Nonlinear activation layer and nonlinear activation method for optical neural networks

Abstract

The embodiment of the application provides a kind of nonlinear activation layer and nonlinear activation method for optical neural network, it is related to optical signal processing technical field.Nonlinear activation layer includes: binary nonlinear activation device, containing the deflection of two-dimensional material liquid crystal under the action of electric field;Polarizer, polarizer is arranged at the first side of binary nonlinear activation device;Analyser, analyser is arranged at the second side of binary nonlinear activation device, second side is the side opposite to first side, the polarization direction of analyser is perpendicular to the polarization direction of polarizer;Wherein, linearly polarized light is incident polarizer, polarizer filters linearly polarized light, filtered linearly polarized light is incident binary nonlinear activation device, and the deflection of two-dimensional material liquid crystal is modulated to filtered linearly polarized light, modulated linearly polarized light is incident analyser, to realize optical nonlinear activation.The nonlinear activation layer of the application has dynamic adjustable control characteristics, and is suitable for all-optical architecture.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical signal processing, and in particular to a nonlinear activation layer for an optical neural network and a nonlinear activation method. BACKGROUND

[0002] An optical neural network is an architecture paradigm that uses optical devices such as optical modulators and optical detectors to realize the function of computing and reasoning by transmitting and processing parameters such as amplitude, phase and frequency of light. Compared with a traditional electrical neural network, the optical neural network has advantages such as low energy consumption, low delay, parallelism and anti-interference because the transmission carrier of the signal is a photon, and has received high attention from the academic and industrial circles. A nonlinear activation layer is one of the core components of an optical neural network and is used for feature learning and representation. However, common optical devices such as a spatial light modulator and a digital micromirror used in a free-space optical neural network cannot be used as a nonlinear activation layer because they do not have a nonlinear optical effect. At present, there are two common technical paths applied to a free-space optical neural network to realize nonlinear activation. One is to convert an optical signal into an electrical signal through a photodiode or an electronic circuit to indirectly realize nonlinear activation. However, this method cannot realize the processing and calculation of an optical signal in the whole link, increases the response time and loses the rich information carried by the optical signal. The other is to realize nonlinear activation by using the absorption and transmission properties of a nonlinear optical material. However, the current nonlinear optical activation is mainly based on light intensity coding and cannot realize negative value activation or only has an odd function activation form about the origin, which makes it difficult for the network to capture the real distribution of data and may lose important features. In addition, the nonlinear optical material only has fixed response characteristics, and the activation mode cannot be dynamically regulated, so it cannot adapt to various and more complex learning tasks and network architectures.

[0003] In summary, how to develop a nonlinear activation layer with a dynamically adjustable characteristic and suitable for a full optical architecture has become a technical problem to be solved. SUMMARY

[0004] The main purpose of the embodiments of the present application is to propose a nonlinear activation layer for an optical neural network and a nonlinear activation method. The nonlinear activation layer has a dynamically adjustable characteristic and is suitable for a full optical architecture.

[0005] To achieve the above purpose, a first aspect of the embodiments of the present application proposes a nonlinear activation layer for an optical neural network, which comprises:

[0006] A binary nonlinear activation device, which comprises a two-dimensional material liquid crystal that deflects under the action of an electric field;

[0007] A polarizer, which is arranged on a first side of the binary nonlinear activation device.

[0008] a polarizer disposed on a second side of the binary nonlinear active device, the second side being opposite to the first side, a polarization direction of the polarizer being perpendicular to the polarization direction of the polarizer;

[0009] wherein linearly polarized light is incident on the polarizer, the polarizer filters the linearly polarized light, the filtered linearly polarized light is incident on the binary nonlinear active device, the filtered linearly polarized light is modulated by the deflected two-dimensional material liquid crystal, and the modulated linearly polarized light is incident on the polarizer to achieve optical nonlinear activation.

[0010] Optionally, the binary nonlinear active device includes a first nonlinear active device and a second nonlinear active device with perpendicular electric field directions, the polarizer includes a first polarizer and a second polarizer with perpendicular polarization directions, the polarizer includes a first polarizer and a second polarizer with perpendicular polarization directions, the polarization direction of the first polarizer is the same as the polarization direction of the second polarizer, and the polarization direction of the second polarizer is the same as the polarization direction of the first polarizer.

[0011] The nonlinear active layer further includes:

[0012] a light splitting prism disposed on one side of the first nonlinear active device and on one side of the second nonlinear active device;

[0013] wherein linearly polarized light is incident on the light splitting prism, the light splitting prism splits the linearly polarized light into first linearly polarized light and second linearly polarized light, the polarization direction of the first linearly polarized light is perpendicular to the polarization direction of the second linearly polarized light; the first linearly polarized light is incident on the first polarizer, and the second linearly polarized light is incident on the second polarizer.

[0014] Optionally, the first nonlinear active device is used for positive modulation, the second nonlinear active device is used for negative modulation, the included angle between the polarization direction of the first polarizer and the electric field direction of the first nonlinear active device is 45°, and the included angle between the polarization direction of the second polarizer and the electric field direction of the second nonlinear active device is 45°.

