Information processing device, information processing method, and program

Abstract

To provide an information processing apparatus, an information processing method, and a program capable of improving the performance of a machine learning model while reducing the huge learning cost of the machine learning model.SOLUTION: An information processing apparatus according to an embodiment includes: an acquisition unit that acquires time series data to be tested; and a processing unit that recognizes an object from the time series data to be tested by inputting the time series data to be tested to reservoir computing in which an input layer, at least one reservoir layer, and an output layer are included, wherein the processing unit controls the degree of influence exerted on the reservoir layer from the time series data input via the input layer.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] This invention relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] Currently, the mainstream image recognition technology is convolutional neural networks. Convolutional neural networks are machine learning models that boast high recognition performance, far surpassing image recognition technologies that existed before they became mainstream. On the other hand, achieving high performance requires training that incurs enormous computational costs, making their operation difficult in situations where computational resources are limited, such as in embedded systems. In recent years, low-training machine learning models called reservoir computing have attracted attention as an alternative to machine learning models that require enormous training costs, such as convolutional neural networks, and research on utilizing these for image recognition has been reported. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Herbert Jaeger, Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the "echo state network" approach, GMD-Forschungszentrum Informationstechnik, 2002. [Non-Patent Document 2] N. Schaetti, M. Salomon, and R. Couturier, “Echo state networks-based reservoir computing for MNIST handwritten digits recognition”, IEEE International Conference on Computational Science and Engineering, August 2016. [Non-Patent Document 3] Q. An, K. Bai, L. Liu, F. Shen, and Y. Yi, “A unified information perceptron using deep reservoir computing”, Computers and Electrical Engineering, vol. 85, July 2020. [Non-Patent Document 4] Z. Tong and G. Tanaka, “Reservoir computing with untrained convolutional neural networks for image recognition”, 2018 24th International Conference on Pattern Recognition, pp. 1289-1294, August 2018. [Non-Patent Document 5] Y. Yonemura and Y. Katori, “Image recognition model based on convolutional reservoir computing”, The 34th Annual Conference of the Japanese Society for Artificial Intelligence, June 2020. [Overview of the project] [Problems that the invention aims to solve]

[0004] However, while previous studies have achieved a certain level of performance in simple handwritten digit recognition tasks, there is room for improvement in tasks such as clothing image recognition and image recognition of animals and cars. Furthermore, these challenges are not limited to image data, but are common to all time-series data that can be input to machine learning models to which reservoir computing is applied.

[0005] One aspect of the present invention has been made in consideration of these circumstances, and one of its objectives is to provide an information processing device, an information processing method, and a program that can improve the performance of a machine learning model while reducing the enormous training cost of the machine learning model. [Means for solving the problem]

[0006] One aspect of the present invention is an information processing device that recognizes an object from a time-series data of a test by inputting the time-series data of a test to the reservoir computing device, which is trained on a training dataset in which the time-series data of a training dataset is labeled with the correct object to be recognized from the time-series data of the training dataset, the reservoir computing device being trained on a training dataset in which the time-series data of a training dataset is labeled with the correct object to be recognized from the time-series data of the training dataset, the reservoir computing device being trained on a training dataset in which the time-series data of a training dataset is labeled with the correct object to be recognized from the time-series data of the training dataset, the reservoir computing device being trained on a training dataset in which the time-series data of a training dataset is labeled with the correct object to be recognized from the time-series data of the training dataset, the reservoir computing device being trained on a training dataset in which the time-series data of a training dataset is labeled with the correct object to be recognized from the time-series data of the training dataset, the reservoir computing device being trained on a time-series data of [Effects of the Invention]

[0007] According to the above embodiment, it is possible to improve the performance of a machine learning model while reducing the enormous training cost of the machine learning model. [Brief explanation of the drawing]

