical-electrical-optical" approach is not conducive to the realization of all-optical neural networks and the increase of system throughput. Summary of the invention

[0005] The present disclosure proposes an all-optical nonlinear layer, an all-optical neural network system and a calculation method.

[0006] According to one aspect of the present disclosure, an all-optical nonlinear layer is provided, comprising a transparent carrier and at least one optical limiting material covered on the transparent carrier, wherein the all-optical nonlinear layer is used to perform a nonlinear transformation on a first optical signal when the optical intensity of the incident first optical signal is greater than a preset threshold value, and transmit a second optical signal, wherein when the optical intensity of the first optical signal is greater than the preset threshold value, the absorption rate of the optical limiting material on the first optical signal increases as the optical intensity of the first optical signal increases.

[0007] In a possible implementation manner, different regions of the transparent carrier are covered with different light limiting materials, and different light limiting materials correspond to different nonlinear transformation parameters.

[0008] In a possible implementation, the nonlinear transformation parameters include a first transformation parameter and a second transformation parameter, the first transformation parameter and the second transformation parameter are determined by the light limiting material, and a nonlinear transformation is performed on the first light signal to transmit a second light signal, including: performing a nonlinear transformation on the first light signal according to the first transformation parameter and the second transformation parameter to obtain the transmitted second light signal.

[0009] In a possible implementation manner, the second optical signal is positively correlated with the first transformation parameter and negatively correlated with the second transformation parameter.

[0010] In a possible implementation, the light confinement material includes quantum dots, and the first transformation parameter and the second transformation parameter are adjusted by changing at least one of the composition, structure, concentration, and surface ligands of the quantum dots.

[0011] According to one aspect of the present disclosure, an all-optical neural network system is provided, which includes at least one optical diffraction processing unit connected in series, wherein the optical diffraction processing unit includes a linear diffraction layer and the all-optical nonlinear layer as described above, and optical communication is performed between the linear diffraction layer and the all-optical nonlinear layer through an optical transmission space.

[0012] In a possible implementation, the linear diffraction layer outputs a first optical signal by diffracting coherent light loaded with information to be processed; the all-optical nonlinear layer performs nonlinear transformation on the first optical signal and transmits a second optical signal, which includes a fluorescent signal.

[0013] In a possible implementation, the linear diffraction layer includes N phase adjustment nodes, each phase adjustment node has a different diffraction angle, the all-optical nonlinear layer includes M regions, each region is covered with a different light limiting material, wherein the coherent light loaded with the information to be processed is diffracted by light of all N phase adjustment nodes, and the first optical signal is output to a single region of the M regions of the all-optical nonlinear layer, and M and N are integers greater than 1.

[0014] In one possible implementation, the light-confining material includes quantum dots, and the adjustable parameters of the all-optical neural network system include the diffraction angle of each phase adjustment node, and at least one of the composition, structure, concentration, and surface ligands of the quantum dots covered in each area.

[0015] In one possible implementation, the all-optical neural network system includes a fully connected neural network system, wherein the linear diffraction layer performs linear operations in the fully connected neural network system, and the all-optical nonlinear layer performs nonlinear operations in the fully connected neural network system.

[0016] According to one aspect of the present disclosure, there is provided a calculation method, which is applied to the all-optical neural network as described above, and the method comprises: the all-optical neural network receives coherent light loaded with information to be processed, and the information to be processed comprises at least one of audio data, image data, video data, voice data, analog data and digital data represented in an optical manner; the all-optical neural network calculates the information to be processed to obtain a calculation result, wherein the linear diffraction layer of the all-optical neural network is used to perform linear operations, and the all-optical nonlinear layer of the all-optical neural network is used to perform nonlinear operations.

[0017] In a possible implementation manner, the calculation method is used to identify the information to be processed or classify the information to be processed.

[0018] In a possible implementation, the all-optical neural network calculates the information to be processed to obtain a calculation result, including: the linear diffraction layer outputs a first light signal by diffracting coherent light loaded with the information to be processed; the all-optical nonlinear layer performs a nonlinear transformation on the first light signal and transmits a second light signal, wherein the second light signal includes a fluorescence signal; and the calculation result of the all-optical neural network is determined based on the second light signal.

[0019] The all-optical nonlinear layer of the disclosed embodiment includes a transparent carrier and at least one optical limiting material covered on the transparent carrier, wherein the all-optical nonlinear layer is used to perform a nonlinear transformation on the first optical signal when the light intensity of the incident first optical signal is greater than a preset threshold value, and transmit a second optical signal, wherein when the light intensity of the first optical signal is greater than the preset threshold value, the absorption rate of the optical limiting material on the first optical signal increases as the light intensity of the first optical signal increases. The all-optical nonlinear layer of the disclosed embodiment is used as an all-optical passive structure, which relies on light intensity for information processing and can achieve nonlinear response of light intensity without electrical or magnetic assistance. The all-optical nonlinear layer can be integrated into various optical neural network architectures that rely on light intensity for information processing. As a nonlinear layer in various optical neural networks, it provides more flexibility and possibilities for the construction of optical neural networks. Users can select a suitable all-optical nonlinear layer according to specific application requirements to achieve personalized nonlinear operations.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only and do not limit the present disclosure. Other features and aspects of the present disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used to illustrate the technical solutions of the present disclosure together with the specification.