[0015] If the polarization azimuth angle of the linearly polarized light is 45°, the first linearly polarized light is incident on the first polarizer, and then sequentially passes through the first nonlinear active device and the first polarizer.

[0016] If the polarization azimuth angle of the linearly polarized light is -45°, the second linearly polarized light is incident on the second polarizer, and then sequentially passes through the second nonlinear active device and the second polarizer.

[0017] Optionally, the nonlinear activation layer further comprises: a quarter-wave plate, the quarter-wave plate is arranged on one side of the light splitting prism,

[0018] Wherein, the circularly polarized light is incident on the quarter-wave plate, and the quarter-wave plate is used to convert the circularly polarized light into linearly polarized light.

[0019] Optionally, the first nonlinear activation device is used for positive modulation, the second nonlinear activation device is used for negative modulation, the angle between the polarization direction of the first polarizer and the electric field direction of the first nonlinear activation device is 45°, and the angle between the polarization direction of the second polarizer and the electric field direction of the second nonlinear activation device is 45°.

[0020] If the phase difference of the circularly polarized light is π / 2, then after the first linearly polarized light is incident on the first polarizer, it successively passes through the first nonlinear activation device and the first polarizer.

[0021] If the phase difference of the linearly polarized light is -π / 2, then after the second linearly polarized light is incident on the second polarizer, it successively passes through the second nonlinear activation device and the second polarizer.

[0022] Optionally, the first nonlinear activation device and the second nonlinear activation device comprise a Kerr cell of two-dimensional material liquid crystal.

[0023] Optionally, the Kerr cell of two-dimensional material liquid crystal comprises:

[0024] A Kerr cell container, which is internally provided with a functional layer containing two-dimensional material liquid crystal;

[0025] Two electrodes, the two electrodes are respectively arranged on opposite sides of the Kerr cell container;

[0026] Wherein, the two electrodes are connected to a power supply to apply an electric field to the functional layer, and the orientation of the two-dimensional material liquid crystal is deflected under the action of the electric field.

[0027] Optionally, the two-dimensional material liquid crystal is a two-dimensional vermiculite aqueous dispersion liquid, which can realize π / 2 phase modulation under the action of an electric field.

[0028] To achieve the above-mentioned purpose, a second aspect of the embodiment of the present application proposes a nonlinear activation method for an optical neural network, applied to the nonlinear activation layer as described in the first aspect, the method comprises:

[0029] Controlling linearly polarized light to be incident on the polarizer, filtering the linearly polarized light by the polarizer, and filtering the linearly polarized light to be incident on the nonlinear activation device;

[0030] An electric field is applied to the nonlinear activation device to modulate the filtered linearly polarized light by the two-dimensional material liquid crystal deflected under the action of the electric field, and the modulated linearly polarized light is incident on the polarizer to realize optical nonlinear activation.

[0031] Optionally, before the control linearly polarized light is incident on the polarizer, the method further comprises:

[0032] The circularly polarized light is controlled to be incident on a quarter-wave plate, and the circularly polarized light is converted into linearly polarized light by the quarter-wave plate.

[0033] The application provides a nonlinear activation layer and a nonlinear activation method for an optical neural network. The nonlinear activation process comprises the following steps: linearly polarized light is incident on a polarizer, the polarizer filters the linearly polarized light, the filtered linearly polarized light is incident on a binary nonlinear activation device, the filtered linearly polarized light is modulated by two-dimensional material liquid crystal deflected, and the modulated linearly polarized light is incident on a polarizer to realize optical nonlinear activation. The process utilizes an optical path formed by the polarizer, the binary nonlinear activation device and the polarizer, and utilizes the fact that the binary nonlinear activation device can regulate the transmittance of different incident light by changing the electric field intensity, can follow Malus' law and realize reconfigurable optical activation. Moreover, the two-dimensional material liquid crystal has a large electro-optic Kerr coefficient, and the phase delay is positively correlated with the electro-optic Kerr coefficient, so that the binary nonlinear activation device can realize full-precision full modulation of the phase in a small electric field range. In addition, the two-dimensional material will not be photodegraded, and has good stability to ultraviolet irradiation, so that it can be used for ultraviolet computing. In conclusion, the nonlinear activation layer has dynamic regulation characteristics and is suitable for all-optical architecture, can improve the performance of the optical neural network in the face of complex tasks, and thus promotes the development of the field of optical neural networks. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 FIG. 1 is a structural schematic diagram of a nonlinear activation layer provided by an embodiment of the application;

[0035] Figure 2 FIG. 4 is a photo of a two-dimensional vermiculite liquid crystal provided by an embodiment of the application;

[0036] Figure 3 FIG. 6 is a binary nonlinear activation device with a diagonal size of 46.1 mm provided by an embodiment of the application;

[0037] Figure 4 FIG. 8 is a phase full modulation effect diagram of a two-dimensional vermiculite aqueous dispersion liquid with a concentration of 9 g / L before and after 840 hours under different electric field intensities provided by an embodiment of the application;

[0038] Figure 5The modulation effect diagram of the two-dimensional vermiculite water phase dispersion liquid with a concentration of 9 g / L provided by the embodiment of the application before and after 12 hours of strong ultraviolet light irradiation is as follows:

