Information processing device, information processing procedure and information processing program

Abstract

An information processing device (1) comprises a neural network with an input layer (10), a physical reservoir layer (20) as an intermediate layer and an output layer (30), an action transmission unit (40) for transmitting an action to the physical reservoir layer (20), thereby causing a change in an internal connection state of the physical reservoir layer corresponding to the action, an adjustment unit (31) for setting weight coefficients (Wj) of the output layer (30) based on an output value of the output layer and a control unit (50, 70) for controlling the operation of the action transmission unit (40).

Description

TECHNICAL AREA

[0001] The present disclosure relates to an information processing device, an information processing method and an information processing program. STATE OF THE ART

[0002] A computer technology called "reservoir computing" has been proposed as a type of hierarchical or interconnected neural network. Reservoir computing reduces the load by replacing a recurrent neural network with a reservoir layer, setting the weight coefficient of the reservoir layer to a fixed value, and performing the learning in only one output.

[0003] Furthermore, a computer technology called “Physical Reservoir Computing” has also been proposed, which implements sophisticated processing such as data classification or time series predictions with respect to the input into an input layer by forming the reservoir layer with a physical reservoir layer (e.g., a material with physical properties, an optical waveguide, or the like) and performing a small amount of learning calculations in the output layer (see, for example, patent references 1 to 3). STATE-OF-THE-ART REFERENCE PATENT LITERATURE Reference 1: Japanese Patent No. 6701247 Reference 2: Japanese Patent No. 7108987 Reference 3: Japanese patent application no. 2022-80891 SUMMARY OF THE INVENTION PROBLEM TO BE SOLVED BY THE INVENTION

[0004] However, under the existing conditions, Physical Reservoir Computing has the problem that it is impossible to change the inference target, since the weighting coefficient of the physical reservoir layer is fixed.

[0005] The purpose of the present disclosure is to provide an information processing device, an information processing method and an information processing program with which it is possible to change the inference target. MEANS TO SOLVE THE PROBLEM

[0006] An information processing device according to the present disclosure comprises a neural network with an input layer, a physical reservoir layer as an intermediate layer, and an output layer; an action transmission unit to transmit an action to the physical reservoir layer, thereby causing a change in an internal connection state of the physical reservoir layer corresponding to the action; an adjustment unit to adjust weight coefficients of the output layer based on an output value of the output layer; and a control unit to control the operation of the action transmission unit.

[0007] An information processing method in the present disclosure is a method executed by an information processing device comprising a neural network with an input layer, a physical reservoir layer as an intermediate layer, and an output layer; and an action transmission unit for transmitting an action to the physical reservoir layer. The information processing method includes a step for adjusting weight coefficients of the output layer based on an output value of the output layer; and a step in which the action transmission unit transmits an action to the physical reservoir layer, thereby causing a change in an internal connection state of the physical reservoir layer corresponding to the action. IMPACT OF THE INVENTION

