Optical computation device

Abstract

This optical computation device (100) comprises: an optical neural network unit (110) that uses an optical neural network that performs computation processing by causing light beams to interfere with each other, and outputs a result of the computation processing as a light intensity distribution; and a determination unit (120) that determines the result of the computation processing from the light intensity distribution. The optical neural network includes at least one intermediate layer comprising at least one optical modulation unit (111) paired with at least one light diffusion / dispersion unit (112), wherein the optical modulation unit (111) modulates light and the light diffusion / dispersion unit (112) diffuses, disperses, or diffracts the light output from the optical modulation unit (111).

Description

optical calculation device

[0001] This disclosure relates to an optical computing device.

[0002] There is an optical computing technology that performs calculations optically using the diffraction and interference of light (see, for example, Patent Document 1). Using this optical computing technology, it is possible to optically process two-dimensional input light and classify or recognize the characteristics of the input light, or recognize or detect unique changes in the time series.

[0003] Optical computing technology, compared to conventional electronic calculations using semiconductor integrated circuit chips, utilizes the physical properties of light, enabling reduced power consumption and high-speed computation processing due to the propagation of light.

[0004] International Publication No. 2022 / 130690

[0005] Optical computing technology has been proposed for application in various systems. For example, an optical neural network can be constructed using optical diffraction elements that enable spatially parallel optical computation on two-dimensional input light.

[0006] In optical neural networks that use optical diffraction elements to perform calculations optically by interfering with each other with light transmitted through an intensity-phase modulation section in a two-dimensional space, it is necessary to take a relatively large propagation distance between the optical diffraction elements, which presents the challenge of increasing the overall size of the device.

[0007] Furthermore, the design constraint of matching the lattice pitch of the optical diffraction element to the required diffraction angle necessitated miniaturization beyond the resolution required for identifying and classifying the input light, making it difficult to apply general-purpose phase modulation elements.

[0008] Furthermore, when the phase modulation elements in the spatial intensity phase modulation section were made low-resolution, there was a problem in that the coupling of the optical neural network between the optical diffraction elements weakened.

[0009] Therefore, one or more aspects of this disclosure aim to enable the realization of a strongly coupled optical neural network by shortening the optical propagation distance using general-purpose elements.

[0010] An optical computing device according to one aspect of the present disclosure comprises an optical neural network unit that performs computational processing by interfering light with each other and outputs the result of the computational processing as a distribution of light intensity, and a determination unit that determines the result of the computational processing from the distribution, wherein the optical neural network includes at least one intermediate layer comprising a pair of at least one optical modulation unit that modulates light and at least one optical diffusion dispersion unit that diffuses, disperses or diffracts the light output from the optical modulation unit.

[0011] According to one or more aspects of this disclosure, a strongly coupled optical neural network can be realized by shortening the optical propagation distance using general-purpose elements.

[0012] This is a schematic block diagram showing the configuration of the optical computing device according to Embodiment 1. This is a schematic diagram for explaining the resolution of the optical modulation section. This is a schematic diagram showing an example in which multiple intermediate layers are constructed using a low-resolution diffraction element. This is a schematic diagram showing an example in which multiple intermediate layers are constructed with multiple pairs of optical modulation sections and optical diffusion / dispersion sections, as in Embodiment 1. (A) and (B) are schematic diagrams showing an example of the installation of the optical modulation section when there are multiple intermediate layers. (A) to (C) are schematic diagrams showing the division elements of the optical modulation section. (A) and (B) are block diagrams showing an example of hardware configuration. This is a schematic block diagram showing the configuration of the optical computing device according to Embodiment 2. This is a schematic block diagram showing the configuration of a modified example of the optical computing device according to Embodiment 1.

[0013] Embodiment 1. Figure 1 is a block diagram schematically showing the configuration of the optical computing device 100 according to Embodiment 1. The optical computing device 100 comprises an optical neural network unit 110, a determination unit 120, a learning unit 121, a modulation control unit 122, and a storage unit 123.

[0014] The optical neural network unit 110 uses an optical neural network that performs computational processing by interfering light with each other, and outputs the result of the computational processing as a distribution of light intensity. Here, the optical neural network includes at least one intermediate layer 114 comprising a pair of optical modulation units 111 that modulate light and optical diffusion / dispersion units 112 that diffuse, disperse, or diffract the light output from the optical modulation unit 111.