[0039] Figure 6 The phase modulation comparison diagram of the two-dimensional vermiculite liquid crystal and the two-dimensional graphene oxide liquid crystal with a concentration of 9 g / L provided by the embodiment of the application in the same electric field intensity range is as follows:

[0040] Figure 7 The binary nonlinear activation device with a diagonal size of 127.3 mm provided by the embodiment of the application is as follows:

[0041] Figure 8 The structure diagram of the nonlinear activation layer for linearly polarized light provided by the embodiment of the application is as follows:

[0042] Figure 9 The structure diagram of the nonlinear activation layer for circularly polarized light provided by the embodiment of the application is as follows:

[0043] Figure 10 The comparison diagram of the binary activation effect and the linear rectifier function activation effect of the two-dimensional vermiculite liquid crystal with a concentration of 9 g / L of the application is as follows. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the application clearer and more apparent, the application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0045] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a manner different from the module division in the device or the order in the flowchart. The terms "first", "second", and the like in the specification and claims and the above-described drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs. The terms used herein are only for the purpose of describing the embodiments of the application and are not intended to limit the application.

[0047] First, the several terms involved in the application are analyzed:

[0048] Optical neural network: is a new type of network formed by using optical technology, such as optical connection technology and optical device technology. It has the ability of super parallel processing and transmission of information, high density lead ability and the unique advantage of being able to directly process images. Its basic component unit is optical neuron device and optical synapse device, among which optical neural chip is crucial. Similar to electrical neural network, optical neural network also needs nonlinear activation processing, so it is necessary to set a nonlinear activation layer for optical neural network.

[0049] Nonlinear activation layer for optical neural network: mainly introduces nonlinear characteristics, so as to be able to process more complex tasks and data patterns.

[0050] The embodiments of the present application provide a nonlinear activation layer and a nonlinear activation method for an optical neural network, which are specifically described through the following embodiments. First, the nonlinear activation layer for the optical neural network in the embodiments of the present application is described.

[0051] Referring to Figure 1 In an embodiment, the nonlinear activation layer comprises:

[0052] The binary nonlinear activation device 100 comprises a two-dimensional material liquid crystal which deflects under the action of an electric field.

[0053] The polarizer 200 is arranged on the first side of the binary nonlinear activation device 100.

[0054] The analyzer 300 is arranged on the second side of the binary nonlinear activation device 100, and the second side is opposite to the first side. The polarization direction of the analyzer 300 is perpendicular to the polarization direction of the polarizer 200.

[0055] The process of nonlinear activation can include that linearly polarized light X is incident on the polarizer 200, the polarizer 200 filters the linearly polarized light X, the filtered linearly polarized light is incident on the binary nonlinear activation device 100, the filtered linearly polarized light is modulated by the two-dimensional material liquid crystal which deflects, the modulated linearly polarized light is incident on the analyzer 300, and the activated linearly polarized light is obtained. In this way, optical nonlinear activation can be realized.

[0056] It should be noted that the azimuth angle of the linearly polarized light can be positive or negative. When the azimuth angle of the linearly polarized light is positive, the electric field is specifically the first electric field. When the azimuth angle of the linearly polarized light is negative, the electric field is specifically the second electric field. The electric field intensity of the first electric field is different from that of the second electric field, so the deflection of the two-dimensional material liquid crystal under the first electric field is different from that under the second electric field. By applying the first electric field, positive modulation of the filtered linearly polarized light is realized. By applying the second electric field, negative modulation of the filtered linearly polarized light is realized. The present embodiment can at least overcome the following technical obstacles: current nonlinear optical activation is mainly based on light intensity coding, which cannot realize negative activation and only has odd function activation form with point symmetry, resulting in that the network is difficult to capture the real distribution of data and important features may be lost.

[0057] In an embodiment, the binary nonlinear activation device 100 is a first nonlinear activation device for positive modulation, or a second nonlinear activation device for negative modulation. The first nonlinear activation device and the second nonlinear activation device both include a Kerr box of two-dimensional material liquid crystal. After the linearly polarized light filtered by the polarizer 200, the filtered linearly polarized light is incident on the first side of the Kerr box, and is transmitted from the second side of the Kerr box to obtain the modulated linearly polarized light. The second side of the Kerr box is opposite to the first side of the Kerr box.

[0058] It should be noted that the Kerr box is an important device realized in optics by the Kerr effect. The Kerr effect refers to the phenomenon that the refractive index of certain materials changes under the action of an electric field. Generally, the change of the refractive index is proportional to the square of the electric field strength. This effect causes the phase and amplitude of the optical signal to change when passing through the material. The structure of the Kerr box: composed of a two-dimensional material liquid crystal with Kerr effect. By applying an electric field, the refractive index of the material can be controlled to modulate the optical signal.

[0059] In an embodiment, referring to Figure 1 , the Kerr box of the two-dimensional material liquid crystal includes: a Kerr box container 110, which is internally provided with a functional layer containing two-dimensional material liquid crystal; and two electrodes 120, which are respectively arranged on opposite sides of the Kerr box container 110. Specifically, the two electrodes are connected to a power supply to apply an electric field to the functional layer, and the orientation of the two-dimensional material liquid crystal in the functional layer is deflected under the action of the electric field.