[0022] Figure 1 A schematic diagram showing an all-optical nonlinear layer according to an embodiment of the present disclosure.

[0023] Figure 2 A schematic diagram showing curve fitting of the all-optical nonlinear layer according to an embodiment of the present disclosure.

[0024] Figure 3 A schematic diagram of an all-optical neural network system according to an embodiment of the present disclosure is shown.

[0025] Figure 4 A schematic diagram showing the prediction results of the all-optical neural network system according to an embodiment of the present disclosure.

[0026] Figure 5 A schematic diagram showing a confusion matrix according to an embodiment of the present disclosure.

[0027] Figure 6 A schematic diagram showing a comparison of the effects of an all-optical neural network system according to an embodiment of the present disclosure is shown.

[0028] Figure 7 A flow chart showing a calculation method according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0029] Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0030] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0031] The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can represent including any one or more elements selected from the set consisting of A, B, and C.

[0032] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present disclosure.

[0033] Figure 1 A schematic diagram showing an all-optical nonlinear layer according to an embodiment of the present disclosure is shown in FIG. Figure 1 As shown, the all-optical nonlinear layer 3 includes a transparent carrier 1 and at least one light limiting material 2 covered on the transparent carrier 1. The all-optical nonlinear layer 3 is used to i When the light intensity is greater than a preset threshold, the first light signal I i Perform nonlinear transformations Transmitting the second light signal I o , wherein the light confinement material 2 is i When the light intensity is greater than a preset threshold, the first light signal I i The absorption rate of the first optical signal I i increases with the increase of light intensity.

[0034] The all-optical nonlinear layer 3 of the embodiment of the present disclosure serves as an all-optical passive structure and relies on light intensity for information processing. It can realize nonlinear response of light intensity without electrical or magnetic assistance. The all-optical nonlinear layer 3 can be integrated into various optical neural network architectures that rely on light intensity for information processing. As a nonlinear layer in various optical neural networks, it provides more flexibility and possibilities for the construction of optical neural networks. Users can choose a suitable all-optical nonlinear layer 3 according to specific application requirements to realize personalized nonlinear operations.

[0035] Optionally, a light limiting material 2 may be covered on the transparent carrier 1. For example, a uniform thin film formed by a light limiting material 2 may be covered on the transparent carrier 1. In this case, the mapping relationship of each nonlinear transformation of the all-optical nonlinear layer 3 is the same.

[0036] In this case, when the all-optical nonlinear layer 3 is applied to an optical neural network, for example, the all-optical nonlinear layer 3 can be placed on the output node of the linear layer. Since the all-optical nonlinear layer 3 is composed of a single type of optical confinement material 2, the all-optical nonlinear layer 3 has the same nonlinear transformation function (also called nonlinear activation function) at all spatial positions.

[0037] Optionally, the transparent carrier 1 may be divided into a plurality of regions, for example, P rows and Q columns, each region may be covered with a light limiting material 2, for example, the first row and first column region of the transparent carrier 1 may be covered with a uniform film formed by a first light limiting material 2, the first row and second column region of the transparent carrier 1 may be covered with a uniform film formed by a second light limiting material 2, and so on, the Pth row and Qth column region of the transparent carrier 1 may be covered with a uniform film formed by a P×Qth light limiting material 2. In this case, the mapping relationship of the nonlinear transformation of the same region of the all-optical nonlinear layer 3 is the same, and the mapping relationship of the nonlinear transformation of different regions is different, and the specific mapping relationship of the nonlinear transformation of each region is determined by the light limiting material 2 covered by the region.

[0038] In this case, when the all-optical nonlinear layer 3 is applied to the optical neural network, for example, the all-optical nonlinear layer 3 can be placed on the output node of the linear layer. Since the all-optical nonlinear layer 3 uses different types of optical confinement materials 2 at different spatial positions, different spatial positions of the all-optical nonlinear layer 3 have different nonlinear transformation functions.

[0039] It should be understood that the embodiments of the present disclosure do not limit the types and quantity of the light limiting materials 2 and the specific area covered. According to the actual application scenario, any light limiting material 2 can be covered at any position of the transparent carrier 1 until the entire transparent carrier 1 is covered.

[0040] Optionally, the transparent carrier 1 includes any thin layer structure made of transparent material and having various sizes and shapes. These transparent thin layers allow light to directly penetrate without or only slightly absorb, scatter or reflect due to their transparency. Therefore, in the transparent thin layer, light can pass through the thin layer structure along a relatively straight path without losing energy or significantly losing energy due to absorption, scattering or interface reflection inside the transparent material. Among them, the transparent carrier 1 may include glass, crystal, acrylic, etc., and the embodiments of the present disclosure do not limit the type of the transparent carrier 1.

[0041] Optionally, the light confinement material 2 may include quantum dots. Quantum dots are semiconductor nanocrystals that confine excitons in three spatial directions. They are composed of only a few atoms and have nanometer-scale dimensions in all three spatial directions. These quantum dots can absorb a portion of the incident light and generate a portion of the transmitted light. When the incident first light signal I i When the intensity of the transmitted second light signal I o The intensity of the first light signal I i However, once the incident first light signal I i When the intensity exceeds the preset threshold, the absorption rate of the quantum dots will increase, resulting in the transmission of the second light signal I oThe intensity of the first light signal I i The intensity of the first light signal I i The intensity of the second optical signal I o The amplitude of the intensity enhancement gradually decreases, and finally the second optical signal I o The intensity tends to a stable value and no longer changes with the first light signal I i The light limiting mechanism is determined by the light limiting material 2 covering the transparent carrier 1.