[0015] In this embodiment, the intermediate layer 114 consists of a pair of light modulation section 111 and light diffusion / dispersion section 112. However, the intermediate layer 114 may include other configurations as long as it has a pair of light modulation section 111 and light diffusion / dispersion section 112.

[0016] Furthermore, the optical modulation unit 111 and the optical diffusion / dispersion unit 112 do not necessarily have to be individual components; the functions performed by the optical modulation unit 111 and the optical diffusion / dispersion unit 112 may be divided into multiple components. For example, the effects that the optical modulation unit 111 and the optical diffusion / dispersion unit 112 should exert on light may be performed by a combination of multiple components. Specifically, the optical diffusion / dispersion unit 112 is composed of two diffraction elements that diffract light in a simple one-dimensional direction, and these two diffraction elements may be arranged so that the diffraction direction of one diffraction element (also called the second diffraction element) is rotated by 10 to 90 degrees around the optical axis with respect to the one-dimensional direction of the diffraction direction of the other diffraction element (also called the first diffraction element). As a result, the optical diffusion and dispersion section 112 is composed of a first diffraction element that diffracts in one dimension, and a second diffraction element that is positioned at an angle to the optical axis from the diffraction direction of the first diffraction element, enabling the light output from the optical modulation section 111 to be diffracted in two dimensions. In such an optical diffusion and dispersion section 112, the light is diffracted in two dimensions by two diffraction elements. By diffracting the light in two dimensions, the optical diffusion and dispersion section 112 increases the diffraction direction of the light from one dimension to two dimensions, thereby strengthening the coupling of the optical neural network described later. Alternatively, for example, the optical diffusion and dispersion section 112 can be configured so that the light is diffused or diffracted in stages by multiple diffraction elements with low resolution. Such a configuration has the advantage that the optical diffusion and dispersion section 112 can be composed of components with even lower resolution.

[0017] The optical neural network section 110 comprises an optical modulation section 111, an optical diffusion / dispersion section 112, and a light receiving section 113. The optical modulation section 111, the optical diffusion / dispersion section 112, and the light receiving section 113 function as an optical neural network. In Figure 1, one intermediate layer 114 is shown, each comprising a pair of optical modulation sections 111 and optical diffusion / dispersion sections 112. However, the more pairs there are, the more complex calculations can be performed. For this reason, the optical neural network section 110 may comprise multiple intermediate layers 114, each comprising a pair of optical modulation sections 111 and optical diffusion / dispersion sections 112.

[0018] The optical modulation unit 111 modulates light. For example, the optical modulation unit 111 changes at least one of the intensity, phase, and direction of vibration of the input light and outputs it. Here, the optical modulation unit 111 changes at least one of the intensity, phase, and direction of vibration of the input light and outputs it in accordance with a control signal from the modulation control unit 122, which will be described later.

[0019] Specifically, the optical modulation unit 111 can be constructed using a material whose refractive index changes in response to external physical forces such as stress, electromagnetic fields, or temperature changes, thereby changing the amount of optical modulation. In this case, the modulation control unit 122 can change these external physical forces using a control signal.

[0020] For example, liquid crystal can be used as a material whose refractive index can be varied. Amorphous silicon (a-Si) can also be used as such a material, as can GeSbTe-based materials (ternary compounds), AgInSbTe (quaternary compounds), and VO ２ Phase change materials such as semiconductors (vanadium oxide) can also be used, such as LiNbO ３ Ferroelectric crystal materials such as those mentioned above can also be used, and Pockels effect materials can also be used. Furthermore, polymer light modulation materials can be used as components with a variable refractive index.

[0021] Furthermore, the optical modulation unit 111 can be constructed using a material whose birefringence changes when an external magnetic force or external electric field is applied. In this case, the modulation control unit 122 can change its external magnetic force or external electric field using a control signal.

[0022] For example, a waveplate has two orthogonal phase-advancing axes (fast axis) and a phase-slow axis (slow axis), and the polarization of incident light changes because a phase difference of light occurs in each polarization component due to the difference in refractive index (birefringence) between the phase axes. Liquid crystals, on the other hand, are materials whose birefringence changes due to an external electric field.