[0060] It should be noted that the filtered polarized light is incident on one side of the Kerr box of the two-dimensional material liquid crystal, the orientation of the two-dimensional material liquid crystal in the functional layer is deflected under the action of the electric field, the filtered polarized light passes through the two-dimensional material liquid crystal to produce a phase delay, and is transmitted from the Kerr box. Since the binary nonlinear activation device uses photons as the transmission carrier of the signal, the calculation speed v can be expressed as where c is the speed of light in free space, n is the refractive index of the medium to the frequency electromagnetic wave, the response time of the binary nonlinear active device is only picoseconds. The binary nonlinear active device can modulate the light intensity of the transmitted light based on the following rule, that is where I0 is the incident light, θ is the angle between the two-dimensional material liquid crystal optical axis and the filtered polarized light, δ is the phase delay, so the binary nonlinear active device can control the transmittance of different incident light by changing the electric field intensity, and realize the reconfigurable optical activation. The two-dimensional material liquid crystal has a large electro-optic Kerr coefficient K, and the phase delay δ = 2πLKE 2 where L is the optical path perpendicular to the electric field direction, E is the electric field intensity, so the binary nonlinear active device can realize full-precision full modulation of the phase in a small electric field range. In addition, the two-dimensional material will not be photodegraded, and has good stability to ultraviolet radiation, and can be used for ultraviolet computing.

[0061] It should be noted that the binary nonlinear active device can be used for positive modulation and negative adjustment, and different electric field intensities can be set. In order to reduce the setting complexity, a first nonlinear active device for positive modulation and a second nonlinear active device for negative modulation can be specially set. The first nonlinear active device and the second nonlinear active device both belong to the binary nonlinear active device.

[0062] In an embodiment, the two-dimensional material liquid crystal includes inorganic two-dimensional material and dispersant.

[0063] In another embodiment, the two-dimensional material liquid crystal is a two-dimensional vermiculite aqueous dispersion, which can realize π / 2 phase modulation under the action of an electric field. In this way, the adjustment accuracy is improved, thereby improving the nonlinear activation accuracy.

[0064] The two-dimensional vermiculite aqueous dispersion refers to a water dispersion of two-dimensional vermiculite material. In an example, the dispersion concentration of the two-dimensional vermiculite material is 9 g / L. The optical photograph of the two-dimensional vermiculite material liquid crystal is as shown in Figure 2 .

[0065] In an example, the Kerr box container 110 can adopt an isotropic cuvette with an external size of 12.5*12.5*45 mm. The electrode 120 can adopt a 45*10*0.06 mm double-sided conductive copper foil tape. After the Kerr box container 110 is added with the water dispersion of the two-dimensional vermiculite material, a teflon cover is used to seal the opening of the cuvette, and a binary nonlinear active device with a diagonal line size of 46.1 mm is obtained, as shown in Figure 3 .

[0066] In one embodiment, this embodiment provides a method for determining whether the binary nonlinear activation device in the above example has full phase modulation capability. Specifically, since the modulation object is an analog optical signal, full precision modulation can be achieved by reducing the electric field interval. After 35 days (840 hours) of testing, the phase modulation intervals of the binary nonlinear activation device with a test duration of 0 hours and 840 hours are basically the same, as shown in FIG. Figure 4 As shown, it has good cycle stability. Figure 4 In the figure, the abscissa is the electric field intensity, and the ordinate is the optical activation.

[0067] In one embodiment, this embodiment provides a method for measuring the UV stability of the binary nonlinear active device in the above example. Specifically, at 4.1×10 7 mW m -2 Under the irradiation of extremely strong ultraviolet light of the same level, the modulation effect of the binary nonlinear activation device was tested every 3 hours. After a total of 12 hours, the modulation effect of the binary nonlinear activation device was attenuated by only 1.29%. Figure 5 As shown, it can meet the application requirements under ultraviolet light. Figure 5 In the figure, the abscissa is the electric field intensity, and the ordinate is the optical activation.

[0068] In a comparative example, this comparative example provides a Kerr cell of a two-dimensional material liquid crystal, Figure 3 The difference between the corresponding examples is that the material of the functional layer is different, that is, the functional layer material is a two-dimensional graphene oxide aqueous dispersion. In the same electric field strength range, the phase modulation effect of the comparative example is tested, such as Figure 6 As shown. Figure 6 In the figure, the horizontal axis represents electric field intensity, and the vertical axis represents optical activation. The test results show that the two-dimensional vermiculite aqueous dispersion has a higher response sensitivity and can achieve full phase modulation within a small electric field intensity range.

[0069] In another example, the Kerr cell container 110 can be a custom container made of isotropic material with an outer dimension of 95*15*92.5 mm, resulting in a binary nonlinear activation device with a diagonal dimension of 127.3 mm (5 inches), such as Figure 7 shown.

[0070] In another example, the Kerr cell container 110 may have an outer dimension of 1*0.04*1 mm, resulting in a binary nonlinear activation device with a diagonal dimension of 1.4 mm.

[0071] In one embodiment, the electric field applied to the functional layer is an alternating current electric field. Specifically, the waveform of the alternating current electric field can be a sine wave, a square wave, etc.; the frequency of the alternating current electric field can be 10 1 Hz to 10 7Hz; the intensity of the alternating electric field can be selected as 10 2 V / m to 10 8 V / m.