[0042] In one possible implementation, Figure 1 As shown, the first optical signal I with a wavelength of λ i The second optical signal I with a wavelength of λ can be obtained by injecting it into the all-optical nonlinear layer 3. o . second optical signal I o At the corresponding spatial position (x, y) and the input first optical signal I i The relationship is:

[0043]

[0044] Among them, I i (x, y) represents the first optical signal I injected into the all-optical nonlinear layer 3 i The light intensity at the spatial position (x,y), I o (x, y) represents the second optical signal I transmitted by the all-optical nonlinear layer 3 o The light intensity at the spatial position (x,y), Represents a nonlinear transformation function. The specific form of is determined by the light confinement material 2 used in the all-optical nonlinear layer 3.

[0045] In a possible implementation, the light confinement mechanism of the light confinement material 2 may include reverse saturation absorption (RSA) and two-photon absorption (TPA). According to the mechanism of the light confinement material 2, when reverse saturation absorption (RSA) or two-photon absorption (TPA) is dominant, the second light signal I transmitted by the light confinement material 2 is o Light intensity and incident first light signal I i , the relationship between the light intensities can be expressed as:

[0046]

[0047] Among them, I i represents the first optical signal injected into the all-optical nonlinear layer 3, I orepresents the second optical signal transmitted by the all-optical nonlinear layer 3, z represents the distance between the test point (focusing point of the test optical system) and the tested all-optical nonlinear layer 3, α0 represents the linear absorption coefficient, β0 represents the nonlinear absorption coefficient, wherein the nonlinear absorption coefficient β0 reflects the absorption capacity of the light-limiting material 2 to high-intensity light, and the larger β0 is, the stronger the nonlinear absorption is. The linear absorption coefficient α0 reflects the absorption capacity of the light-limiting material 2 under low light intensity.

[0048] In the nonlinear working range of quantum dots, formula (2) is simplified and the output second optical signal I o With the first optical signal I i The relationship can be expressed as:

[0049]

[0050] As shown in formula (3), I i represents the first optical signal injected into the all-optical nonlinear layer 3, I o represents the second optical signal transmitted by the all-optical nonlinear layer 3. The first transformation parameter a and the second transformation parameter b are nonlinear transformation parameters. The first transformation parameter a is equivalent to The second transformation parameter b is equivalent to The first transformation parameter a and the second transformation parameter b are determined by the light confinement material 2 and are constants related to the working mechanism of the light confinement material 2. Different light confinement materials 2 may correspond to different first transformation parameters a and second transformation parameters b.

[0051] In a possible implementation manner, the first optical signal I i Perform nonlinear transformation and transmit the second optical signal I o , comprising: according to the first transformation parameter a and the second transformation parameter b, transforming the first optical signal I i Perform nonlinear transformation to obtain the transmitted second light signal I o .

[0052] Obviously, formula (3) corresponds to a nonlinear convex function, which indicates that the optical confinement mechanism of the optical confinement material 2 can be applied to optical neural networks to realize all-optical nonlinear computing.

[0053] In a possible implementation, as shown in formula (3), the second optical signal I o It is positively correlated with the first transformation parameter a and negatively correlated with the second transformation parameter b.

[0054] In the example, when the first optical signal I is injected into the all-optical nonlinear layer 3 iWhen the light intensity is greater than the preset threshold and remains unchanged, if the value of the first conversion parameter a is larger, the second light signal I transmitted by the all-optical nonlinear layer 3 is o If the value of the first conversion parameter a is smaller, the second optical signal I transmitted by the all-optical nonlinear layer 3 is o On the contrary, if the value of the second transformation parameter b is larger, the second optical signal I transmitted by the all-optical nonlinear layer 3 is o If the value of the second transformation parameter b is smaller, the second optical signal I transmitted by the all-optical nonlinear layer 3 is o The greater the light intensity.

[0055] Since the second optical signal I o The first transformation parameter a is positively correlated with the first transformation parameter a and the second transformation parameter b is negatively correlated with the second transformation parameter b. The first transformation parameter a and the second transformation parameter b are determined by the light limiting material 2. Therefore, in practical applications, the first transformation parameter a and the second transformation parameter b can be changed by adjusting the light limiting material 2 to flexibly control the second optical signal I o The output of the optical signal can be precisely controlled (for example, including the enhancement or attenuation of the optical signal). This positive and negative correlation provides more flexibility and operability for the processing of optical signals.

[0056] In a possible implementation, different regions of the transparent carrier 1 are covered with different light limiting materials 2, and different light limiting materials 2 correspond to different nonlinear transformation parameters.