[0023] Here, the resolution of the optical modulation unit 111 will be explained using Figure 2. In Figure 2, the resolution is explained using two adjacent optical modulation units 111A and 111B. Note that the optical modulation unit 111A is located upstream of the optical modulation unit 111B with respect to the direction of incident light (the direction of propagation of incident light). The optical modulation unit 111A is also called the first optical modulation unit, and the optical modulation unit 111B is also called the second optical modulation unit.

[0024] As shown in Figure 2, let λ be the wavelength of light incident on the optical modulation unit 111A, Λ be the size of the microcell (also called region) constituting the optical modulation unit 111A, and θ be the diffraction angle of the first diffracted light diffracted by the optical modulation unit 111A. In such a case, the diffraction angle θ can be expressed by equation (1) below. Note that, regarding the wavelength λ of light, if there is a range of wavelengths to be used depending on the processing task, the approximate center wavelength of that range should be used. (1)

[0025] Here, the resolution of the optical modulation unit 111A decreases as the size Λ of the microcell increases relative to the wavelength λ, in other words, as the diffraction angle θ of the primary diffracted light diffracted by the optical modulation unit 111A decreases. Conversely, the resolution of the optical modulation unit 111A increases as the size Λ of the microcell increases relative to the wavelength λ, in other words, as the diffraction angle θ of the primary diffracted light diffracted by the optical modulation unit 111A increases.

[0026] To strengthen the coupling of the optical neural network, it is conceivable to increase the resolution of the optical modulation unit 111 used. On the other hand, an optical modulation unit 111 with high resolution requires many processing steps and high-precision processing.

[0027] Therefore, as shown in Figure 2, if the distance between the light modulation unit 111A and the light modulation unit 111B is d, and the length of the side of the light modulation unit 111B is L, then the maximum diffraction angle θ due to the light modulation unit 111A is ｍａｘ This can be expressed by equation (2) below. Here, the length L of the side of the light modulation section 111B is the length of the side that intersects the direction of propagation of the light incident on the light modulation section 111. The length L of the side of the light modulation section 111B is the length of the longest side among the four sides of the surface that obstructs the propagation of the light. (2)

[0028] As described above, the coupling of the optical neural network becomes stronger as the diffraction angle θ of the primary diffracted light diffracted by the optical modulation unit 111A increases. θ = θ ｍａｘ When the size Λ of the microcell is reduced with respect to the wavelength λ to satisfy the condition, the resolution of the optical modulation section 111A also increases, and the coupling strength can be maximized. However, for optical neural network coupling, for example, θ > 1 ÷ 2 × θ ｍａｘ This is desirable. In other words, θ > 1 ÷ 2 × θ ｍａｘ A light modulation unit 111 that results in such a value can be said to have high resolution. On the other hand, for example, θ ≤ 1 ÷ 2 × θ ｍａｘ A light modulation section 111A that behaves in this way can be said to have low resolution.

[0029] Furthermore, from the size Λ of one microcell in the optical modulation unit 111A, the diffraction angle θ≦θ of the primary diffracted light diffracted by the optical modulation unit 111A is obtained by equation (3) below. ｍｉｎ If the diffraction angle θ of the optical modulation unit 111A is set in such a way that the diffraction angle at the optical modulation unit 111A is almost eliminated. (3) Even when such diffraction angles hardly occur, the optical modulation unit 111A can be said to have low resolution.

[0030] However, as in the present embodiment, even when using a light modulation unit 111 with low resolution, the coupling of the optical neural network can be strengthened by adopting a configuration in which the light modulation unit 111 and the light diffusion and dispersion unit 112 are paired. And by using such a light modulation unit 111 with low resolution, the number of processing steps and the like can be suppressed, and the manufacturing cost of the light modulation unit 111 can be reduced.

[0031] Note that the resolution of the light modulation unit 111 may be set lower than a predetermined threshold value. The above-mentioned 1÷2×θ ｍａｘ and θ ｍｉｎ are examples of the predetermined threshold value. The predetermined threshold value can be determined according to the scale or performance of the optical arithmetic device 100.

[0032] Hereinafter, the advantages of setting the resolution of the light modulation unit 111 to be low will be described. When the resolution is high, in other words, when the size of each of the plurality of microcells that modulate light is small, the difficulty of manufacturing the device of the light modulation unit 111 increases, the manufacturing yield of the device deteriorates, and the cost increases. For example, with respect to the wavelength used for processing in microfabrication technologies such as lithography, the closer the size of the microcells of the light modulation unit 111 is to the wavelength used for processing or smaller than that, the higher the processing difficulty becomes. Therefore, setting the size of the microcells of the light modulation unit 111 of the optical arithmetic device 100 of the present embodiment to be large and reducing the resolution has the advantage of reducing the processing difficulty of the device of the light modulation unit 111.