[0008] [Figure 1] This diagram shows an example of the configuration of the information processing system 1 according to the embodiment. [Figure 2] This figure shows an example of the configuration of the information processing device 100 according to the embodiment. [Figure 3] This is a diagram illustrating the overview of the reservoir computing MDL according to the embodiment. [Figure 4] This diagram illustrates a typical example of a reservoir computing MDL. [Figure 5] This is a flowchart showing an example of the specific processing flow of the information processing device 100 according to the embodiment. [Figure 6] This figure shows an example of time-series data provided to the reservoir layer LR of the reservoir computing MDL according to the embodiment. [Figure 7] This diagram shows the internal configuration of the reservoir computing MDL according to this embodiment. [Figure 8] This figure shows an example of the configuration of the learning device 200 according to the embodiment. [Figure 9] This is a flowchart illustrating an example of the specific processing flow of the learning device 200 according to the embodiment. [Figure 10] This diagram illustrates the learning method for reservoir computing MDL. [Figure 11] This diagram shows an example of the parameters for each configuration of the reservoir computing MDL. [Figure 12] This figure shows a comparison between the accuracy rate of this embodiment and the accuracy rate of the conventional method. [Figure 13] This figure shows the input images for each dataset alongside the feature images output by the reservoir computing MDL of this embodiment. [Figure 14] This figure shows the input images for each dataset alongside the feature images output by the reservoir computing MDL of this embodiment. [Figure 15] This figure shows the input images for each dataset alongside the feature images output by the reservoir computing MDL of this embodiment. [Modes for carrying out the invention]

[0009] Hereinafter, embodiments of the information processing apparatus, information processing method, and program of the present invention will be described with reference to the drawings.

[0010] [Configuration of the Information Processing System] Figure 1 is a diagram showing an example of the configuration of an information processing system 1 according to an embodiment. The information processing system 1 according to an embodiment includes, for example, an information processing device 100 and a learning device 200. These devices are connected via a network NW. The network NW is, for example, a WAN (Wide Area Network) or a LAN (Local Area Network).

[0011] The information processing device 100 is a device that processes various tasks such as recognition, classification, and prediction using a machine learning model called Reservoir Computing MDL. In other words, Reservoir Computing MDL can be applied to a variety of tasks. As an example, the following explanation will describe how Reservoir Computing MDL is applied to the task of image recognition.

[0012] The information processing device 100 may be a single device, or it may be a system in which multiple devices connected via a network (NW) cooperate with each other. In other words, the information processing device 100 may be implemented by multiple computers (processors) included in a distributed computing system or a cloud computing system. Below, as an example, the information processing device 100 will be described as a single device.

[0013] The learning device 200 is a device that learns the reservoir computing MDL used by the information processing device 100. The learning device 200 may also be a single device, or it may be a system in which multiple devices connected via a network NW work together.

[0014] [Configuration of the information processing device] Figure 2 is a diagram showing an example of the configuration of an information processing device 100 according to an embodiment. As shown in Figure 2, the information processing device 100 includes, for example, a communication unit 102, an input unit 104, an output unit 106, a processing unit 110, and a storage unit 130.

[0015] The communication unit 102 may include, for example, a NIC (Network Interface Card) or a wireless communication module including a receiver and a transmitter. The communication unit 102 may communicate with the learning device 200 or other devices via a network NW.

[0016] The input unit 104 receives various input operations from the user, converts the received input operations into electrical signals, and outputs them to the processing unit 110. For example, the input unit 104 includes a mouse, keyboard, trackball, switch, button, joystick, touch panel, etc. The input unit 104 may also be a user interface that receives audio input, such as a microphone.

[0017] The output unit 106 includes, for example, a display and a speaker. The display shows images generated by the processing unit 110, as well as a GUI (Graphical User Interface) for receiving various input operations from the user. For example, the display may be an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence) display. The speaker outputs the information input from the processing unit 110 as sound.

[0018] The processing unit 110 includes, for example, an acquisition unit 112, an image processing unit 114, a task processing unit 116, and an output control unit 118.

[0019] These components of the processing unit 110 are realized, for example, by a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) executing a program stored in the memory unit 130. Some or all of the components of the processing unit 110 may be realized by hardware such as an LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), or FPGA (Field-Programmable Gate Array), or by the cooperation of software and hardware.

[0020] The memory unit 130 is implemented by, for example, an HDD (Hard Disk Drive), flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), ROM (Read Only Memory), or RAM (Random Access Memory). Various programs, such as firmware and application programs, are stored in the memory unit 130. In addition to one or more programs referenced by the processor, the memory unit 130 also stores model data. Model data is information (programs or algorithms) that defines the reservoir computing MDL.

[0021] Figure 3 is a diagram illustrating the overview of the reservoir computing MDL according to the embodiment. The reservoir computing MDL according to the embodiment includes, for example, an input layer L IN and multiple reservoir layers L R (L in the example shown) R1、 L R2 ) and output layer L OUT This includes the input layer L of the reservoir computing MDL when applied to the task of image recognition. IN An image is input to this. Multiple reservoir layers L Rextracts high-dimensional features related to an object (numbers in the illustrated example) from an input image. Output layer L OUT combines the features extracted by each of the plurality of reservoir layers L R , converts the combined features into an image recognition result, and outputs the result. The basic processing of the reservoir layer L R will be described below.

