Information processing device

Abstract

This information processing device comprises an input layer, a reservoir layer, an output layer, an evaluation circuit, and an adjustment circuit. The reservoir layer is connected to the input layer, and generates a feature space including information on first signals input from the input layer. The output layer is connected to the reservoir layer, and applies connection weights to second signals output from the reservoir layer. The evaluation circuit calculates the distribution of connection weights in the output layer, and evaluates whether or not the distribution of connection weights is a defined distribution. The adjustment circuit changes an adjustment parameter for adjusting the first signals when the distribution of connection weights is not the defined distribution.

Description

[Technical field]

[0001] The present invention relates to an information processing device. [Background technology]

[0002] A neuromorphic device is an element that mimics the human brain by using neural networks. Neuromorphic devices artificially mimic the relationship between neurons and synapses in the human brain.

[0003] A neuromorphic device, for example, has nodes (neurons in the brain) arranged in a hierarchical structure and transmission means (synapses in the brain) that connect them. Neuromorphic devices increase the rate of correct answers to questions by learning the transmission means (synapses). Learning is the process of finding knowledge that can be used in the future from information, and neuromorphic devices weight the input data.

[0004] Recurrent neural networks are known as one type of neural network. Recurrent neural networks contain recurrent connections and can handle time-series data. Time-series data is data whose values ​​change over time, such as stock prices. Recurrent neural networks can also have nonlinear activation parts inside. Mathematically, the processing in the activation part can be considered as a projection into a nonlinear space. By projecting data into a nonlinear space, recurrent neural networks can extract the characteristics of complex signal changes that time-series signals have. By returning the processing results in neurons in the later layer to neurons in the earlier layer, recurrent neural networks can achieve recursive processing. By performing recursive processing, recurrent neural networks can acquire the rules and governing factors behind time-series data.

[0005] Reservoir computing is a type of recurrent neural network that includes recurrent connections and nonlinear activation functions (e.g., Non-Patent Document 1). Reservoir computing is a neural network that has been developed as an implementation method for liquid state machines.

[0006] Reservoir computing has a reservoir layer. The "layer" here is a conceptual layer, and does not necessarily have to be formed as a physical structure. The reservoir layer has a graph structure that includes a large number of nonlinear nodes and recursive connections between the nodes. In reservoir computing, the reservoir layer mimics the neuron connections of the human brain and expresses states as transitions of interference states.

[0007] One of the features of reservoir computing is that the reservoir layer is not the learning target. Since reservoir computing requires few computer resources for learning, it is attracting attention as a system for handling time series signals in the IoT (Internet of Things) where hardware resources are limited, and at the edge. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] U. Ozertem, D. Erdogmus, and I. Santamaria, Detection of nonlinearly distorted signals using mutual information, European Signal Processing Conference. IEEE, 2005. Summary of the Invention [Problem to be solved by the invention]

[0009] The accuracy of fitting the output values ​​from reservoir computing to the training data varies depending on the parameter settings, and a systematic method for designing the parameters of reservoir computing has not yet been established.

[0010] The present invention has been made in consideration of the above circumstances, and has an object to provide an information processing device with a high rate of correct answers. [Means for solving the problem]

[0011] (1) An information processing device according to a first aspect includes an input layer, a reservoir layer, an output layer, an evaluation circuit, and an adjustment circuit. The reservoir layer is connected to the input layer and generates a feature space including information of a first signal input from the input layer. The output layer is connected to the reservoir layer and applies a connection weight to a second signal output from the reservoir layer. The evaluation circuit obtains a distribution of connection weights in the output layer and evaluates whether the distribution of connection weights is a specified distribution. The adjustment circuit changes an adjustment parameter for adjusting the first signal when the distribution of connection weights is not the specified distribution.

[0012] (2) In the information processing device according to the above aspect, the specified distribution may be a normal distribution.

[0013] (3) In the information processing device according to the above aspect, when updating the connection weights to be applied to the second signal, the evaluation circuit may evaluate whether or not a distribution of the connection weights is a specified distribution if an amount of change in each of the connection weights before and after the update is equal to or less than a threshold.

[0014] (4) In the information processing device according to the above aspect, when the number of times the adjustment parameters have been changed reaches a prescribed number, the adjustment circuit may select an optimal adjustment parameter from among the adjustment parameters that have been changed so far.

[0015] (5) In the information processing device according to the above aspect, the adjustment parameter may be a connection weight by which an input signal applied to the input layer is multiplied.

[0016] (6) In the information processing device according to the above aspect, the adjustment parameter may be a filter coefficient of a filter that selectively passes a frequency component that constitutes an input signal applied to the input layer.

[0017] (7) In the information processing device according to the above aspect, the distribution of the adjustment parameters may be a normal distribution.

[0018] (8) In the information processing device according to the above aspect, the distribution of the adjustment parameters may be a uniform distribution. Effect of the Invention

[0019] The information processing device according to the above aspect has a high rate of correct answers. [Brief description of the drawings]

[0020] [Figure 1] 1 is a conceptual diagram of an information processing device according to a first embodiment. [Diagram 2] 1 is an example of a distribution of connection weights. [Diagram 3] FIG. 2 is a flow diagram of the information processing device according to the first embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] The present embodiment will be described in detail below with reference to the drawings as appropriate. The drawings used in the following description may show characteristic parts in an enlarged scale for the sake of convenience in order to make the characteristics easier to understand, and the dimensional ratios of each component may differ from the actual ones. The materials, dimensions, etc. exemplified in the following description are merely examples, and the present invention is not limited thereto. They may be modified as appropriate within the scope of the effects of the present invention.