[0008] According to the present disclosure, the inference target can be switched by changing the internal connection state of the physical reservoir layer. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 is a functional block diagram showing the configuration of an information processing device (learning device) according to a first embodiment. Fig. Figure 2 is a diagram showing an example of the hardware configuration of the information processing device (learning device) according to the first embodiment. Fig. Figure 3 is a schematic diagram showing that the state of a neural network in a physical reservoir layer changes as a result of an action. Fig. Figure 4 is a schematic diagram indicating that the state of the neural network in the physical reservoir layer changes due to a mechanical action (external force). Fig. 5A and Fig. Figure 5B are schematic diagrams indicating that the state of the neural network of the physical reservoir layer changes due to an optical action (light pattern). Fig. 6A and Fig. Figure 6B shows schematic representations that illustrate how the state of the neural network in the physical reservoir layer changes due to a magnetic action (magnetic flux). Fig. Figure 7 is a flowchart showing a learning process of the information processing device (learning device) according to the first embodiment. Fig. Figure 8 is a functional block diagram showing the configuration of an information processing device (inference device) according to the first embodiment. Fig. Figure 9 is a diagram showing an example of the hardware configuration of the information processing device (inference device) according to the first embodiment. Fig. Figure 10 is a flowchart showing an inference process of the information processing device (inference device) according to the first embodiment. Fig. Figure 11 is a functional block diagram showing a further configuration of the information processing device (inference device) according to the first embodiment. Fig. Figure 12 is a functional block diagram showing the configuration of an information processing device (learning device) according to a second embodiment. Fig. Figure 13 is a flowchart showing the learning process of the information processing device (learning device) according to the second embodiment. Fig. Figure 14 is a functional block diagram showing the configuration of an information processing device (inference device) according to the second embodiment. Fig. Figure 15 is a flowchart showing the inference process of the information processing device (inference device) according to the second embodiment. Fig. Figure 16 is a functional block diagram showing a further configuration of the information processing device (inference device) according to the second embodiment. MODE FOR EXECUTING THE INVENTION

[0009] An information processing device, an information processing method, and an information processing program according to each embodiment are described below with reference to the drawings. The following embodiments are only examples, and it is possible to combine embodiments appropriately and to modify each embodiment appropriately. In the drawings, identical components or components with the same function are assigned the same reference numerals. (First embodiment)

[0010] A first embodiment relates to an information processing device that makes it possible to switch the neural network to a different learning object by changing an internal connection state of a specific physical reservoir layer.

[0011] Fig. Figure 1 is a functional block diagram showing the configuration of an information processing device 1 (learning device) according to the first embodiment. The information processing device 1 is a device capable of performing an information processing procedure (learning procedure) according to the first embodiment. A control unit 50 (learning control unit) of the information processing device 1 is, for example, a computer.

[0012] The information processing device 1 comprises a neural network including an input layer 10, a physical reservoir layer 20 as an intermediate layer, an output layer 30, an action transmission unit 40 that transmits an action to the physical reservoir layer 20, thereby causing a change in the internal connection state of the physical reservoir layer 20 corresponding to the assigned action, an adjustment unit 31 that sets weighting coefficients Wj (j: positive integer assigned to each of a plurality of weighting coefficients) of the output layer 30 based on an output value of the output layer 30, and the control unit 50 that controls the operation of the action transmission unit 40.The information processing device 1 is connected to a storage unit 51 as a storage device for storing information. The storage unit 51 can also be part of the information processing device 1.

[0013] In the first embodiment, the weighting coefficients Wj of output layer 30 are updated such that in each neural network, as a result of switching by changing the action (i.e., in each of a multitude of actions), a desired inference result (e.g., an output value close to a target value) is obtained.

[0014] The control unit 50 generates a training model for deriving the output value as an inference result from input information (or input data) regarding the inference target, which is fed into the input layer 10, using training data that includes training input data fed into the input layer 10 and a target value as training data. The control unit 50 stores the weighting coefficients Wj for each of the plurality of actions in the memory unit 51 as parameters of the training model.

[0015] Physical reservoir computing using a physical reservoir layer is described, for example, in the following non-patent references 1 and 2:

[0016] Nicht-Patentreferenz 1: Gouhei Tanaka, „(Keywords You Should Know No. 136) Reservoir Computing“, The Journal of the Institute of Image Information and Television Engineers, Vol. 74, No. 3, pp.532-534, 2020. https: / / www.ite.or.jp / contents / keywords / 2005keyword.pdf Nicht-Patentreferenz 2: Toshiyuki Yamane, „Machine Learning Device by Physical Reservoir Computing and Application to Edge Computing“, ProVISON No. 95, pp.61-65 (Technical Description 5), 2019. https: / / www.ibm.com / downloads / cas / 8MZNZZGX

[0017] The physical reservoir layer 20 is responsible for converting information supplied by the input layer 10 into high-dimensional information according to the intrinsic nonlinear characteristics of its material and structure, and is represented by a recurrent neural network for time-series data. The neural network depends on the internal connection state, which is the state of the connections between virtual nodes in the physical reservoir layer 20.