[0022] FIG. 4 is a diagram illustrating an example of a typical reservoir computing model. As illustrated, the reservoir computing model includes an input layer L IN , one reservoir layer L R , and an output layer L OUT .

[0023] Time-series data u(t) is input to the input layer L IN . The time-series data u(t) includes a plurality of observation data obtained by continuously observing temporal changes of a certain phenomenon. For example, the time-series data u(t) is a multidimensional vector whose elements include temporally continuous observation data such as observation data u(t) observed at a certain time t, observation data u(t+1) observed at time t+1, and observation data u(t+2) observed at time t+2. The time-series data u(t) input to the input layer L IN is output to the reservoir layer L R . At this time, the time-series data u(t) is weighted by a certain connection weight matrix W IN between the input layer L R and the reservoir layer L in . The connection weight matrix W in is initialized with random numbers and is not updated through learning.

[0024] Each reservoir layer L R non-linearly converts the weighted time-series data u(t) output from the input layer L IN into higher-dimensional time-series data. The reservoir layer L R is an intermediate layer that performs recursive processing having sparse and random connections (nodes). Specifically, the reservoir layer L RThis is the recurrent coupling load matrix W rec The time series data u(t) is processed based on the recurrent bond weight matrix W. rec The bond weight matrix W in Similarly, it is initialized with random numbers and is not updated by learning. However, the recurrent bond weight matrix W rec It is sparsified to satisfy an arbitrary bond density ρ and scaled to satisfy the Echo State Property.

[0025] Reservoir layer L R This process stores and remembers past input information in order to process it based on the temporal relationships (dependencies) of time-series data u(t). This reservoir layer L stores such past input information. R The internal state of the reservoir layer L can be represented by the state vector x(t). The state vector x(t) is R This represents the recursive processing result, and the subsequent output layer L OUT This represents the data itself that is output.

[0026] Output layer L OUT The reservoir layer L R The state vectors x(t) output at each time step are linearly combined, and the result y(t) of that linear combination is output. When the state vectors x(t) are linearly combined, a certain combination weight matrix W is used. out The weights are assigned by the following: Joint weight matrix W OUT This is the only thing to learn in reservoir computing MDL.

[0027] [Processing flow of information processing equipment] The specific processing flow of the information processing device 100 will be explained below with reference to the flowchart. Figure 5 is a flowchart showing an example of the specific processing flow of the information processing device 100 according to the embodiment. The processing in this flowchart may be repeated at a predetermined cycle. Furthermore, if the information processing device 100 is implemented by multiple computers included in a distributed computing system or a cloud computing system, some or all of the processing in this flowchart may be processed in parallel by the multiple computers.

[0028] First, the acquisition unit 112 acquires the input image M0 that is the subject of the image recognition test (step S100). For example, the acquisition unit 112 may acquire the input image M0 from another device (e.g., network storage) via the communication unit 102. Also, if the input image M0 is stored in the storage unit 130, the acquisition unit 112 may read the input image M0 from the storage unit 130. Furthermore, if a non-transient storage medium (e.g., fresh memory) containing the input image M0 is connected to the drive device of the information processing device 100, the acquisition unit 112 may read the input image M0 from the storage medium.

[0029] Next, the image processing unit 114 generates time-series data u(t) from the input image M0 acquired by the acquisition unit 112 (step S102).

[0030] Figure 6 shows the reservoir layer L of the reservoir computing MDL according to the embodiment. R This figure shows an example of time-series data to be given. First, the image processing unit 114 extracts a local region M1 from the input image M0. The shape of this extracted local region M1 is (C,K h ,K w ) is represented by . C represents a channel, and K h ,K w The vertical and horizontal dimensions are represented. If the input image M0 is a color image with three colors (red, green, and blue), channel C will be 3, and if the input image M0 is a grayscale image, channel C will be 1. Next, the image processing unit 114 cuts out the local region M1 into strips. For example, the image processing unit 114 cuts out Kw The local region M1 is cut out in strips along the axial direction, one or several pixels at a time. Each of the strips cut out from the local region M1 will be referred to as strip region M2. The image processing unit 114 then uses the strip region M2 as time-series data u(t). In other words, the time-series data u(t) is C × K h Dimension K w This will be time-series data. Note that the time-series data u(t) is C × K w Dimension K h Time-series data of specific times is also acceptable.

[0031] Returning to the explanation of the flowchart in Figure 5, the task processing unit 116 inputs the time-series data u(t) generated from the input image M0 by the image processing unit 114 to the reservoir computing MDL (step S104).