[0033] In addition, this makes it easier to apply a device called a general-purpose product rather than a custom product to the light modulation unit 111. For example, in the case of a liquid crystal device, there is also an advantage that a liquid crystal device having general-purpose specifications equivalent to or lower than those used in already mature liquid crystal displays can be used as the light modulation unit 111.

[0034] The light diffusion / dispersion unit 112 diffuses, disperses, or diffracts the input light and outputs it. Specifically, the light diffusion / dispersion unit 112 is a phase modulation element that modulates the phase in a fixed manner. For example, the light diffusion / dispersion unit 112 is a diffraction element that diffracts light in a fixed pattern such as a line type or a geometric pattern such as vertical or horizontal, a diffuser containing scattering particles which are particles that scatter light, or a diffuser plate that diffuses light in a speckled manner.

[0035] Here, if the optical modulation section 111 is composed of multiple microcells whose refractive index can be independently changed, and the optical diffusion / dispersion section 112 is a diffraction element with a geometric pattern such as a vertical or horizontal orientation, it is desirable that the microcells of the diffraction element be smaller in size than the microcells of the optical modulation section 111. In other words, it is desirable that the resolution of the optical modulation section 111 be lower than the resolution of the optical diffusion / dispersion section 112.

[0036] The light diffusion and dispersion section 112 is composed of two diffraction elements that diffract light in a simple one-dimensional direction. These two diffraction elements may be arranged such that the diffraction direction of one diffraction element (also called the second diffraction element) is rotated by 10 to 90 degrees around the optical axis relative to the diffraction direction of the other diffraction element (also called the first diffraction element). In such a configuration, the light diffusion and dispersion section 112 diffracts light in two dimensions using the two diffraction elements. By diffracting light in two dimensions, the light diffusion and dispersion section 112 increases the diffraction direction of light from one dimension to two dimensions, thereby strengthening the coupling of the optical neural network described later. In addition, the light diffusion and dispersion section 112 can be configured such that light is diffused or diffracted in stages using multiple diffraction elements with low resolution.

[0037] As described above, in Embodiment 1, one intermediate layer 114 is formed by a pair of a light modulation unit 111 with low resolution and a light diffusion / dispersion unit 112. For example, as shown in FIG. 3, when a plurality of intermediate layers are formed using a diffraction element 101 with low resolution, diffusion or dispersion becomes insufficient, and the bonding between the layers becomes weak. On the other hand, by forming one intermediate layer with a pair of a light modulation unit 111 with low resolution and a light diffusion / dispersion unit 112, the light diffusion / dispersion unit 112 can sufficiently diffuse or disperse light. Therefore, even when using a light modulation unit 111 with low resolution, such problems can be solved.

[0038] Further, as shown in FIG. 3, when forming a plurality of intermediate layers using only diffraction elements 101#1, 101#2, 101#3 with low resolution, in order to obtain sufficient bonding, the distance between the diffraction elements 101#1, 101#2, 101#3 must be increased. On the other hand, as shown in Embodiment 1, one intermediate layer 114 is formed by a pair of a light modulation unit 111 with low resolution and a light diffusion / dispersion unit 112, and a plurality of intermediate layers 114 are arranged along the direction of the optical axis of the intermediate layer 114. In other words, as shown in FIG. 4, when the optical neural network unit 110 is configured to include a plurality of intermediate layers 114#1, 114#2, 114#3, even if the distance between the intermediate layers 114#1, 114#2, 114#3 is shortened, the light output is sufficiently diffused or dispersed by the light diffusion / dispersion units 112#1, 112#2, 112#3, and the bonding between the layers becomes strong. In FIG. 4, the intermediate layer 114#1 includes a light modulation unit 111#1 and a light diffusion / dispersion unit 112#1, the intermediate layer 114#2 includes a light modulation unit 111#2 and a light diffusion / dispersion unit 112#2, and the intermediate layer 114#3 includes a light modulation unit 111#3 and a light diffusion / dispersion unit 112#3, but the present embodiment is not limited to such an example. Note that in FIG. 3, only the ±1st order light including the 0th order light is shown as diffracted light, and higher order diffracted lights are omitted. For the sake of explanation, diffracted lights are shown with limited orders in other drawings as well.