[0072] In an embodiment, the alternating electric field described above can be a sinusoidal alternating electric field, the frequency of the alternating electric field can be 10 4 Hz, and the intensity of the alternating electric field can be 10 3 V / m.

[0073] In an embodiment, the feature size of the binary nonlinear activation device 100 can be adjustable, specifically, the diagonal size can be in the order of microns to centimeters.

[0074] In an embodiment, the diagonal size of the binary nonlinear activation device 100 can be 1.4 mm, 46.1 mm, or 127.3 mm (5 inches).

[0075] In an embodiment, the incident direction of the linearly polarized light is selected as the third side of the Kerr cell, i.e. the adjacent surface of the installed electrode. The incident angle of the linearly polarized light is selected as perpendicular to the surface of the Kerr cell. Specifically, the linearly polarized light is orthogonal to the direction of the applied electric field in the functional layer.

[0076] In an embodiment, the linearly polarized light can be selected as a laser with a wavelength of 300 nm-700 nm. For example, the wavelength of the linearly polarized light can be 303 nm, 450 nm, and 532 nm.

[0077] In an embodiment, with reference to Figure 1 and Figure 8The binary nonlinear activation device 100 includes a first nonlinear activation device 1001 and a second nonlinear activation device 1002 with electric field directions perpendicular to each other, the polarizer 200 includes a first polarizer 2001 and a second polarizer 2002 with polarization directions perpendicular to each other, and the analyzer 300 includes a first analyzer 3001 and a second analyzer 3002 with polarization directions perpendicular to each other. The polarization direction of the first polarizer 2001 is the same as the polarization direction of the second analyzer 3002, and the polarization direction of the second polarizer 2002 is the same as the polarization direction of the first analyzer 3001. The nonlinear activation layer further includes a beam splitter prism 400 arranged on one side of the first nonlinear activation device 1001 and on one side of the second nonlinear activation device 1002. The linearly polarized light is incident on the beam splitter prism 400, and the beam splitter prism 400 splits the linearly polarized light into first linearly polarized light and second linearly polarized light with polarization directions perpendicular to each other. The first linearly polarized light is incident on the first polarizer 2001, and the second linearly polarized light is incident on the second polarizer 2002. Regardless of whether the azimuth angle of the polarized light is positive or negative, the polarized light will be split by the beam splitter prism 400 into first linearly polarized light and second linearly polarized light with the same positive azimuth angle or the same negative azimuth angle, and will be incident on the first polarizer and the second polarizer, respectively. The first linearly polarized light with a positive azimuth angle can be transmitted through the first polarizer 2001, but the first linearly polarized light with a negative azimuth angle cannot be transmitted through the first polarizer 2001 (the polarization direction of the first linearly polarized light with a negative azimuth angle is perpendicular to the polarization direction of the first polarizer, so it cannot be transmitted through the first polarizer), and the second linearly polarized light with a negative azimuth angle can be transmitted through the second polarizer 2002, but the second linearly polarized light with a positive azimuth angle cannot be transmitted through the second polarizer 2002 (the polarization direction of the second linearly polarized light with a positive azimuth angle is perpendicular to the polarization direction of the second polarizer, so it cannot be transmitted through the second polarizer). The embodiment has the advantage of automatic positive and negative differentiation.

[0078] The first linearly polarized light and the second linearly polarized light have the same light intensity in addition to the perpendicular polarization directions. The first linearly polarized light is incident on the first polarizer, the first polarizer filters the first linearly polarized light, the filtered first linearly polarized light is incident on the nonlinear activation device, the nonlinear activation device modulates the filtered first linearly polarized light, and the modulated first linearly polarized light is incident on the analyzer to achieve optical nonlinear activation. The second linearly polarized light is incident on the second polarizer, the second polarizer filters the second linearly polarized light, the filtered second linearly polarized light is incident on the nonlinear activation device, the nonlinear activation device modulates the filtered second linearly polarized light, and the modulated second linearly polarized light is incident on the analyzer to achieve optical nonlinear activation.

[0079] In an embodiment, a linearly polarized light source provides linearly polarized light as an input optical signal, and the polarization azimuth angle of the linearly polarized light is encoded as positive at 45° and as negative at -45°.

[0080] In an embodiment, the angle between the polarization direction of the first polarizer 2001 and the electric field direction of the first nonlinear activation device 1001 is 45°, and the angle between the polarization direction of the second polarizer 2002 and the electric field direction of the second nonlinear activation device 1002 is 45°. If the polarization azimuth angle of linearly polarized light is 45°, the first linearly polarized light is incident on the first polarizer 2001, and then sequentially passes through the first nonlinear activation device 1001 and the first polarizer 3001; if the polarization azimuth angle of linearly polarized light is -45°, the second linearly polarized light is incident on the second polarizer 2002, and then sequentially passes through the second nonlinear activation device 1002 and the second polarizer 3002.

[0081] The embodiment has the benefit that the nonlinear activation layer can automatically activate linearly polarized light with a polarization azimuth angle of 45° and -45°, improving the degree of automation of nonlinear activation and facilitating the integration of optical neural networks.