[0032] Figure 7 is a diagram showing the internal configuration of the reservoir computing MDL according to this embodiment. The reservoir computing MDL shown in Figure 7 is the preceding part (reservoir layer L) of the reservoir computing MDL exemplified in Figure 3. R1 This corresponds to the reservoir computing MDL in the preceding stage, for example, the input layer L IN and binding layer L C Multiple reservoir layers L R1 This is a parallel-connected model. The coupled layer L C This refers to multiple reservoir layers L R1 This is an intermediate layer that combines the respective outputs x. For example, a reservoir layer L. R1 If there are N such elements, the binding layer L C This refers to each reservoir layer L from 1 to N. R1 The outputs x1 to xN are combined. In the example in Figure 7, multiple input layers L IN The diagram schematically illustrates this, but it is a single input layer L IN That is the case.

[0033] Reservoir layer L in this embodiment R1 The time evolution of the state vector x(t) is calculated by equation (1), where x(0)=0.

[0034]

number

[0035] In equation (1), f represents the activation function, and for example, the hyperbolic tangent (tanh) can be used. δ is the reservoir layer L R1 This represents the control coefficient (hereinafter referred to as the leakage rate) that controls the update rate of the state vector x(t). The maximum value of the leakage rate δ is 1. Multiple reservoir layers L R1 Each of these is assigned a different leakage rate δ. For example, the uppermost reservoir layer L R1 A leakage rate δ1 is set, and the bottom reservoir layer L R1 The leakage rate δ N This is set. The leakage rate δ is an example of a "time constant".

[0036] Equation (1) represents a reservoir layer L with a certain state vector x(t). R1 When observation data u(t) at a certain time t is input to the reservoir layer L R1 The degree to which it is influenced by the observed data u(t), and the reservoir layer L when the observed data u(t) at time t is input. R1 Based on the state vector x(t) and the leakage rate δ, the reservoir layer L is input when the observed data u(t+1) at time t+1 is input. R1 The state vector x(t+1) is being calculated.

[0037] Reservoir layer L of state vector x(t) R1 The degree to which it is influenced by the observed data u(t) is the bond weight matrix W in The observed data u(t) at time t weighted by (i.e., W) in u(t) and the recurrent coupling weight matrix W rec The reservoir layer L is weighted by R1 State vector x(t) (i.e., W rec It can be expressed as a sum with x(t). Below, W in u(t) and W recThe term that is added with x(t) is referred to as the influence term and will be explained accordingly.

[0038] As shown in equation (1), based on the sum of the state vector x(t) multiplied by (1-δ) and the influence term multiplied by δ, the reservoir layer L at the next time t+1 is determined. R1 The state vector x(t+1) is calculated. As the leakage rate δ approaches zero, the influence term approaches zero, and the state of the reservoir layer L at the next time t+1 is calculated. R1 The state vector x(t+1) is the state of the reservoir layer L at the previous time t. R1 The state vector x(t) approaches the state vector x(t). On the other hand, as the leakage rate δ increases, the influence term also increases, and the reservoir layer L at the next time t+1 R1 The state vector x(t+1) of the reservoir layer L at the current time t is R1 This results in a significant deviation from the state vector x(t). By appropriately setting or adjusting the leakage rate δ, the accuracy of reservoir computing MDL can be improved.

[0039] When the leakage rate δ is reduced, the reservoir layer L R1 Because the internal state changes proceed relatively slowly (the state update is slow), reservoir layer L R1 In this case, low spatial frequency features are extracted. On the other hand, when the leakage rate δ is increased, the reservoir layer L R1 Because the internal state changes of the reservoir layer L proceed rapidly (because the state is updated quickly), R1 High spatial frequency features are extracted in this case.

[0040] Therefore, for example, a reservoir layer L with a leakage rate δ set to 0.1 R1 In the input layer L IN When time-series data u(t) is input via this method, low spatial frequency features are extracted from the time-series data u(t), and a reservoir layer L with a leakage rate δ set to 0.9 is created. R1 In the input layer L IN When time-series data u(t) is input via this, high spatial frequency features are extracted from the time-series data u(t). In this way, multiple reservoir layers L R1By setting different leakage rates δ for each of the s, features of different frequencies can be extracted from each other.