[0057] Optionally, the light confinement material 2 may include quantum dots, and different regions of the transparent carrier 1 are covered with different quantum dots, and different quantum dots may correspond to different nonlinear transformation parameters. Among them, the quantum dots may include a combination of one or more of TiS2 quantum dots, TiSe2 quantum dots, TaS2 quantum dots, TaSe2 quantum dots, VS2 quantum dots, VSe2 quantum dots, NbS2 quantum dots, NbSe2 quantum dots, and PbS quantum dots, or a combination of one or more quantum dots doped with a certain element (such as Mn, Cu), and the embodiments of the present disclosure are not limited to this.

[0058] Exemplarily, the transparent carrier 1 can be divided into a plurality of regions, such as P rows and Q columns, and P×Q quantum dots can be covered in the P×Q regions respectively. For example, a uniform film formed by the first quantum dots can be covered in the first row and first column region of the transparent carrier 1, and a uniform film formed by the second quantum dots can be covered in the first row and second column region of the transparent carrier 1. By analogy, a uniform film formed by the P×Q quantum dots can be covered in the Pth row and Qth column region of the transparent carrier 1. In this case, the P×Q regions can correspond to different first transformation parameters a and second transformation parameters b, respectively, and the values ​​of the first transformation parameters a and second transformation parameters b corresponding to each region are determined by the quantum dots covered in the region.

[0059] In practical applications, since the first transformation parameter a and the second transformation parameter b are related to the composition, structure, concentration, surface ligands, etc. of the quantum dots, in order to more accurately obtain the first transformation parameter a and the second transformation parameter b, the first transformation parameter a and the second transformation parameter b can be determined by data fitting.

[0060] Figure 2 A schematic diagram showing the curve fitting of the all-optical nonlinear layer 3 according to an embodiment of the present disclosure is shown as follows: Figure 2 As shown, taking PbS (lead sulfide) quantum dots as an example, a 5 nm PbS quantum dot solution can be spin-coated on a transparent carrier 1 (e.g., glass) to form a uniform film to form an all-optical nonlinear layer 3 containing Pbs quantum dots. Then, within the nonlinear working range of the PbS quantum dots, incident light P with different light intensities can be injected into the all-optical nonlinear layer 3 containing Pbs quantum dots. in , collect the transmitted light P of the all-optical nonlinear layer 3 out , we will get multiple groups of PbS quantum dot measurement data points (P in , P out ), the measured data points (P in , P out ), and the nonlinear transformation function of PbS quantum dots is obtained, namely:

[0061]

[0062] Among them, I i represents the first optical signal injected into the all-optical nonlinear layer 3 composed of PbS quantum dots, I o It represents the second light signal transmitted by the all-optical nonlinear layer 3 composed of PbS quantum dots, 1.73393 is the first transformation parameter a of the fitted nonlinear transformation function of the PbS quantum dots, and 0.70939 is the second transformation parameter b of the fitted nonlinear transformation function of the PbS quantum dots.

[0063] Data fitting is to find the appropriate first transformation parameter a and second transformation parameter b so that formula (3) is as close as possible to or matches the actual observed measurement data point (P in , P out ), the embodiments of the present disclosure do not limit the specific data fitting method.

[0064] In a possible implementation, the light confinement material 2 may include quantum dots, and the first transformation parameter a and the second transformation parameter b are adjusted by changing at least one of the composition, structure, concentration, and surface ligands of the quantum dots.

[0065] For example, appropriate elements (such as Mn or Cu) may be doped into quantum dots to change their energy level distribution and recombination mechanism, thereby changing the magnitudes of the first transformation parameter a and the second transformation parameter b.

[0066] For another example, the concentration of quantum dots may be changed, that is, by changing the number of quantum dots in a unit space, the magnitudes of the first transformation parameter a and the second transformation parameter b may be changed.

[0067] For another example, for quantum dots with a "core-shell-shell" structure, the carrier trap can be changed by changing the shell thickness or adopting a gradient shell design, thereby changing the size of the first transformation parameter a and the second transformation parameter b.

[0068] For another example, long chains or functionalized ligands may be selected to change the effect of their surface states on the optical properties, thereby changing the magnitudes of the first transformation parameter a and the second transformation parameter b.

[0069] It can be seen that by changing at least one of the composition, structure, concentration, and surface ligands of the quantum dots, the first transformation parameter a and the second transformation parameter b can be flexibly adjusted.

[0070] In summary, by using quantum dots with different compositions, different structures, different concentrations or different surface ligands in different spatial positions of the all-optical nonlinear layer 3 of the embodiment of the present disclosure, different nonlinear transformation functions can be simulated, and applying such an all-optical nonlinear layer 3 to an optical neural network system can improve the design flexibility of the optical neural network system and further improve the complexity and characterization capability of the optical neural network system.

[0071] Figure 3 A schematic diagram of an all-optical neural network system according to an embodiment of the present disclosure is shown as follows: Figure 3 As shown, the all-optical neural network system includes at least one optical diffraction processing unit connected in series, and the optical diffraction processing unit includes a linear diffraction layer 4 and an all-optical nonlinear layer 3, and optical communication is performed between the linear diffraction layer 4 and the all-optical nonlinear layer 3 through an optical transmission space.

[0072] Optionally, the all-optical nonlinear layer 3 of the all-optical neural network system includes a transparent carrier 1, and at least one light-limiting material 2 covered on the transparent carrier 1. For details, please refer to the above and will not be repeated here.