[0082] In an embodiment, referring to Figure 9 , the nonlinear activation layer further comprises: a quarter-wave plate 500, the quarter-wave plate 500 being arranged on one side of the light splitting prism 400. Wherein, the circularly polarized light Y is incident on the quarter-wave plate 500, and the quarter-wave plate 500 is used to convert the circularly polarized light Y into linearly polarized light. After the linearly polarized light is obtained by conversion, the linearly polarized light is incident on the light splitting prism. The light path after the linearly polarized light is incident on the light splitting prism has been described in the above (such as the embodiment shown in Figure 7 ).

[0083] It should be noted that after the incident circularly polarized light is converted into linearly polarized light by the quarter-wave plate 500 (also referred to as 1 / 4 wave plate), the linearly polarized light selectively enters from one branch (the light path of the first polarizer or the light path of the second polarizer) from the side of the Kerr cell of the two-dimensional material liquid crystal with a vertical electric field direction after passing through the light splitting prism 400, the first polarizer 2001 and the second polarizer 2002, the orientation of the two-dimensional material liquid crystal in the functional layer of the Kerr cell is deflected under the action of the electric field, the filtered linearly polarized light passes through the two-dimensional material liquid crystal to generate a phase delay, and is transmitted from the Kerr cell, achieving different activation effects on linearly polarized light with positive or negative azimuth angles.

[0084] In an embodiment, the circularly polarized light source provides circularly polarized light as an input light signal, and the left-handed circularly polarized light phase difference is -π / 2 encoded as negative, and the right-handed circularly polarized light phase difference is π / 2 encoded as positive.

[0085] In an embodiment, the angle between the polarization direction of the first polarizer 2001 and the electric field direction of the first nonlinear activation device 1001 is 45°, and the angle between the polarization direction of the second polarizer 2002 and the electric field direction of the second nonlinear activation device 1002 is 45°. If the phase difference of the circularly polarized light is π / 2, the first linearly polarized light is incident on the first polarizer 2001, and then sequentially passes through the first nonlinear activation device 1001 and the first polarizer 3001; if the phase difference of the linearly polarized light is -π / 2, the second linearly polarized light is incident on the second polarizer 2002, and then sequentially passes through the second nonlinear activation device 1002 and the second polarizer 3002.

[0086] The embodiment has the benefits that the nonlinear activation layer can automatically perform nonlinear activation on the circularly polarized light with a phase difference of π / 2 and -π / 2, respectively, improves the degree of automation of nonlinear activation, and facilitates the integration of the optical neural network.

[0087] It can be understood that the application uses the optical path design to complete the reconfigurable binary nonlinear activation layer in free space by encoding the polarization characteristics as positive and negative signs and combining the incident light and the outgoing light relationship of the Kerr cell of the two-dimensional material liquid crystal, and realizes the reconfigurable binary optical activation on the positive and negative input signals (linearly polarized light or circularly polarized light) in free space. The size of the binary nonlinear activation device can be expanded from microns to inch level, the response speed is in the order of picoseconds, and the response range covers the visible light range. The reconfigurable binary nonlinear activation layer solves the problems of delay and fixed response of optical activation, and cannot perform different processing on positive and negative inputs, expands the function of optical activation, and provides the possibility for the implementation of the all-optical neural network facing complex tasks.

[0088] In an embodiment, a nonlinear activation method for an optical neural network is applied to the nonlinear activation layer as described above, and the method comprises:

[0089] The linearly polarized light is incident on the polarizer, and the linearly polarized light is filtered by the polarizer. The filtered linearly polarized light is incident on the nonlinear activation device.

[0090] An electric field is applied to the nonlinear activation device, so that the filtered linearly polarized light is modulated by the two-dimensional material liquid crystal deflected under the action of the electric field. The modulated linearly polarized light is incident on the polarizer to realize optical nonlinear activation.

[0091] Specifically, the nonlinear activation layer based on the incident linearly polarized light is set to encode positive and negative based on the polarization azimuth angle of the incident linearly polarized light, and the negative value response law is independently defined to realize the reconfigurable binary optical activation in the full range.

[0092] In one embodiment, before controlling the linear polarized light to be incident on the polarizer, the method further includes: controlling the circular polarized light to be incident on a quarter wave plate, and converting the circular polarized light into linear polarized light by the quarter wave plate. After the linear polarized light is converted, the linear polarized light is incident on a beam splitter prism. The optical path after the linear polarized light is incident on the beam splitter prism has been described above (e.g. Figure 3 The embodiment shown in FIG3 is described in detail and will not be repeated here.