[0018] In the first embodiment, the internal connection state, defined as the state between virtual nodes present in the material supporting the physical reservoir layer 20 and its structure, is not fixed but is proactively modified (e.g., updated). By transmitting an action to all or part of the physical reservoir layer 20, the action transmission unit 40 enables the modification of a nonlinear characteristic (i.e., nonlinear dynamics) resulting from a deformation, a change of state, or the like of the material supporting the physical reservoir layer 20 and the structure of the physical reservoir layer 20. In other words, a processing task of the information processing device 1, which includes the physical reservoir layer 20, can be switched by changing the action applied to the physical reservoir layer 20.

[0019] Fig. Figure 2 is a diagram showing an example of the hardware configuration of the information processing device 1 (learning device) according to the first embodiment. The information processing device 1 according to the first embodiment comprises, for example, a processor 101, such as a CPU (Central Processing Unit), a memory 102 as a storage device, such as RAM (Random Access Memory), a storage device 103 as a non-volatile storage device, such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and an interface 104. These components correspond to the control unit 50 and the storage unit 51 in Figure 2. Fig. 1. These components can also be configured with dedicated processing circuits.

[0020] The processor 101 is capable of executing a program for processing information according to the first embodiment. The information processing program is acquired, for example, in the form of a recording on a recording medium such as an SD (Secure Digital) memory card or a USB (Universal Serial Bus) memory card, or by downloading it over a network. The in Fig. The hardware configuration shown is just an example, and various modifications to the hardware configuration are possible.

[0021] Fig. Figure 3 is a schematic diagram indicating that the state of the neural network in the physical reservoir layer 20 changes as a result of the action. By transmitting the action via the action transmission unit 40 to all or part of the action receiving area of ​​the physical reservoir layer 20, the internal connection state of the physical reservoir layer 20 can be changed. Furthermore, it is possible to use a physical reservoir layer 20 in which the internal connection state of the physical reservoir layer 20 differs when the action is transmitted via the action transmission unit 40 to the entire action receiving area of ​​the physical reservoir layer 20, compared to the internal connection state of the physical reservoir layer 20 when the action is transmitted via the action transmission unit 40 to only part of the action receiving area (e.g.,a half area, a quarter area, a circular area, a square area, a triangular area, or the like) of the physical reservoir layer 20. It is also possible to use a physical reservoir layer 20 in which the internal connection state depends on the shape, area, or the like of the area of ​​the action receiving surface to which the action is transferred (e.g., half area, 1 / 4 area, circular area, square area, triangular area, or the like), for example, if the action is not required for part of the area of ​​action of the physical reservoir layer 20. The action includes at least one factor of light, electricity, magnetism, heat, external force, and a chemical substance that is transferred (applied) to the physical reservoir layer 20.

[0022] Fig. Figure 4 is a schematic diagram indicating that the state of the neural network in the physical reservoir layer 20 changes due to a mechanical action (external force).

[0023] The Fig. 5A and Fig. Figures 5B are schematic representations showing that the state of the neural network formed by the physical reservoir layer 20 changes due to an optical action (light pattern). In the example in the Fig. 5A and Fig. 5B is the action transmission unit 40, a light irradiation device that irradiates the physical reservoir layer 20 with light. In Fig. 5A makes it possible to apply structured light to a sub-area of ​​the physical reservoir layer 20 and to vary the internal connection state depending on the area of ​​the physical reservoir layer 20 to which the action (light pattern) is transferred. That is, in Fig. 5A is the internal connection state of an area irradiated with structured light, which differs from the internal connection state of an area not irradiated with structured light. The structured light can be generated using an optical element whose sub-area has an optical filter characteristic that transmits, reflects, or absorbs light from a light source. Alternatively, the patterned light can be generated using a spatial light modulation element capable of dynamically switching the light transmission / reflection / absorption properties. The physical reservoir layer 20 can also be a layer that converts the input light into intermediate light, which becomes speckle light through repeated scattering and diffusion of the light within the layer. The light input into the physical reservoir layer 20 can also be scattered light.This is because the interference pattern of the scattered light has the characteristic of changing sensitively depending on the interference conditions.