[0073] Optionally, it is known that diffraction refers to the phenomenon that when light propagates in a straight line in the air and encounters an obstacle or a small hole, the light will deviate from the straight-line propagation path, resulting in an uneven distribution of light intensity. Any obstacle can cause diffraction of light, but obvious diffraction phenomenon will only be observed when the size of the obstacle or hole is smaller than or similar to the wavelength of light. Diffraction will produce many small waves at small apertures. These wavelets can be superimposed on each other, and the degree of mutual weakening during the superposition process can regularly become lighter or heavier, thus forming light and dark stripes. The coherent superposition of infinite continuous wavelets such as diffraction can correspond to mathematical integrals. Therefore, the phenomenon of light diffraction can be used to design the linear diffraction layer 4 of the optical neural network to realize linear multiplication and summation operations in the optical neural network.

[0074] In this way, for the signal communication between the linear diffraction layer 4 and the all-optical nonlinear layer 3, there is no need to perform photoelectric conversion and electro-optical conversion, and the transmission process of the optical signal from the linear diffraction layer 4 to the all-optical nonlinear layer 3 is carried out in the optical domain. Since this process does not require the participation of electronic conversion equipment, the conversion loss caused by the electronic conversion equipment in the related technology can be avoided, thereby improving the communication efficiency.

[0075] In one possible implementation, Figure 3 As shown, each optical diffraction processing unit may also include an input interface 5 and an output interface 6 interconnected with other optical diffraction processing units. The input interface 5 may be used to load the information to be processed on the amplitude of the spatial coherent light, and the phase of the light field of the coherent light at different spatial positions remains consistent. Among them, a laser light source coupled with an encodable device may be used as the input interface 5.

[0076] In a possible implementation, the linear diffraction layer 4 performs diffraction processing on the coherent light loaded with the information to be processed, and outputs a first optical signal I i The all-optical nonlinear layer 3 is used to treat the first optical signal I i Perform nonlinear transformation and transmit the second optical signal I o , the second optical signal I o Includes fluorescent signal.

[0077] For example, the linear diffraction layer 4 can diffract the coherent light loaded with the information to be processed in space by modulating the phase of the wavefront, and output the first optical signal I iIn an example, the linear diffraction layer 4 may include a plurality of phase adjustment nodes, each of which may serve as a neuron. Different neurons may be physically connected to a single output node through light diffraction, and the first optical signal I may be transmitted to the output node through the output node. i Transmitted to the all-optical nonlinear layer 3. The weight controlling the connection strength is determined by the phase modulation of the wavefront.

[0078] The all-optical nonlinear layer 3 can be used to generate the first optical signal I output by the linear diffraction layer 4. i Perform nonlinear transformation (also called nonlinear activation) and transmit the second optical signal I o Among them, the nonlinear transformation involves two processes, one is the mapping from the amplitude and phase information of the light field to the light intensity, and the other is the nonlinear response to the light intensity value.

[0079] The first optical signal I at all output nodes of the linear diffraction layer 4 i , the second optical signal I obtained by the nonlinear transformation of the all-optical nonlinear layer 3 o , can be transmitted to other optical diffraction processing units through the output interface 6, or the calculation result output of the entire all-optical neural network system can be realized. Among them, a photodetector can be used as the output interface 6.

[0080] The working method of the optical diffraction processing unit is described mathematically below. For the linear diffraction layer 4, it is assumed that the input light field at the kth phase adjustment node is in is the light field amplitude of the coherent light loaded with the information to be processed, and θ0 represents the constant phase value of the input plane coherent light wavefront. All light fields input to the phase adjustment node are modulated by the elements on the spatial phase plate or the spatial light modulator to control the weight of the neuron connection. According to the fully connected neuron model, the light field at the lth output node in the linear diffraction layer 4 is:

[0081]

[0082] Among them, the light field at the lth output node in the linear diffraction layer 4 is That is, the first optical signal I outputted by the linear diffraction layer 4 at the lth output node i , which can be used as the input signal of the all-optical nonlinear layer 3, is the modulation coefficient of the modulator on the phase of the coherent light field; is the connection weight between the kth phase adjustment node and the lth output node, which can be derived and calculated by the Rayleigh-Sommerfeld diffraction theory:

[0083]

[0084] Where λ is the working wavelength of the optical diffraction processing unit, Δ r represents the Euclidean distance between the kth phase adjustment node and the lth output node, and d represents the axial distance between the kth phase adjustment node and the lth output node.

[0085] According to the Rayleigh-Sommerfeld diffraction theory, the light field at the output node of the linear diffraction layer 4 can be The first optical signal I is regarded as the secondary wave source. i In each optical diffraction processing unit, after the light field at the lth output node of the linear diffraction layer 4 is nonlinearly activated by the all-optical nonlinear layer 3, the transmitted second optical signal I o Can be a fluorescent signal Right now: in, is a function of the relationship between the fluorescence intensity of the light-confining material (such as quantum dots) and the intensity of the excitation light, see the above formula (3) representing the nonlinear transformation function. The light intensity distribution of the fluorescence of the light-confining material (such as quantum dots) in space is the calculation result of the optical diffraction processing unit, which can be transmitted to the input interface 5 of the next optical diffraction processing unit through the output interface 6 of the optical diffraction processing unit.