[0093] In one example, a process for implementing nonlinear activation for incident linearly polarized light may include:

[0094] The nonlinear activation layer based on incident linearly polarized light is shown in the following diagram: Figure 8 As shown in FIG. The nonlinear activation layer includes a beam splitter prism 400, a first polarizer 2001, a first nonlinear activation device 1001, a first analyzer 3001, a second polarizer 2002, a second nonlinear activation device 1002, and a second analyzer 3002. Linearly polarized light X passes through the beam splitter prism 400 and is provided as incident light to the first nonlinear activation device 1001 and the second nonlinear activation device 1002 in perpendicular optical paths. The first polarizer 2001 is disposed between the beam splitter prism 400 and the first nonlinear activation device 1001, and the second polarizer 2002 is disposed between the beam splitter prism 400 and the second nonlinear activation device 1002. The polarization direction of the first polarizer 2001 in the two perpendicular optical paths is perpendicular to the polarization direction of the second polarizer 2002 and forms a 45° angle with the electric field direction. The first analyzer 3001 is disposed at the end of the optical path, on the opposite side from the first nonlinear activation device 1001 and the first polarizer 2001. The second analyzer 3002 is disposed at the end of the optical path and on the opposite side of the second nonlinear active device 1002 and the second polarizer 2002. The polarization direction of the first analyzer 3001 is orthogonal to the polarization direction of the first polarizer 2001. The polarization direction of the second analyzer 3002 is orthogonal to the polarization direction of the second polarizer 2002.

[0095] In this example, linearly polarized light X is used as an input optical signal and is encoded as positive and negative at polarization angles of 45° and -45°, respectively. It is split into two perpendicular paths with equal light intensity and the same phase by a beam splitter prism 400. The first polarizer 2001 and the second polarizer 2002 filter the positive and negative inputs, respectively. The filtered (also called beam splitting) incident light is selectively incident from one side of the binary nonlinear activation device (the first nonlinear activation device 1001 or the second nonlinear activation device 1002) through one branch in a direction perpendicular to the electric field. The two-dimensional liquid crystal material in the functional layer of the Kerr cell is 10 4Hz, 1V / mm electric field deflection occurs, the input optical signal phase delay occurs from the two-dimensional material liquid crystal, and the transmitted light is emitted from the Kerr box. In this setting, the nonlinear activation characteristic curve of the binary nonlinear activation device is as shown in Figure 10 Figure 10 In the figure, the abscissa is the electric field intensity, and the ordinate is the optical activation characteristic. It mainly shows the comparison of the binary activation of the two-dimensional mica aqueous dispersion liquid and the traditional linear rectifier function activation effect.

[0096] In an example, a nonlinear activation implementation process for incident circularly polarized light can include:

[0097] A nonlinear activation layer based on incident circularly polarized light, the structure diagram is as shown in Figure 9 The nonlinear activation layer includes: a quarter-wave plate 500, a beam splitter prism 400, a first polarizer 2001, a first nonlinear activation device 1001, a first analyzer 3001, a second polarizer 2002, a second nonlinear activation device 1002, and a second analyzer 3002. Circularly polarized light Y provides incident light to the first nonlinear activation device 1001 and the second nonlinear device 1002 in the vertical light path in turn after passing through the quarter-wave plate 500 and the beam splitter prism 400. The first polarizer 2001 is arranged between the beam splitter prism 400 and the first nonlinear activation device 1001, and the second polarizer 2002 is arranged between the beam splitter prism 400 and the second nonlinear activation device 1002. The polarization direction of the first polarizer 2001 in the two vertical light paths is perpendicular to the polarization direction of the second polarizer 2002, and forms a 45° angle with the electric field direction. The first analyzer 3001 is arranged at the end of the light path and opposite to the first nonlinear activation device 1001 and the first polarizer 2001. The second analyzer 3002 is arranged at the end of the light path and opposite to the second nonlinear activation device 1002 and the second polarizer 2002. The polarization direction of the first analyzer 3001 is orthogonal to the polarization direction of the first polarizer 2001. The polarization direction of the second analyzer 3002 is orthogonal to the polarization direction of the second polarizer 2002.

[0098] In this example, when the circularly polarized light Y as the input optical signal phase difference is π / 2 and -π / 2, it is encoded as positive and negative respectively, converted into linearly polarized light by the quarter-wave plate 500, and divided into two vertical paths with equal light intensity and the same phase by the beam splitter prism 400. The first polarizer 2001 and the second polarizer 2002 filter the positive and negative inputs respectively, and the incident light after beam splitting is selectively incident from one branch to the side of the binary nonlinear activation device (the first nonlinear activation device 1001 or the second nonlinear activation device 1002) with a vertical electric field direction. The two-dimensional material liquid crystal in the functional layer of the Kerr box is in a 10 4 ​Hz, 1V / mm electric field, the input optical signal (phase difference is π / 2 corresponding to polarization azimuth angle 45°, phase difference is-π / 2 corresponding to polarization azimuth angle-45°) generates phase delay through the two-dimensional material liquid crystal, and is transmitted out of the Kerr box, so that the positive and negative input of the polarization azimuth angle realizes different activation effects.

[0099] The embodiments described in the embodiments of the present application are used to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0100] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and can include more or fewer steps than those shown in the figures, or combine certain steps or different steps.

[0101] Those skilled in the art can understand that all or some steps in the above disclosed method, the functions of the modules / units in the system and the device can be implemented as software, firmware, hardware and their appropriate combinations.

[0102] The terms "first", "second", "third", "fourth" and the like (if any) in the specification of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be exchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0103] It should be understood that in the present application, "at least one" refers to one or more, and "multiple" refers to two or more. "And / or" is used to describe the association relationship of the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases of only A, only B, and A and B existing at the same time, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after it. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can mean a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0104] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed to multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment of the present application according to actual needs.