[0039] Incidentally, as shown in FIG. 4, when a plurality of intermediate layers 114#1, 114#2, 114#3 are constituted by a plurality of pairs of light modulation units 111#1, 111#2, 111#3 with low resolution and light diffusion and dispersion units 112#1, 112#2, 112#3, for example, as shown in FIG. 5(A), the space between the light modulation unit 111C and the light modulation unit 111D adjacent to the light modulation unit 111C may be offset in a direction perpendicular to the optical axis. The offset is preferably less than the length in the direction perpendicular to the optical axis of one microcell of the light modulation unit 111C and the light modulation unit 111D. Incidentally, the intermediate layer including the light modulation unit 111C is also referred to as the first intermediate layer, and the intermediate layer including the light modulation unit 111D is also referred to as the second intermediate layer.

[0040] Thus, the light modulation unit 111C and the light modulation unit 111D are offset so that, when viewed from the direction of the optical axis of the intermediate layer 114, the regions of their respective microcells are partially displaced, or rather, the entire regions of their respective microcells do not overlap. As a result, the light that passes through the region of one microcell of the light modulation unit 111C and travels straight is dispersed and combined in the regions of four adjacent microcells of the light modulation unit 111D. Since the four microcell regions each give a different modulation amount to the light, the coupling of the optical neural network can be substantially increased.

[0041] Further, as shown in FIG. 5(B), the space between the light modulation unit 111C and the light modulation unit 111D adjacent to the light modulation unit 111C may be offset in a direction of rotation with respect to the optical axis. The offset is preferably less than the length in the direction perpendicular to the optical axis of one microcell of the light modulation unit 111C and the light modulation unit 111D. In this case, the divided elements of the light modulation unit 111C are as shown in FIG. 6(A), the divided elements of the light modulation unit 111D are as shown in FIG. 6(B), and the overlapping of the divided elements of the light modulation unit 111C and the light modulation unit 111D is as shown in FIG. 6(C).

[0042] Thus, the optical modulation section 111C and the optical modulation section 111D are offset such that, when viewed from the direction of the optical axis of the intermediate layer 114, the regions of their respective microcells are partially misaligned, or in other words, their entire microcell regions do not overlap. As a result, light that passes through the region of one microcell in the optical modulation section 111C and travels in a straight line is dispersed and coupled into multiple adjacent microcell regions of the optical modulation section 111D. Since each of these multiple microcell regions imparts a different modulation amount to the light, the coupling of the optical neural network can be effectively increased.

[0043] When multiple intermediate layers 114 are formed by multiple pairs of optical modulation units 111 and optical diffusion / dispersion units 112, the multiple optical modulation units 111 included in the multiple intermediate layers 114 may have the same modulation pattern or different patterns. Similarly, the multiple optical diffusion / dispersion units 112 included in the multiple intermediate layers 114 may have the same diffusion, dispersion, or diffraction performance or different patterns.

[0044] The light-receiving unit 113 detects the distribution of light intensity output from the light diffusion / dispersion unit 112. For example, the light-receiving unit 113 can be configured as a two-dimensional image sensor. However, it is not limited to a two-dimensional image sensor; it may also be a sensor array with multiple light-receiving areas, or a unit of light-receiving elements arranged in two dimensions. As an example, in the case of an optical computing device that processes an identification task, the light-receiving unit 113 may consist of a light-receiving element with light-receiving areas corresponding to multiple classification classes according to the identification task, or a separate light-receiving element for each of the multiple classes.

[0045] The determination unit 120 determines the result of the calculation processing of the optical neural network unit 110 based on the distribution of light intensity detected by the light receiving unit 113. Here, the determination unit 120 identifies the target indicated by the input light from the intensity distribution detected by the light receiving unit 113.

[0046] The learning unit 121 changes the control signal that the modulation control unit 122 sends to the optical modulation unit 111 so that the identification result in the determination unit 120 matches the target of the input light indicated in the target data. For example, the target data indicates the light intensity value for each pixel and the target to be identified in that case, and the learning unit 121 changes the control signal that the modulation control unit 122 sends to the optical modulation unit 111 so that the difference between the light intensity value for each pixel detected by the light receiving unit 113 and the light intensity value for each pixel indicated in the target data becomes smaller.