[0024] Furthermore, the physical system of the physical reservoir layer 20 can also be a system that contains molecules or particles as contained materials within the system and forms a neural network that is optically connected through the contained materials.

[0025] The Fig. 6A and Fig. Figure 6B shows schematic representations indicating that the state of the neural network in the physical reservoir layer 20 changes due to a magnetic action (magnetic flux). In the example in the Fig. 6A and Fig. 6B is the action transfer unit 40 a coil as a magnetic flux action transfer unit that generates a magnetic flux (e.g. magnetic flux density distribution pattern 41) which is to be applied to the physical reservoir layer 20. Fig. Figure 6B shows an example of the change in the state of the neural network formed by the physical reservoir layer 20, using the magnetic flux density distribution pattern, which is the magnetic flux density distribution with a bias or a region where the magnetic flux density is partially high or low (i.e., patterned magnetic flux density), similar to the optical action in Fig. 5A is provided. This means that the internal state of a region where the magnetic flux is strong differs from the internal state of a region where the magnetic flux is weak. The number of coils for transmitting the magnetic flux density distribution pattern can also be three or more.

[0026] Examples of the action transferred to the physical reservoir layer 20 of the information processing device 1 in the first embodiment are shown below. Mechanical actions include pressure, deformation, torsional distortion, and so on. Optical actions include light intensity, wavelength, light plasma, and so on. Electrical actions include an electric field, electric current, voltage, electric charge, etc. Thermal actions include heat absorption, heating, cooling, etc. Magnetic actions include a magnetic field, magnetic force, etc. Chemical actions include liquid molecules, gas molecules, etc.

[0027] Fig. Figure 7 is a flowchart illustrating a learning process of the information processing device 1 (learning device) according to the first embodiment. In the learning process of the information processing device 1, an action is transmitted to the physical reservoir layer 20, and the learning data is input into the input layer 10 (step S11). The weighting coefficients Wj are learned as parameters of the learning model (step S12), and the weighting coefficients are output as parameters (step S13). The output weighting coefficients Wj are stored in the memory unit 51.

[0028] Furthermore, there are cases where a learning model is used to increase the difference between two patterns that are difficult to distinguish and to transform the two patterns into patterns that are easily distinguishable. In such cases, it is possible to create a learning model that determines whether input patterns are identical or not based on a first and a second action. This is achieved by determining the weighting coefficients as parameters of the learning model by performing the learning process while changing the weighting coefficients. Subsequently, an action is transferred to the physical reservoir layer 20, resulting in the first action, which derives a first target value for a first inference target. A further action is then transferred to the physical reservoir layer 20, resulting in the second action, which obtains a second target value for a second inference target.

[0029] Fig. Figure 8 is a functional block diagram showing the configuration of an information processing device 1a (inference device) according to the first embodiment. The information processing device 1a is a device capable of performing an information processing procedure (inference procedure) according to the first embodiment. A control unit 60 (inference control unit) of the information processing device 1a is, for example, a computer. The control unit 60 of the information processing device 1a derives an output value as an inference result output by the output layer 30 from the input information regarding the inference target that was entered into the input layer 10, using the learning model corresponding to the action mediated by the action transmission unit 40. Meanwhile, a storage unit 61 stores the learning model in the storage unit 51 in Fig. If storage unit 1 is used, then storage unit 61 can be the same unit as storage unit 51.