[0105] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0106] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including multiple instructions for causing an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program storage media.

[0107] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, but this does not limit the scope of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A nonlinear activation layer for an optical neural network, characterized in that The nonlinear activation layer comprises: A binary nonlinear activation device, comprising a two-dimensional liquid crystal that deflects under the action of an electric field, the binary nonlinear activation device comprising a first nonlinear activation device and a second nonlinear activation device with the electric field directions perpendicular to each other; a beam splitter prism, the beam splitter prism being disposed on one side of the first nonlinear activation device and on one side of the second nonlinear activation device; A polarizer, the polarizer being arranged on a first side of the binary nonlinear activation device, the polarizer comprising a first polarizer and a second polarizer having perpendicular polarization directions; wherein, when linearly polarized light is incident on the beam splitter prism, the beam splitter prism splits the linearly polarized light into a first linear polarized light and a second linear polarized light, the polarization direction of the first linear polarized light and the polarization direction of the second linear polarized light being perpendicular; the first linear polarized light is incident on the first polarizer, and the second linear polarized light is incident on the second polarizer; an analyzer, the analyzer being arranged on a second side of the binary nonlinear activation device, the second side being a side opposite to the first side, the analyzer comprising a first analyzer and a second analyzer having polarization directions perpendicular thereto, the polarization direction of the first polarizer being the same as the polarization direction of the second polarizer, and the polarization direction of the second polarizer being the same as the polarization direction of the first analyzer; The linearly polarized light is incident on the polarizer, the polarizer filters the linearly polarized light, the filtered linearly polarized light is incident on the binary nonlinear activation device, the filtered linearly polarized light is modulated by the deflected two-dimensional material liquid crystal, and the modulated linearly polarized light is incident on the analyzer to achieve optical nonlinear activation; The first nonlinear activation device is used for positive modulation, the second nonlinear activation device is used for negative modulation, the angle between the polarization direction of the first polarizer and the electric field direction of the first nonlinear activation device is 45°, and the angle between the polarization direction of the second polarizer and the electric field direction of the second nonlinear activation device is 45°; If the polarization azimuth angle of the linearly polarized light is 45°, then after the first linearly polarized light is incident on the first polarizer, it passes through the first nonlinear activation device and the first analyzer in sequence; If the polarization azimuth angle of the linearly polarized light is -45°, the second linearly polarized light enters the second polarizer and then passes through the second nonlinear activation device and the second analyzer in sequence.

2. The nonlinear activation layer according to claim 1, characterized in that The nonlinear activation layer further includes a quarter wave plate, which is arranged on one side of the beam splitter prism. Wherein, the circularly polarized light is incident on the quarter-wave plate, and the quarter-wave plate is used to convert the circularly polarized light into linearly polarized light.

3. The nonlinear activation layer according to claim 2, characterized in that The first nonlinear activation device is used for positive modulation, the second nonlinear activation device is used for negative modulation, the angle between the polarization direction of the first polarizer and the electric field direction of the first nonlinear activation device is 45°, and the angle between the polarization direction of the second polarizer and the electric field direction of the second nonlinear activation device is 45°; If the phase difference of the circularly polarized light is π / 2, then after the first linearly polarized light is incident on the first polarizer, it passes through the first nonlinear activation device and the first analyzer in sequence; If the phase difference of the linearly polarized light is -π / 2, the second linearly polarized light enters the second polarizer and then passes through the second nonlinear activation device and the second analyzer in sequence.

4. The nonlinear activation layer according to any one of claims 2 to 3, characterized in that The first nonlinear activation device and the second nonlinear activation device include: Kerr cells made of two-dimensional liquid crystal.

5. The nonlinear activation layer according to claim 4, characterized in that The Kerr cell of the two-dimensional liquid crystal material comprises: A Kerr cell container having a built-in functional layer, wherein the functional layer contains a two-dimensional material liquid crystal; two electrodes, the two electrodes being respectively arranged on opposite sides of the Kerr cell container; The two electrodes are connected to an external power source to apply an electric field to the functional layer, and the orientation of the two-dimensional material liquid crystal is deflected under the action of the electric field.

6. The nonlinear activation layer according to claim 5, characterized in that The two-dimensional liquid crystal material is a two-dimensional vermiculite aqueous dispersion, which can achieve π / 2 phase modulation under the action of an electric field.

7. A nonlinear activation method for optical neural networks, characterized in that Applied to the nonlinear activation layer according to any one of claims 1 to 6, the method comprises: Controlling linearly polarized light to be incident on the polarizer, filtering the linearly polarized light by the polarizer, and allowing the filtered linearly polarized light to be incident on the nonlinear activation device; An electric field is applied to the nonlinear activation device to modulate the filtered linearly polarized light through the two-dimensional material liquid crystal that is deflected under the action of the electric field. The modulated linearly polarized light is incident on the analyzer to achieve optical nonlinear activation.

8. The nonlinear activation method according to claim 7, characterized in that: Before controlling the linearly polarized light to be incident on the polarizer, the method further includes: The circularly polarized light is controlled to be incident on a quarter wave plate, and the quarter wave plate converts the circularly polarized light into linearly polarized light.

Images