[0047] Then, when the learning completion conditions are met, for example, when the difference between the light intensity value indicated by the target data and the light intensity value detected by the light receiving unit 113 becomes smaller than a predetermined threshold, or when the number of changes to the control signal in the modulation control unit 122 exceeds a predetermined threshold, the learning unit 121 notifies the modulation control unit 122 that learning has ended. Upon receiving notification of the end of learning, the modulation control unit 122 stores parameters in the storage unit 123 to identify the control signal at the time learning has ended.

[0048] The modulation control unit 122 changes the control signal supplied to the optical modulation unit 111 in response to instructions from the learning unit 121. The storage unit 123 stores the parameters indicating the control signal at the time learning is completed.

[0049] Some or all of the determination unit 120, learning unit 121, and modulation control unit 122 described above can be configured, for example, with a memory 10 and a processor 11 such as a CPU (Central Processing Unit) that executes the program stored in the memory 10, as shown in Figure 7(A). Such a program may be provided via a network or by being recorded on a recording medium. That is, such a program may be provided, for example, as a computer program product.

[0050] Furthermore, part or all of the determination unit 120, learning unit 121, and modulation control unit 122 can also be composed of processing circuits 12 such as a single circuit, a composite circuit, a program-operated processor, a program-operated parallel processor, an ASIC (Application Specific Integrated Circuit), or an FPGA (Field Programmable Gate Array), as shown in Figure 7(B). As described above, the determination unit 120, learning unit 121, and modulation control unit 122 can be realized by a processing circuit network.

[0051] The memory unit 123 can be implemented using storage such as an HDD (Hard Disk Drive), SSD (Solid State Drive), ROM (Read Only Memory), or RAM (Random Access Memory).

[0052] In the embodiment 1 described above, in order to enable learning by the learning unit 121, the optical modulation unit 111 is a component that can change the amount of optical modulation in accordance with a control signal from the modulation control unit 122. However, when learning has already been completed and inference is to be performed by the optical neural network, a component that can output a fixed optical output of the learned optical modulation unit 111 may be used as the optical modulation unit 111. Examples of such components include a diffraction element (holographic optical element) that changes the amount of optical modulation depending on the thickness of the microcell, or a polarizer that transmits only light that vibrates in a specific direction.

[0053] Embodiment 2. Embodiment 1 was described on the premise that there is one task for identifying an object, but Embodiment 2 corresponds to multiple tasks for identifying an object. Figure 8 is a schematic block diagram showing the configuration of the optical computing device 200 according to Embodiment 2. The optical computing device 200 comprises an optical neural network unit 110, a determination unit 220, a learning unit 221, a modulation control unit 222, a storage unit 223, and a task control unit 224.

[0054] The optical neural network section 110 in Embodiment 2 is the same as the optical neural network section 110 in Embodiment 1, but in Figure 8 it is shown as a configuration that includes a plurality of intermediate layers 114#1, ..., 114#N, each comprising a plurality of pairs of optical modulation sections 111#1, ..., 111#N (where N is an integer of 2 or more) and optical diffusion dispersion sections 112#1, ..., 112#N.

[0055] Furthermore, when a plurality of intermediate layers 114#1, ..., 114#N are constructed, each comprising a plurality of pairs of optical modulation sections 111#1, ..., 111#N and optical diffusion / dispersion sections 112#1, ..., 112#N, the optical modulation sections 111#1, ..., 111#N included in the intermediate layers 114#1, ..., 114#N may have the same modulation pattern or different patterns. Similarly, the optical diffusion / dispersion sections 112#1, ..., 112#N may have the same diffusion, dispersion, or diffraction performance or different patterns.

[0056] In Embodiment 2, in order to learn multiple tasks, for example, as shown in Figure 8, the optical output device 240 can output input light corresponding to the task. As a result, the optical neural network unit 110 receives input light corresponding to each of the multiple tasks. Here, it is assumed that input light indicating a location or multiple objects to be identified for each task is output.

[0057] Specifically, the optical output device 240 comprises a light source 241, an optical modulation unit 242, and an input optical modulation control unit 243 that outputs input light according to the task to the optical modulation unit 242. The refractive index of the optical modulation unit 242 changes in response to external physical forces such as stress, electromagnetic fields, or temperature changes, and the input optical modulation control unit 243 changes the light output from the optical modulation unit 242 according to the input data by changing the control signal to the optical modulation unit 242.