[0041] All the cut-out local regions M1 are formed into strips and input to the input layer L IN as time-series data u(t), and when all input is completed, in the coupling layer L of the reservoir computing MDL C , the reservoir layer L R1 couples all the output state vectors x(t) output thereby into one, resulting in a feature image in which only features of high spatial frequency and low spatial frequency of the input image are extracted. The coupling layer L C develops the feature image in the channel direction and outputs it to a plurality of subsequent reservoir layers L R2 . When the number of reservoir layers L R1 is N, and the number of nodes of each reservoir layer L R1 is R, the channel size of the output image output to the subsequent reservoir layer L R2 is N×R.

[0042] Each subsequent reservoir layer L R2 , similarly to the preceding reservoir layer L R1 , extracts features of mutually different frequencies such as low spatial frequency and high spatial frequency from the feature image having each spatial frequency. The subsequent reservoir layer L R2 , similarly to the preceding reservoir layer L R1 performs the calculation of formula (1) to extract features from the feature image.

[0043] The output layer L OUT couples the features extracted by each subsequent reservoir layer L R2 , converts the coupled features into a recognition result of the target object, and outputs the result.

[0044] Returning to the description of the flowchart in FIG. 5. Next, the task processing unit 116 acquires the recognition result of the target object from the reservoir computing MDL to which the time-series data u(t) has been input (step S106).

[0045] Next, the output control unit 118 outputs the recognition result of the object (step S108). For example, the output control unit 118 may display the recognition result on the display of the output unit 106. Alternatively, the output control unit 118 may transmit the recognition result to another device via the communication unit 102. This completes the processing of this flowchart.

[0046] [Configuration of the learning device] Figure 8 is a diagram showing an example of the configuration of a learning device 200 according to an embodiment. The learning device 200 includes, for example, a communication unit 202, a processing unit 210, and a storage unit 230.

[0047] The communication unit 202 may include, for example, a NIC or a wireless communication module including a receiver and a transmitter. The communication unit 202 may communicate with the information processing device 100 or other devices via a network NW.

[0048] The processing unit 210 includes, for example, an acquisition unit 212, a learning unit 214, and a communication control unit 216.

[0049] These components of the processing unit 210 are realized, for example, by a processor such as a CPU or GPU executing a program stored in the memory unit 230. Some or all of the components of the processing unit 210 may be realized by hardware such as an LSI, ASIC, or FPGA, or by the cooperation of software and hardware.

[0050] The memory unit 230 is implemented by, for example, an HDD, flash memory, EEPROM, ROM, RAM, etc. Various programs such as firmware and application programs are stored in the memory unit 230. In addition to one or more programs referenced by the processor, the memory unit 230 also stores model data that defines unlearned reservoir computing MDLs, etc.

[0051] [Processing flow of the learning device (training using reservoir computing)] The specific processing flow of the learning device 200 will be explained below with reference to the flowchart. Figure 9 is a flowchart showing an example of the specific processing flow of the learning device 200 according to the embodiment. The processing in this flowchart may be repeated at a predetermined cycle. Furthermore, if the learning device 200 is implemented by multiple computers included in a distributed computing system or a cloud computing system, some or all of the processing in this flowchart may be processed in parallel by the multiple computers.

[0052] First, the acquisition unit 212 acquires a training dataset for training the reservoir computing MDL (step S200). The training dataset is a dataset in which the time series data to be trained is labeled with the correct (ideal) object that should be recognized from the time series data to be trained. The correct object may be interpreted as the object to which the training signal is associated. For example, suppose time series data is generated from an input image with a certain number drawn on it using the method described in Figure 6. In this case, the training dataset is a dataset in which the time series data derived from the input image is labeled with the number drawn on the input image. Thus, the training dataset is a dataset in which the time series data to be trained is used as input data and the correct recognition result of the object obtained from the time series data to be trained is used as output data.

[0053] For example, the acquisition unit 212 may acquire training datasets from other devices (e.g., open data sources on a network) via the communication unit 202. Also, if training datasets are stored in the storage unit 230, the acquisition unit 212 may read the training datasets from the storage unit 230. Furthermore, if a non-transient storage medium (e.g., fresh memory) containing training datasets is connected to the drive device of the learning device 200, the acquisition unit 212 may read the training datasets from the storage medium.

[0054] Next, the learning unit 214 learns the reservoir computing MDL using the training dataset acquired by the acquisition unit 212 (step S202).

[0055] Figure 10 is a diagram illustrating the learning method of the reservoir computing MDL. For example, the learning unit 214 inputs the time series data u(t) of the target to be trained, which is included as input data in the training dataset, into the reservoir computing MDL, calculates the error between the recognition result of the object output by the reservoir computing MDL at that time and the correct recognition result, and performs ridge regression or the like to reduce this error, thereby improving the output layer L OUT The joint load matrix W in OUT Adjust the output layer L. OUT For example, this is a fully connected layer. This allows the reservoir computing MDL to output the recognition result of the correct object that should be recognized from the time series data when time series data is input to it.