[0030] Furthermore, the control unit 60 can also select a learning model corresponding to the action transmitted by the action transmission unit 40 from learning models for deriving the output value from the input information fed into the input layer 10, which was generated by a learning device using learning data, comprising learning input data and learning training data, and derive the output value as the inference result output by the output layer 30 from the input information regarding the inference target fed into the input layer 10 using the selected learning model. If the information processing device 1a is a learning inference device.

[0031] Fig. Figure 9 is a diagram showing an example of the hardware configuration of the information processing device 1a (inference device) according to the first embodiment. The information processing device 1a according to the first embodiment comprises a processor 111, for example a CPU, a memory 112 as a storage device, for example RAM, a storage device 113 as a non-volatile storage device, for example an HDD or an SSD, and an interface 114. These components correspond to the control unit 60 and the storage unit 61 in Figure 9. Fig. 8. These components can also be formed with dedicated processing circuits.

[0032] The processor 111 is capable of executing a program for processing information according to the first embodiment. The program for processing information is obtained, for example, in the form of a recording on a write medium such as an SD memory card or a USB memory card, or by downloading it over a network. The in Fig. The hardware configuration shown is just an example, and various modifications to the hardware configuration are possible.

[0033] Fig. Figure 10 is a flowchart illustrating an inference process of the information processing device 1a (inference device) according to the first embodiment. In the inference process of the information processing device 1, an action is forwarded to the physical reservoir layer 20, and input data relating to the inference target is fed into the input layer 10 (step S21). The inference is performed using the parameters (weighting coefficients) of the learning model (step S22), and the inference result is output (step S23).

[0034] Furthermore, if a difference is identified between two patterns that are difficult to distinguish, the control unit 60 first switches the inference to a first inference target, loads a condition that reproduces a first action state, transmits the first action state to the physical reservoir layer, and executes the inference. Subsequently, the control unit 60 can switch the inference to a second inference target, load a condition that reproduces a second action state, transmit the second action state to the physical reservoir layer, execute the inference, and make the decision based on the two inference results.

[0035] Fig. Figure 11 is a functional block diagram showing another configuration of an information processing device 1b (inference device) according to the first embodiment. The information processing device 1b is a device capable of performing an information processing procedure (inference procedure) according to the first embodiment. The control unit 60 (inference device) of the information processing device 1b is, for example, a computer. The control unit 60 is capable of performing the inference based on the learning model using the input information and output value mentioned above, without requiring the weighting coefficients Wj of the output layer 30 to be updated, so that an output value is obtained that is closest to a target value.

[0036] As described above, the information processing devices 1, 1a, and 1b, based on Physical Reservoir Computing, make it possible to process various learning and inference processes by changing the action. Furthermore, the information processing devices 1, 1a, and 1b enable highly sensitive data acquisition based on Physical Reservoir Computing. (Second embodiment)

[0037] A second embodiment relates to an information processing device that enables the neural network to be switched to a different learning object by changing the internal connection state of a specific physical reservoir layer 20, and to derive a feature that infers a change in the internal connection state of the physical reservoir layer 20 from the change in the weight coefficients of the output layer, which are updated such that a desired inference result is obtained via the neural network. In the second embodiment, a use different from the concept of conventional physical reservoir computing is employed, as described below.

[0038] In contrast to the first embodiment, in the second embodiment the internal connection state, as the connection state of virtual nodes in the neural network, is changed by transmitting an action to the physical reservoir layer 20, and the mode of the action is derived from the change in the weight coefficients Wj of the output layer 30 after the learning effect caused by the change in the internal connection state. In other words, in the second embodiment the action transmitted to the physical reservoir layer 20 is the input information as the inference target, and the output data determined based on the weight coefficients Wj is the inference result.

[0039] Furthermore, in the second embodiment, it is generally common for there to be a plurality of weighting coefficients Wj, and the second embodiment is more advantageous in cases where the difference between the characteristic of the first state and the characteristic of the second state is identified based on information regarding the weighting coefficients Wj as part of an inference step, compared with cases where the difference in the characteristic between the first state and the second state is identified based on the final output result (inference result) from the output layer 30.