[0086] In this way, linear light intensity response can be achieved using the linear diffraction layer 4, and nonlinear light intensity response can be achieved using the all-optical nonlinear layer 3 without the need for electrical or magnetic assistance. Through the linear diffraction layer 4 and the all-optical nonlinear layer 3, an all-optical neural network system can be constructed more flexibly.

[0087] In a possible implementation, the linear diffraction layer 4 includes N phase adjustment nodes, each of which has a different diffraction angle, so that each phase adjustment node has a different degree of phase delay for a light wave (e.g., coherent light loaded with information to be processed), thereby resulting in different light wave diffraction conditions. The all-optical nonlinear layer 3 includes M regions, each of which is covered with a different light limiting material, wherein the coherent light loaded with the information to be processed is diffracted by all N phase adjustment nodes to output the first optical signal I i To a single region among the M regions of the all-optical nonlinear layer 3, where M and N are integers greater than 1. The transmission space between the linear diffraction layer 4 and the all-optical nonlinear layer 3 needs to satisfy the complete connectivity condition, that is, under the spatial diffraction condition, the distance between the two can make the diffraction angle of the phase adjustment node (input neuron) of each linear diffraction layer 4 cover all output nodes, that is, the M regions of the all-optical nonlinear layer 3.

[0088] Exemplarily, when the coherent light loaded with the information to be processed is diffracted by all N phase adjustment nodes, the first optical signal I is output. iWhen the first optical signal I is transferred to the first region of the M regions of the all-optical nonlinear layer 3, the first optical signal I can be converted into the first optical signal I according to the nonlinear transformation parameter corresponding to the first region. i When the coherent light loaded with the information to be processed is diffracted by all N phase adjustment nodes, the first optical signal I is output. i When the first optical signal I is transmitted to the second region of the M regions of the all-optical nonlinear layer 3, the first optical signal I can be transformed according to the nonlinear transformation parameter corresponding to the second region. i By analogy, when the coherent light loaded with the information to be processed is diffracted by all N phase adjustment nodes, the first optical signal I is output. i When the first optical signal I is in the Mth region of the M regions of the all-optical nonlinear layer 3, the first optical signal I can be transformed according to the nonlinear transformation parameter corresponding to the Mth region. i Perform nonlinear transformation.

[0089] In this way, the same all-optical nonlinear layer 3 can provide different nonlinear transformations, which is more flexible.

[0090] In one possible implementation, the all-optical neural network system includes a fully-connected neural network (FCNN), wherein the linear diffraction layer 4 performs linear operations in the fully-connected neural network system, and the all-optical nonlinear layer 3 performs nonlinear operations in the fully-connected neural network system.

[0091] Exemplarily, the fully connected neural network system is a deep neural network composed of a series of fully connected layers, and the fully connected layers may include adjacent linear diffraction layers 4 and all-optical nonlinear layers 3, and any two neurons in two adjacent layers are connected.

[0092] Assuming that the linear diffraction layer 4 includes N phase adjustment nodes, each phase adjustment node serves as an input neuron, the all-optical nonlinear layer 3 may include M regions covered with different light confinement materials (such as quantum dots), each region serving as an output neuron.

[0093] It can be seen that the fully connected layer consists of M input neurons, which can be recorded as And N output neurons, which can be recorded as The weight of the connection from the kth (k is any integer from 1 to M) input neuron to the lth (l is any integer from 1 to N) output neuron is Then, the lth output neuron can output Among them, the weight is As can be seen from the above formula (6), the nonlinear transformation function Please refer to the above formula (3), which will not be elaborated here.

[0094] In this way, an all-optical fully connected neural network system can be realized. Since the input and output can be directly connected in the fully connected neural network system, the optical signal can be efficiently transmitted in the network, thereby providing a larger network throughput. Moreover, in a fully connected neural network, the failure of a link will only result in the loss of a node, without affecting the operation of the entire network, and has high reliability.

[0095] Exemplarily, an all-optical neural network system (for example, including 5 optical diffraction processing units) can be constructed to verify the function of the optical diffraction processing unit and the effectiveness of the all-optical nonlinear layer 3. In order to ensure the physical feasibility of the proposed architecture, the parameters of each part of the optical diffraction processing unit are designed with full consideration of the current manufacturing capabilities. The linear diffraction layer 4 of each optical diffraction processing unit can have 100×100 input and output nodes. The size of the linear diffraction layer 4 is 63.3μm, and the size of each node is 316.5nm. The distance between the linear diffraction layer 4 and the all-optical nonlinear layer 3 is set to 9.495μm. The input interface 5 of the optical diffraction processing unit can be a laser with a programmable beam expansion system, and the output interface 6 can be realized by a photodetector. Among them, the wavelength of the laser beam used can be 633nm.

[0096] It should be understood that the above parameters are only examples and the parameters can be set according to actual application scenarios, and the embodiments of the present disclosure are not limited to this.

[0097] The classification result of the all-optical neural network system corresponds to the probability distribution of the area with the highest light intensity in the output layer (the output of the terminal optical diffraction processing unit). The method adopted is to maximize the normalized signal in the detector area corresponding to the target while minimizing the total signal outside the target area. This is achieved by using the mean square error (MSE) of the output and target images as the loss function.