[0056] Next, the communication control unit 216 transmits the model data defining the reservoir computing MDL learned by the learning unit 214 to the information processing device 100 via the communication unit 202 (step S204). This completes the processing in this flowchart.

[0057] According to the embodiments described above, there are multiple reservoir layers L R By setting different leakage rates δ for each of these factors, it is possible to improve the performance of the machine learning model while reducing the enormous training cost of the machine learning model.

[0058] Specifically, a reservoir layer L of a certain state vector x(t) R When observation data u(t) at a certain time t is input to the reservoir layer L R The degree to which it is influenced by the observed data u(t), and the reservoir layer L when the observed data u(t) at time t is input. RBased on the state vector x(t) and the leakage rate δ, the reservoir layer L is input when the observed data u(t+1) at time t+1 is input. R By calculating the state vector x(t+1), it is possible to improve the performance of the machine learning model while reducing the enormous training cost of the machine learning model.

[0059] Furthermore, according to the above embodiment, a local region of the input image is cut out in strips, and each of these strips is used as time-series data. Compared to the case where the local region is treated directly as time-series data, the reservoir layer L R This reduces the number of nodes, thereby lowering computational costs and memory resources. Generally, the larger the dimensionality and time length of time series data, the more reservoir layers L are needed. R Because the amount of input information that must be stored increases, the reservoir layer L R The number of nodes needs to be increased. As a result, computational costs and memory resources increase. On the other hand, as in this embodiment, when a local region is cut out into strips and each of these strips is used as time-series data, the number of dimensions and time length of the time-series data can be reduced compared to when the entire image is input to the reservoir, thus reducing the reservoir layer L R This reduces the amount of input information that needs to be stored, and the reservoir layer L R This allows us to reduce the number of nodes. As a result, we can reduce computational costs and memory resources.

[0060] [Examples] The following examples illustrate experimental results using the reservoir computing MDL of the above embodiment. Figure 11 shows an example of the parameters of each component of the reservoir computing MDL. Reservoir layer L RThe parameters were set as N=5, R=30, ρ=0.8 (however, when a color image was input, only the first layer had R=12, ρ=0.5). Evaluation was performed using the MNIST dataset, the Fashion-MNIST dataset, and the CIFAR-10 dataset. In training, the training dataset for each dataset was fed to the reservoir computing MDL of the embodiment, and only the connection weights of the fully connected layer (Fully Connected 1) were optimized by ridge regression. In the experiment, the training dataset for each dataset was fed to the trained reservoir computing MDL, and the accuracy was calculated.

[0061] Figure 12 shows a comparison of the accuracy rate of this embodiment with that of the conventional method. [2] to [5] in the figure represent the accuracy rates of the respective methods described in Non-Patent Documents 2 to 5. Multiple reservoir layers L R In this embodiment, the method, which sets different leakage rates δ for each of the datasets, was lower than the accuracy rate for each dataset using the conventional method [2-5], but it was higher than the accuracy rate of the conventional method for the Fashion-MNIST dataset and the CIFAR-10 dataset.

[0062] Figures 13 to 15 show the input images of each dataset alongside the feature images output by the reservoir computing MDL of this embodiment. Figure 13(a) shows the input image of the MNIST dataset, figure 13(b) shows the feature image when the leakage rate δ = 0.1, and figure 13(c) shows the feature image when the leakage rate δ = 0.9. Figure 14(a) shows the input image of the Fashion-MNIST dataset, figure 14(b) shows the feature image when the leakage rate δ = 0.1, and figure 14(c) shows the feature image when the leakage rate δ = 0.9. Figure 15(a) shows the input image of the CIFAR-10 dataset, figure 15(b) shows the feature image when the leakage rate δ = 0.1, and figure 15(c) shows the feature image when the leakage rate δ = 0.9.

[0063] Figures 13 to 15(b) show feature images when the leakage rate δ = 0.1, representing low spatial frequency components such as contours and general shapes on the input image. Figures 13 to 15(c) show feature images when the leakage rate δ = 0.9, representing high spatial frequency components such as fine edges on the input image. In this way, multiple reservoir layers L R By setting different leakage rates δ for each of these, it is possible to capture various features of the input image.