[0040] Furthermore, in the second embodiment, when learning the weighting coefficients Wj of the output layer 30, which is performed to obtain the change in the functional characteristic between the first state and the second state, the input information can be used as an input data set as a specific reference standard and a training data set corresponding to the input data set can be used as training data.

[0041] Furthermore, the training dataset can be switched to one of a variety of different training datasets. Each of the different training datasets can be used to switch the scope or type of the functional characteristic to be obtained.

[0042] Fig. Figure 12 is a functional block diagram showing the configuration of an information processing device 2 (learning device) according to the second embodiment. The information processing device 2 is a device capable of performing an information processing procedure (learning procedure) according to the second embodiment. A control unit 70 (learning control unit) of the information processing device 2 is, for example, a computer.

[0043] The information processing device 2 comprises the neural network, which includes the input layer 10, the physical reservoir layer 20 as an intermediate layer, and the output layer 30; the action transmission unit 40, which assigns an action to the physical reservoir layer 20, thereby causing a change in the internal connection state of the physical reservoir layer 20 corresponding to the transmitted action; the adjustment unit 31, which sets the weighting coefficients Wj (j: positive integer assigned to each of a plurality of weighting coefficients) of the output layer 30 based on the output value of the output layer 30; and the control unit 70, which controls the operation of the action transmission unit 40. The information processing device 2 is connected to a storage unit 71 as a storage device for storing information.The storage unit 71 can also be part of the information processing device 2.

[0044] The control unit 70 of the information processing device 2 generates a learning model for deriving an inference result based on the weighting coefficients Wj from the action assigned to the physical reservoir layer 20, using training data that includes training action data and training data, and stores the weighting coefficients Wj for each of the multiple actions in the storage unit 71 as parameters of the learning model. The hardware configuration of the information processing device 2 is the same as in the first embodiment.

[0045] Fig. Figure 13 is a flowchart illustrating a learning process of the information processing device 2 (learning device) according to the second embodiment. In the learning process of the information processing device 2, the learning input data is entered into the input layer 10, and an action is transmitted to the physical reservoir layer 20 as learning data (step S31). The weighting coefficients Wj are learned as parameters of the learning model (step S32), and the weighting coefficients are output as parameters (step S33). The output weighting coefficients Wj are stored in the memory unit 71.

[0046] Furthermore, there are cases where a learning model is used to increase the difference between two patterns that are difficult to distinguish from each other and to convert the two patterns into patterns that are easy to distinguish from each other.In such cases, it is possible to create a learning model that is able to determine whether input actions match or not by setting the internal connection state of the physical reservoir layer 20 to the first state, by transferring an action to the physical reservoir layer 20, performing the learning while changing the weighting coefficients, obtaining a first weighting coefficient group Wg1, setting the internal connection state of the physical reservoir layer 20 to the second state, by transferring another action to the physical reservoir layer 20, performing the learning while changing the weighting coefficients, obtaining a second weighting coefficient group Wg2, and comparing the first weighting coefficient group Wg1 and the second weighting coefficient group Wg2.

[0047] Fig. Figure 14 is a functional block diagram showing the configuration of an information processing device 2a (inference device) according to the second embodiment. The information processing device 2a is a device capable of performing an information processing procedure (inference procedure) according to the second embodiment. A control unit 60 (inference control unit) of the information processing device 2a is, for example, a computer. The control unit 80 of the information processing device 2a derives the output value as the inference result, which consists of the weighting coefficients from the action as the inference target, using a learning model corresponding to the action mediated by the action transmission unit 40. A data processing unit 82 outputs the inference result from the weighting coefficients. The inference result is stored in a storage unit 81.While memory unit 81 stores the learning model in memory unit 71 in . Fig. If storage unit 12 stores the data, storage unit 81 can be the same unit as storage unit 71. Data processing unit 82 is able to calculate an action level representing the intensity of the action transmitted to the physical reservoir layer, based on the difference between a first weighting coefficient when the internal connection state of physical reservoir layer 20 is the first state, and a second weighting coefficient when the internal connection state of physical reservoir layer 20 is the second state, which differs from the first state.