[0058] The determination unit 220 determines the calculation result of the optical neural network unit 110 based on the intensity distribution detected by the light receiving unit 113. Here, the determination unit 220 identifies the target indicated by the input light from the intensity distribution detected by the light receiving unit 113. In Embodiment 2, the determination unit 220 corresponds to multiple tasks for identifying targets, and identifies one or more targets for each task according to the intensity distribution detected by the light receiving unit 113.

[0059] The task control unit 224 controls multiple tasks that identify targets in the optical computing unit 200. For example, in the learning phase, the task control unit 224 provides the learning unit 221 with the corresponding target data for each task to be learned. On the other hand, in the inference phase, the task control unit 224 selects one task from the multiple tasks and instructs the modulation control unit 222 to use that one task as the task to perform inference. As a result, the modulation control unit 222 reads the parameters of the instructed task from the storage unit 223 and inputs the control signal identified by the read parameters to the optical modulation unit 111.

[0060] The learning unit 221 changes the control signal that the modulation control unit 222 provides to the optical modulation unit 111 so that the identification result in the determination unit 220 matches the target of the input light indicated by the target data. In the second embodiment, the learning unit 221 performs learning by changing the control signal that the modulation control unit 222 provides to the optical modulation unit 111 using target data corresponding to each of the multiple tasks. The target data is provided by the task control unit 224.

[0061] Specifically, the learning unit 221 learns, for each of the multiple tasks, a control signal to be input to the optical modulation unit 111, such that the difference between the distribution of light intensity that represents the correct result of the calculation process and the output from the optical neural network unit 110 becomes small.

[0062] Then, for each task, when the learning completion conditions are met, for example, when the difference between the light intensity value indicated by the target data and the light intensity value detected by the light receiving unit 113 becomes smaller than a predetermined threshold, or when the number of changes to the control signal in the modulation control unit 222 exceeds a predetermined threshold, the learning unit 221 notifies the modulation control unit 222 that learning has ended. Upon receiving notification of the end of learning, the modulation control unit 222 stores parameters that identify the control signal at the time learning is completed in the storage unit 223 for each task.

[0063] The modulation control unit 222 changes the control signal supplied to the optical modulation unit 111 in response to instructions from the learning unit 221. In Embodiment 2, the modulation control unit 222 changes the control signal supplied to the optical modulation unit 111 for each task. The storage unit 223 stores the parameters at the time learning is completed for each task.

[0064] The task control unit 224 described above can also be configured, for example, as shown in Figure 7(A), with a memory 10 and a processor 11 that executes the program stored in the memory 10. Alternatively, the task control unit 224 can be configured with a processing circuit 12, for example, as shown in Figure 7(B). Thus, the task control unit 224 can be realized by a processing circuit network.

[0065] As described above, according to embodiments 1 and 2, by using a pair of light modulation unit 111 and light diffusion / dispersion unit 112, the distance between the pair, or the distance between the light diffusion / dispersion unit 112 and the light receiving unit 113 can be shortened, thereby enabling miniaturization and weight reduction of the optical computing devices 100 and 200 using optical neural networks.

[0066] Furthermore, in embodiments 1 and 2, a simple fixed pattern member can be used as the light diffusion / dispersion section 112, eliminating the need for task-specific design.

[0067] Furthermore, embodiments 1 and 2 can use an optical modulation unit 111 with a lower resolution than conventional ones, such as one with a resolution sufficient for the identification and classification of input light. Therefore, the coupling strength of the optical neural network can be maintained without using a phase modulation element with a very fine pitch as the optical modulation unit 111.

[0068] Furthermore, according to embodiments 1 and 2, even when the optical modulation section 111 has a low resolution, the coupling between layers can be made more complex, enabling high-performance computational processing.

[0069] In the first embodiment described above, as shown in Figure 9, an optical processing unit 100-1 may be provided, and a light distribution layer 115 may be added between the light receiving unit 133 and the light diffusion / dispersion unit 112, which can guide at least a portion of the light emitted from the light diffusion / dispersion unit 112 to match the arrangement of the light receiving area of ​​the light receiving unit 113. The light distribution layer 115 is a refractive lens, a diffracting lens, or a multi-lens array, but is not limited to these. Although not shown, in the second embodiment, the above-mentioned light distribution layer may also be added between the light diffusion / dispersion unit 112#N shown in Figure 8 and the light receiving unit 113.