[0098] In one possible implementation, the light-confining material includes quantum dots, and the adjustable parameters of the all-optical neural network system include the diffraction angle of each phase adjustment node, and at least one of the composition, structure, concentration, and surface ligands of the quantum dots covered in each area.

[0099] For example, based on the loss value obtained by the loss function, the diffraction angle of each phase adjustment node in the linear diffraction layer 4 can be adjusted to optimize the linear transformation parameters of the entire linear diffraction layer 4.

[0100] For example, based on the loss value obtained by the loss function, at least one of the composition, structure, concentration, and surface ligands of the quantum dots covered in each area of ​​the all-optical nonlinear layer 3 can be adjusted to optimize the nonlinear transformation parameters of the all-optical nonlinear layer 3. For example, for quantum dots with a "core-shell-shell" structure, carrier traps can be reduced by increasing the shell thickness or adopting a gradient shell design. For another example, appropriate elements (such as Mn or Cu) can be doped into quantum dots to change their energy level distribution and recombination mechanism. For another example, long chains or functionalized ligands can be selected to reduce the impact of the surface state of quantum dots on optical properties. It should be understood that the embodiments of the present disclosure do not limit the specific method for optimizing nonlinear transformation parameters.

[0101] In this way, the linear transformation parameters and nonlinear transformation parameters of the all-optical neural network system can be adjusted during the training process until the training is completed, and an optimized all-optical neural network system is obtained.

[0102] The all-optical neural network system of the disclosed embodiment can perform various complex tasks, and its effectiveness is verified by taking the handwritten digit classification task of a preset database as an example. For example, the all-optical neural network system can use a database consisting of 60,000 training images and 10,000 test images, each with a resolution of 28×28 and belonging to one of the 10 preset categories. The training goal is set to map the input images of the numbers "1" to "10" to 10 predefined areas of the output layer of the last optical diffraction processing unit, respectively, and determine the classification result by finding the target area with the highest illumination intensity.

[0103] During the training process, the random gradient descent algorithm and the adaptive matrix estimation optimizer (Adaptive Moment Estimation, Adam) can be used to optimize the network parameters (such as the linear transformation parameters of the linear diffraction layer 4 and the nonlinear transformation parameters corresponding to each region in the all-optical nonlinear layer 3) to minimize the loss function. The learning rate can be set to 0.0005, and the training batch size is 200. Based on the above network design and recognition tasks, the training phase requires 100 times to obtain a converged model, and the training time is about 2 hours.

[0104] Figure 4 A schematic diagram showing the prediction results of the all-optical neural network system according to an embodiment of the present disclosure, as shown in Figure 4 As shown in the figure, the constructed multi-layer all-optical neural network system achieved a blind test accuracy of 91.58% on 10,000 images in the digital test set. This confirms the effectiveness of the all-optical neural network system. Taking the numbers "0", "1", "4", and "8" in the database as an example, the output image of each optical diffraction processing unit (i.e., the output light intensity distribution pattern of the all-optical nonlinear layer 3) is shown in Figure 4As shown in FIG. 1 , the output image of the last optical diffraction processing unit is the output result of the all-optical neural network system. Ten detection areas are set in the output result frame to extract the light intensity value, and the area number with the highest light intensity value corresponds to the predicted label.

[0105] Figure 5 A schematic diagram of a confusion matrix according to an embodiment of the present disclosure is shown as follows: Figure 5 As shown in the figure, the predicted category shown in the figure is consistent with the label of the actual category, which can successfully complete the handwritten digit image classification task.

[0106] Figure 6 A schematic diagram showing the effect comparison of the all-optical neural network system according to an embodiment of the present disclosure is shown as follows: Figure 6 As shown in the figure, the performance of neural networks using and not using light-confined materials (such as quantum dots) in classifying handwritten images is compared while keeping the network structure unchanged. The test accuracy of the two networks on this task can be found in Figure 6 The curve in shows that the classification accuracy of the network can be improved by 5.69% using the all-optical nonlinear layer 3. This shows that the all-optical nonlinear layer 3 can enhance the representation and generalization of the optical neural network.

[0107] In summary, the all-optical nonlinear layer 3 based on the optical confinement mechanism material (such as quantum dots that conform to the optical confinement mechanism) provided by the embodiment of the present disclosure can realize the nonlinear response of light intensity as a nonlinear layer in the optical neural network. For example, the all-optical nonlinear layer 3 can be coupled with various optical neural networks, and based on the all-optical nonlinear layer 3 and the diffraction computing architecture, an all-optical neural network system composed of at least one optical diffraction processing unit in series is designed. The simulation results show that the blind test accuracy of the all-optical neural network system on the digital test set is very high (for example, an accuracy of 91.58%), and the recognition accuracy of the all-optical neural network system composed of the all-optical nonlinear layer 3 is 5.69% higher than that of the all-optical neural network system without the all-optical nonlinear layer 3. The all-optical nonlinear layer 3 provides reliable optical nonlinear activation and has the potential to be integrated with various optical computing architectures today. The all-optical nonlinear layer 3 can help the optical computing system develop in the all-optical direction, making it possible to have a computing system with huge arithmetic power and excellent computing speed.