[0048] Furthermore, the control unit 60 can also select a learning model, according to the action transmitted by the action transmission unit 40, from learning models for deriving the output value from the input information entered into input layer 10, which is generated by a learning device that uses learning data, comprising learning input data and training data, and deriving the output value as the derivation result that is output by the output layer 30 from the input information regarding the derivation target that is entered into input layer 10, using the selected learning model. In this case, the information processing device 1a is a learning inference device.

[0049] Fig. Figure 15 is a flowchart illustrating the inference process of the information processing device 2a (inference device) according to the second embodiment. In the inference process of the information processing device 2a, input data is fed into the input layer 10 and an action as the inference target is transmitted to the physical reservoir layer 20 (step S41), the inference is performed using the parameters (weighting coefficients) of the learning model (step S42) and the inference result is output (step S43).

[0050] Fig.Figure 16 is a functional block diagram showing a further configuration of an information processing device 2b (inference device) according to the second embodiment. If the control unit 80 updates the learning model based on the action transmitted by the action transmission unit 40 and the inference result, without requiring the weighting coefficients Wj of the output layer 30 to be updated, an output value is obtained that is closest to the target value.

[0051] As described above, the information processing devices 2, 2a, and 2b, which are based on Physical Reservoir Computing, make it possible to handle various learning and inference processes by defining the action as the inference target. Furthermore, the information processing devices 2, 2a, and 2b enable highly sensitive detection based on Physical Reservoir Computing.

[0052] Furthermore, the information processing devices in the first and second embodiments and their modifications described above (hereinafter simply referred to as "the embodiments described above") can be modified appropriately. For example, a modification, addition, or removal of a component with respect to the physical reservoir layer 20 and the action transmission unit 40 can be made in the embodiments described above. Moreover, features or components of the embodiments described above can be appropriately combined in a mode that differs from one of the modes described above. DESCRIPTION OF REFERENCE MARKS

[0053] 1, 2: Information processing device (learning device), 1a, 1b, 2a, 2b: Information processing device (inference device), 10: Input layer, 20: Physical reservoir layer (intermediate layer), 30: Output layer, 40: Action transmission unit, 50, 70: Control unit (learning control unit), 51, 71: Storage unit, 60, 80: Control unit (inference control unit), 61, 81: Storage unit, Wj: Weighting coefficient. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] JP 2022-80891

[0003] Cited non-patent literature

[0000] Gouhei Tanaka, „(Keywords You Should Know No. 136) Reservoir Computing“, The Journal of the Institute of Image Information and Television Engineers, Vol. 74, No. 3, pp.532-534, 2020. https: / / www.ite.or.jp / contents / keywords / 2005keyword.pdf Nicht-Patentreferenz

[0016] Toshiyuki Yamane, „Machine Learning Device by Physical Reservoir Computing and Application to Edge Computing“, ProVISON No. 95, pp.61-65 (Technical Description 5), 2019. https: / / www.ibm.com / downloads / cas / 8MZNZZGX

[0016]