[0064] (Modified version of the embodiment) The following describes modifications of the above-described embodiment. In the above-described embodiment, the reservoir computing MDL is the input layer L IN And multiple reservoir layers L, each with a different leakage rate δ set. R And, output layer L OUT Although it was explained as including this, it is not limited to this. For example, reservoir computing MDL is a leakage rate δ of a certain value δ X A single reservoir layer L is set to this configuration. R It may also include the following. In this case, first the task processing unit 116 determines the leakage rate δ X Reservoir layer L set to R Input layer L IN Time series data u(t) is input via this. In response, the reservoir layer L R The leakage rate δ X The spatial frequency characteristics corresponding to the leakage rate δ are extracted. The task processing unit 116 processes the leakage rate δ. X After the spatial frequency characteristics corresponding to are extracted, the leakage rate δ X Other leakage rates δ Y The settings are reset. Then, the task processing unit 116 determines the leakage rate δ Y Reservoir layer L set to R Input layer L IN Time series data u(t) is input via this. In response, the reservoir layer L R The leakage rate δ Y The spatial frequency characteristics corresponding to the function are extracted. For example, the leakage rate δ X Leakage rate δ Y If the latter is larger, the leakage rate δ YThe spatial frequency characteristics corresponding to this are the leakage rate δ X The spatial frequency will be higher than the spatial frequency characteristics corresponding to the spatial frequency.

[0065] Furthermore, although the above-described embodiment assumed that the information processing device 100 and the learning device 200 are independent and separate devices, this is not limited to this configuration. For example, the processing unit 110 of the information processing device 100 may include the learning unit 214 of the learning device 200.

[0066] Although embodiments for carrying out the present invention have been described above using examples, the present invention is not limited in any way to these embodiments, and various modifications and substitutions can be made without departing from the spirit of the present invention. [Explanation of Symbols]

[0067] 1...Information processing system, 100...Information processing device, 102...Communication unit, 104...Input unit, 106...Output unit, 110...Processing unit, 112...Acquisition unit, 114...Image processing unit, 116...Task processing unit, 118...Output control unit, 130...Storage unit, NW...Network, MDL...Reservoir computing, 200...Learning device, 202...Communication unit, 210...Processing unit, 212...Acquisition unit, 214...Learning unit, 216...Communication control unit, 230...Storage unit

Claims

1. The acquisition unit acquires the time-series data to be tested, A reservoir computing system trained on a training dataset in which the time series data to be trained is labeled with the correct objects to be recognized from the time series data to be trained, the reservoir computing system comprising: an input layer into which the time series data is input; at least one reservoir layer that extracts the features of the objects from the time series data input to the input layer; and an output layer that outputs the recognition result of the objects based on the features extracted by the reservoir layer, the reservoir computing system comprising: a processing unit that recognizes the objects from the time series data to be tested by inputting the time series data to be tested, The aforementioned processing unit, The degree to which the reservoir layer is influenced by the time-series data input via the input layer is controlled. When observation data at a first time point included in the time series data is input to the reservoir layer in a certain state via the input layer, the influence the reservoir layer receives from the observation data at the first time point, the first state which is the state of the reservoir layer when the observation data at the first time point is input, and a time constant for controlling the degree of the influence are used to calculate the second state which is the state of the reservoir layer when observation data at a second time point, which is later than the first time point, is input. The reservoir computing includes a plurality of reservoir layers, each having a different time constant. The aforementioned processing unit, By inputting the time-series data of the object under test via the input layer to the first reservoir layer, in which the time constant is set to a first value, the characteristics of the first spatial frequency are extracted from the time-series data of the object under test. By inputting the time-series data of the subject of test to a second reservoir layer, whose time constant is set to a second value different from the first value, via the input layer, features of the second spatial frequency, which is different from the first spatial frequency, are extracted from the time-series data of the subject of test. Information processing device.

2. The processing unit calculates the degree of influence based on the observed data at the first time, which is weighted between the time it is output from the input layer and input to the reservoir layer, and the first state, which is weighted by the recursive coupling load. The information processing apparatus according to claim 1.

3. The processing unit calculates the second state based on the sum of the first state, which is obtained by multiplying 1 by the difference obtained by subtracting the time constant, and the degree of influence, which is obtained by multiplying by the time constant. The information processing apparatus according to claim 1 or 2.

4. The degree of the aforementioned effect decreases as the time constant decreases and increases as the time constant increases. The information processing apparatus according to claim 3.

5. The aforementioned processing unit, A portion of an image is cut out in the shape of a strip, The strip-shaped regions cut out from the aforementioned partial region are used as the time-series data. The information processing apparatus according to any one of claims 1 to 4.