[0070] 100, 200 Optical computing unit, 110 Optical neural network unit, 111 Optical modulation unit, 112 Optical diffusion and dispersion unit, 113 Light receiving unit, 114 Intermediate layer, 115 Light distribution layer, 120, 220 Judgment unit, 121, 221 Learning unit, 122, 222 Modulation control unit, 123, 223 Memory unit, 224 Task control unit, 240 Optical output device, 241 Light source, 242 Optical modulation unit, 243 Input optical modulation control unit.

Claims

1. An optical computing device comprising: an optical neural network unit that performs computational processing by interfering light with each other and outputs the result of the computational processing as a distribution of light intensity; and a determination unit that determines the result of the computational processing from the distribution, wherein the optical neural network includes at least one intermediate layer comprising a pair of at least one optical modulation unit that modulates light and at least one optical diffusion / dispersion unit that diffuses, disperses or diffracts the light output from the optical modulation unit.

2. The optical computing device according to claim 1, characterized in that the resolution of the optical modulation section is lower than the resolution of the optical diffusion dispersion section.

3. The optical computing device according to claim 1, characterized in that the light diffusion dispersion unit is a phase modulation element that modulates the phase in a fixed manner.

4. The optical computing device according to claim 3, characterized in that the light diffusion dispersion section is a diffraction element that diffracts light in a fixed pattern, a diffuser containing particles that scatter light, or a diffuser plate that diffuses light in a speckled pattern.

5. The optical computing device according to claim 3, wherein the light diffusion and dispersion section comprises a first diffraction element that diffracts in a one-dimensional direction, and a second diffraction element that is tilted around the optical axis from the diffraction direction of the first diffraction element, thereby diffracting the light output from the optical modulation section in two dimensions.

6. A plurality of optical modulation units, each including the optical modulation unit and configured similarly to the optical modulation unit, wherein the diffraction angle θ of a first optical modulation unit included in the plurality of optical modulation units is the maximum diffraction angle θ shown by the relationship between the distance d between the first optical modulation unit and a second optical modulation unit located one position behind the first optical modulation unit in the direction of propagation of light incident on the plurality of optical modulation units, and the length L of the side of the second optical modulation unit in the direction intersecting the direction of propagation. ｍａｘ The following equations (1) and (2) must be satisfied. (1) (2) The optical computing device according to claim 1, characterized by 7. The optical computing device according to any one of claims 1 to 6, wherein the optical modulation unit changes the amount of light modulation in accordance with a given control signal, the optical neural network unit receives input light corresponding to each of a plurality of tasks, and comprises a modulation control unit that changes the control signal given to the optical modulation unit, and a learning unit that learns a control signal given to the optical modulation unit for each of the plurality of tasks such that the difference between the distribution of light intensity indicating the correct result of the calculation process and the output from the optical neural network unit becomes small, corresponding to each of the plurality of tasks.

8. The optical computing device according to claim 7, further comprising a task control unit that selects one task from the plurality of tasks, wherein the modulation control unit inputs a control signal corresponding to the one task to the optical modulation unit.

9. The optical neural network includes the intermediate layer, and includes a plurality of intermediate layers, each configured similarly to the intermediate layer, each of the plurality of optical modulation units provided in the plurality of intermediate layers has a plurality of regions, the plurality of intermediate layers are arranged to be aligned along the direction of the optical axis of the intermediate layer, and the optical modulation unit of a first intermediate layer included in the plurality of intermediate layers and the optical modulation unit of a second intermediate layer included in the plurality of intermediate layers and adjacent to the first intermediate layer are offset in the direction perpendicular to the optical axis by a distance less than the length of each of the plurality of regions perpendicular to the optical axis, characterized in that the optical computing device according to claim 1.

10. The optical neural network includes the intermediate layer, and includes a plurality of intermediate layers, each configured similarly to the intermediate layer, each of the plurality of optical modulation units provided in the plurality of intermediate layers has a plurality of regions, the plurality of intermediate layers are arranged to be aligned along the direction of the optical axis of the intermediate layer, and the optical modulation unit of a first intermediate layer included in the plurality of intermediate layers and the optical modulation unit of a second intermediate layer included in the plurality of intermediate layers and adjacent to the first intermediate layer are offset in a direction rotating with respect to the optical axis, as described in claim 1.

Images