[0108] Figure 7 A flow chart showing a calculation method according to an embodiment of the present disclosure is shown as follows: Figure 7 As shown, the calculation method is applied to the all-optical neural network as described above, and the method includes:

[0109] In step S11, the all-optical neural network receives coherent light loaded with information to be processed, wherein the information to be processed includes at least one of audio data, image data, video data, voice data, analog data, and digital data represented in an optical manner;

[0110] In step S12, the all-optical neural network calculates the information to be processed to obtain a calculation result, wherein the linear diffraction layer of the all-optical neural network is used to perform linear operations, and the all-optical nonlinear layer of the all-optical neural network is used to perform nonlinear operations.

[0111] In a possible implementation manner, the calculation method is used to identify the information to be processed or classify the information to be processed.

[0112] In a possible implementation, step S12 may include: the linear diffraction layer outputs a first light signal by diffracting coherent light loaded with information to be processed; the all-optical nonlinear layer performs nonlinear transformation on the first light signal and transmits a second light signal, wherein the second light signal includes a fluorescence signal; and based on the second light signal, determines a calculation result of the all-optical neural network.

[0113] It can be understood that the above-mentioned various method embodiments mentioned in the present disclosure can be combined with each other to form a combined embodiment without violating the principle logic. Due to space limitations, the present disclosure will not repeat them. It can be understood by those skilled in the art that in the above-mentioned method of the specific implementation method, the specific execution order of each step should be determined according to its function and possible internal logic.

[0114] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0115] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0116] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0117] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

[0118] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. An all-optical nonlinear layer, characterized in that: The all-optical nonlinear layer comprises a transparent carrier and at least one light limiting material covered on the transparent carrier. The all-optical nonlinear layer is used to perform nonlinear transformation on the first optical signal when the light intensity of the incident first optical signal is greater than a preset threshold, and transmit a second optical signal. Wherein, when the light intensity of the first light signal is greater than a preset threshold, the absorption rate of the light confinement material to the first light signal increases as the light intensity of the first light signal increases.

2. The all-optical nonlinear layer according to claim 1, characterized in that: Different areas of the transparent carrier are covered with different light limiting materials, and different light limiting materials correspond to different nonlinear transformation parameters.

3. The all-optical nonlinear layer according to claim 1 or 2, characterized in that: The nonlinear transformation parameters include a first transformation parameter and a second transformation parameter, wherein the first transformation parameter and the second transformation parameter are determined by the light confinement material. Performing a nonlinear transformation on the first optical signal to transmit a second optical signal includes: According to the first transformation parameter and the second transformation parameter, a nonlinear transformation is performed on the first optical signal to obtain a transmitted second optical signal.

4. The all-optical nonlinear layer according to claim 3, characterized in that: The second optical signal is positively correlated with the first transformation parameter and negatively correlated with the second transformation parameter.

5. The all-optical nonlinear layer according to claim 3, characterized in that: The light confinement material includes quantum dots, and the first transformation parameter and the second transformation parameter are adjusted by changing at least one of the composition, structure, concentration, and surface ligands of the quantum dots.

6. An all-optical neural network system, characterized in that: The all-optical neural network system includes at least one optical diffraction processing unit connected in series, and the optical diffraction processing unit includes a linear diffraction layer and an all-optical nonlinear layer as described in any one of claims 1 to 5, and optical communication is performed between the linear diffraction layer and the all-optical nonlinear layer through an optical transmission space.

7. The all-optical neural network system according to claim 6, characterized in that: The linear diffraction layer performs diffraction processing on the coherent light loaded with the information to be processed, and outputs a first optical signal; The all-optical nonlinear layer performs nonlinear transformation on the first optical signal and transmits a second optical signal, where the second optical signal includes a fluorescent signal.

8. The all-optical neural network system according to claim 7, characterized in that: The linear diffraction layer includes N phase adjustment nodes, each of which has a different diffraction angle, and the all-optical nonlinear layer includes M regions, each of which is covered with a different light limiting material. The coherent light loaded with the information to be processed is diffracted by all N phase adjustment nodes to output the first optical signal to a single region of the M regions of the all-optical nonlinear layer, where M and N are integers greater than 1.

9. The all-optical neural network system according to claim 8, characterized in that: The light-confining material includes quantum dots, and the adjustable parameters of the all-optical neural network system include the diffraction angle of each phase adjustment node, and at least one of the composition, structure, concentration, and surface ligands of the quantum dots covered in each area.

10. The all-optical neural network system according to claim 6, characterized in that: The all-optical neural network system includes a fully connected neural network system, wherein the linear diffraction layer performs linear operations in the fully connected neural network system, and the all-optical nonlinear layer performs nonlinear operations in the fully connected neural network system.

11. A calculation method, characterized in that: The calculation method is applied to the all-optical neural network according to any one of claims 6 to 10, and the calculation method comprises: The all-optical neural network receives coherent light loaded with information to be processed, wherein the information to be processed includes at least one of audio data, image data, video data, voice data, analog data, and digital data represented in an optical manner; The all-optical neural network calculates the information to be processed to obtain a calculation result, wherein the linear diffraction layer of the all-optical neural network is used to perform linear operations, and the all-optical nonlinear layer of the all-optical neural network is used to perform nonlinear operations.

12. The calculation method according to claim 11, characterized in that: The calculation method is used for identifying the information to be processed or classifying the information to be processed.

Images