6. The acquisition unit acquires the time-series data to be tested, A reservoir computing system trained on a training dataset in which the time series data to be trained is labeled with the correct objects to be recognized from the time series data to be trained, the reservoir computing system comprising: an input layer into which the time series data is input; at least one reservoir layer that extracts the features of the objects from the time series data input to the input layer; and an output layer that outputs the recognition result of the objects based on the features extracted by the reservoir layer, the reservoir computing system comprising: a processing unit that recognizes the objects from the time series data to be tested by inputting the time series data to be tested, The aforementioned processing unit, The degree to which the reservoir layer is influenced by the time-series data input via the input layer is controlled. When observation data at a first time point included in the time series data is input to the reservoir layer in a certain state via the input layer, the influence the reservoir layer receives from the observation data at the first time point, the first state which is the state of the reservoir layer when the observation data at the first time point is input, and a time constant for controlling the degree of the influence are used to calculate the second state which is the state of the reservoir layer when observation data at a second time point, which is later than the first time point, is input. The reservoir computing includes one reservoir layer whose time constant is set to a first value, The aforementioned processing unit, By inputting the time-series data of the object under test to the reservoir layer via the input layer, the first spatial frequency features are extracted from the time-series data of the object under test. After the first spatial frequency feature is extracted, the time constant is reset to a second value different from the first value. By inputting the time series data of the object under test to the reservoir layer, whose time constant is set to the second value, via the input layer, the characteristics of the second spatial frequency, which is different from the first spatial frequency, are extracted from the time series data of the object under test. Information processing device.

7. A method of information processing performed by a computer, Obtain the time-series data to be tested. A reservoir computing system trained on a training dataset in which the time series data to be trained is labeled with the correct objects to be recognized from the time series data to be trained, the reservoir computing system comprising an input layer into which the time series data is input, at least one reservoir layer that extracts the features of the objects from the time series data input to the input layer, and an output layer that outputs the recognition result of the objects based on the features extracted by the reservoir layer, is used to recognize the objects from the time series data to be tested by inputting the time series data to be tested. Controlling the degree to which the reservoir layer is influenced by the time-series data input via the input layer. The process includes calculating a second state, which is the state of the reservoir layer when observation data for a second time point, which is later than the first time point, is input, based on the influence the reservoir layer receives from the observation data for a first time point, which is included in the time series data, when the reservoir layer is input via the input layer, the first state, which is the state of the reservoir layer when the observation data for the first time point is input, and a time constant for controlling the degree of the influence. The reservoir computing includes a plurality of reservoir layers, each having a different time constant. By inputting the time-series data of the object under test via the input layer into the first reservoir layer, whose time constant is set to a first value, the characteristics of a first spatial frequency are extracted from the time-series data of the object under test. The method further includes inputting the time-series data of the subject of test to a second reservoir layer, whose time constant is set to a second value different from the first value, via the input layer, thereby extracting features of a second spatial frequency different from the first spatial frequency from the time-series data of the subject of test. Information processing methods.

8. A program to be executed by a computer, Obtain the time-series data to be tested. A reservoir computing system trained on a training dataset in which the time series data to be trained is labeled with the correct objects to be recognized from the time series data to be trained, the reservoir computing system comprising an input layer into which the time series data is input, at least one reservoir layer that extracts the features of the objects from the time series data input to the input layer, and an output layer that outputs the recognition result of the objects based on the features extracted by the reservoir layer, is used to recognize the objects from the time series data to be tested by inputting the time series data to be tested. Controlling the degree to which the reservoir layer is influenced by the time-series data input via the input layer. The process includes calculating a second state, which is the state of the reservoir layer when observation data for a second time point, which is later than the first time point, is input, based on the influence the reservoir layer receives from the observation data for a first time point, which is included in the time series data, when the reservoir layer is input via the input layer, the first state, which is the state of the reservoir layer when the observation data for the first time point is input, and a time constant for controlling the degree of the influence. The reservoir computing includes one reservoir layer whose time constant is set to a first value, By inputting the time-series data of the object under test to the reservoir layer via the input layer, the first spatial frequency features are extracted from the time-series data of the object under test. After the first spatial frequency characteristics have been extracted, the time constant is reset to a second value different from the first value. The method further includes inputting the time-series data of the object under test to the reservoir layer, whose time constant is set to the second value, via the input layer, thereby extracting features of a second spatial frequency different from the first spatial frequency from the time-series data of the object under test. program.

Images