Claims

[1] Information processing device, comprising: a neural network comprising an input layer, a physical reservoir layer as an intermediate layer, and an output layer; an action transmission unit for transmitting an action to the physical reservoir layer, thereby causing a change in an internal connection state of the physical reservoir layer corresponding to the action; an adjustment unit for setting weighting coefficients of the output layer based on an output layer output value; and a control unit for controlling the operation of the action transmission unit. [2] Information processing device according to claim 1, wherein the control unit a learning model is generated to derive the output value from the input information fed into the input layer using learning data that includes learning input data and learning training data, and The weighting coefficients are stored for each of a multitude of actions in a storage unit as parameters of the learning model. [3] Information processing device according to claim 2, wherein the control unit derives the output value as an inference result output by the output layer from the input information relating to an inference target that was entered into the input layer, using the learning model that corresponds to the action mediated by the action transmission unit. [4] Information processing device according to claim 3, wherein the control unit updates the learning model using the input information and the output value. [5] Information processing device according to claim 1, wherein the control unit selects a learning model, according to the action transferred by the action transfer unit, from learning models for deriving the output value from input information fed into the input layer generated by a learning device using learning data, comprising learning input data and learning training data, and the output value is derived as an inference result output by the output layer from the input data regarding an inference target that was entered into the input layer, using the selected learning model. [6] Information processing device according to claim 1, wherein the control unit a learning model is generated to derive an inference result based on the weighting coefficients from the action applied to the physical reservoir layer, using learning data, including learning action data and learning training data, and The weighting coefficients are stored for each of a multitude of actions in a storage unit as parameters of the learning model. [7] Information processing device according to claim 6, wherein the control unit derives the inference result based on the weighting coefficients from the action mediated by the action transmission unit using the learning model. [8] Information processing device according to claim 7, wherein the control unit updates the learning model based on the action mediated by the action transmission unit and the inference result. [9] Information processing device according to claim 1, wherein the control unit a learning model is selected, according to the action transferred by the action transfer unit, from learning models to derive an inference result based on the weighting coefficients from the action transferred to the physical reservoir layer, generated by a learning device using learning data, comprising learning action data and learning training data, and derives the inference result based on the weighting coefficients from the action transferred by the action transfer unit using the selected learning model. [10] Information processing device according to any one of claims 1 to 9, wherein the physical reservoir layer includes an optical internal connection state as a component of the neural network, and the internal connection state changes due to the action acting on the physical reservoir layer, including an optical factor. [11] Information processing device according to claim 10, wherein Specklelight is introduced from the inlet layer into the physical reservoir layer, and The physical reservoir layer generates intermediate light of the speckle light by repeating the scattering and diffusion of light within the layer. [12] Information processing device according to any one of claims 1 to 9, wherein the action comprises at least one factor of light, electricity, magnetism, heat, external force and a chemical substance which is transferred to the physical reservoir layer. [13] Information processing device according to any one of claims 1 to 12, wherein the changing of the internal connection state of the physical reservoir layer varies depending on the area of ​​the physical reservoir layer to which the action is transferred. [14] Information processing device according to one of claims 1 to 12, wherein the action is transferred to a part of the area of ​​the physical reservoir layer. [15] Information processing device according to any one of claims 1 to 14, further comprising a data processing unit for calculating an action level representing the intensity of the action exerted on the physical reservoir layer, based on the difference between a first weighting coefficient when the internal connection state of the physical reservoir layer is a first state, and a second weighting coefficient when the internal connection state of the physical reservoir layer is a second state that differs from the first state. [16] Information processing procedure performed by an information processing device comprising a neural network comprising an input layer, a physical reservoir layer as an intermediate layer, and an output layer; and an action transmission unit for transmitting an action to the physical reservoir layer, comprising the information processing procedure: a step to adjust the weighting coefficients of the output layer based on an output layer initial value; and a step to cause the action transfer unit to transmit an action to the physical reservoir layer, thereby causing a change in an internal connection state of the physical reservoir layer corresponding to the action. [17] Program for processing information that causes a computer to comprehensively a neural network comprising an input layer, a physical reservoir layer as an intermediate layer, and an output layer; and an action transmission unit that transmits an action to the physical reservoir layer, executes: a step to adjust the output layer's weighting coefficients based on an output layer initial value; and a step to cause the action transmission unit of the physical reservoir layer to transmit an action, thereby causing a change corresponding to the action in an internal connection state of the physical reservoir layer.

Images