[0022] 1 is a conceptual diagram of an information processing device 100 according to a first embodiment. The information processing device 100 is a reservoir device that realizes reservoir computing. The information processing device 100 may be a device that realizes reservoir computing by software, or a device in which the concept of reservoir computing is implemented in a physical device (hardware).

[0023] The information processing device 100 includes, for example, an input layer 10, a reservoir layer 20, an output layer 30, an evaluation circuit 40, an adjustment circuit 50, a comparison circuit 60, and an instruction circuit 70. The information processing device 100 can perform both learning to increase the rate of correct answers to tasks and calculation (inference) to output answers to tasks based on the learning results. The evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60, and the instruction circuit 70 are used in the learning stage and are not necessary in the calculation (inference) stage.

[0024] The input layer 10 receives input signals Sin1 to Sin n are input. Input signals Sin1 to Sin n The number of n is not important, and n is a natural number.

[0025] The input layer 10 is, for example, a single-layer or multi-layer perceptron. n are the first signals S11 to S1 n The first signals S11 to S1 n The first signals S11 to S1 n The number of input signals Sin1 to Sin n The number of input signals Sin1 to Sin n From the first signal S11 to S1 n The conversion of the signal to is adjusted by the adjustment parameters.

[0026] The input layer 10 includes, for example, a filter 11 and a bias application unit 12. The filter 11 receives, for example, input signals Sin1 to Sin n is divided into signal and noise. The ratio of signal to noise is, for example, nIt is determined by the filter coefficients f1 to f n is one of the adjustment parameters. The filter 11, for example, n A reference number is provided for each of the above.

[0027] Each input signal Sin1~Sin n Filter coefficients f1 to f n Each of the filter coefficients f1 to f may be the same or different. n may show a distribution when plotted, for example, with the horizontal axis representing the value of the filter coefficient and the vertical axis representing the number of filter coefficients of a particular value. n The distribution of is, for example, a normal distribution, a uniform distribution, a Laplace distribution, or the like.

[0028] The bias application unit 12 applies input signals Sin1 to Sin n The connection weights wa1 to wa n Input signals Sin1 to Sin n Each of these has a connection weight wa1 to wa n The bias application unit 12 applies the input signals Sin1 to Sin n Each of the weights wa1 to wa n A multiplication operation is performed to multiply the connection weights wa1 to wa n is one of the adjustment parameters.

[0029] Connection weights wa1~wa n Each of the connection weights wa1 to wa2 may be the same or different. n For example, the distribution may be shown when plotted with the horizontal axis representing the value of the connection weight and the vertical axis representing the number of connection weights of a particular value. n The distribution of the connection weights wa1 to wa2 is, for example, a normal distribution, a uniform distribution, a Laplace distribution, etc. n FIG. 2 shows an example of the distribution of the connection weights wa1 to wa n is an example of a normal distribution.

[0030] In one example of implementing the concept of reservoir computing in a physical device (hardware), the coupling weights wa1 to wa n is, for example, the amplitude, frequency, phase, etc. of the propagating wave. The wave may be generated by vibration, for example, an electromagnetic field, a magnetic field, a spin wave, or an elastic wave. In another example where the concept of reservoir computing is implemented in a physical device (hardware), the coupling weights wa1 to wa n is the resistance value of the variable resistor. The variable resistor is, for example, a resistance variable element called a memristor. For example, a domain wall motion type magnetoresistance effect element in which the resistance value changes depending on the position of the domain wall is an example of a memristor.

[0031] Filter coefficients f1 to f n and connection weights wa1~wa n fluctuates during the learning phase. On the other hand, the filter coefficients f1 to f n and connection weights wa1~wa n In the calculation (inference) stage, is fixed based on the learning results tailored to the task in the learning stage.

[0032] The reservoir layer 20 includes a plurality of nodes 21. The number of the nodes 21 is not particularly limited. The greater the number of the nodes 21, the greater the expressiveness of the reservoir layer 20. For example, the number of the nodes 21 is assumed to be i, where i is any natural number.

[0033] When the concept of reservoir computing is implemented in a physical device (hardware), each of the nodes 21 is substituted with, for example, a physical element. The physical element is, for example, a device that can convert an input signal into vibration, an electromagnetic field, a magnetic field, a spin wave, or the like. The node 21 is, for example, a MEMS microphone. The MEMS microphone can convert between the vibration of a diaphragm and an electric signal. The node 21 may be, for example, a spin torque oscillator (STO). The spin torque oscillator can convert between an electric signal and a high-frequency signal. The node 21 may also be a Schmitt trigger circuit having a hysteresis circuit whose output state changes with hysteresis in response to a change in the potential of the input signal, or an operational amplifier having other nonlinear response characteristics. The node 21 may also be a memristor.

[0034] A signal from each node 21 interacts with signals from surrounding nodes 21. Between each node 21, for example, a connection weight wb m is set. The connection weight wb m The number of connections between the nodes 21 is equal to the number of combinations of connections between the nodes 21. m is, for example, an arbitrary natural number. m In principle, each of the weights wb between the nodes 21 is fixed and does not change due to learning. m Each of the weights wb between the nodes 21 is arbitrary and may be the same or different. m A part of may vary due to learning.

[0035] In one example of implementing the concept of reservoir computing in a physical device (hardware), the coupling weights wb m are, for example, the amplitude, frequency, phase, etc. of the propagating wave. Also, the coupling weight wb m may be the resistance value of a variable resistor.

[0036] The reservoir layer 20 includes first signals S11 to S1 n The first signals S11 to S1 nThe first signals S11 to S1 n The interaction between the first signals S11 and S11 means that a signal propagating to a node 21 affects a signal propagating to another node 21. For example, the first signals S11 to S1 n When propagating between nodes 21, m The reservoir layer 20 changes in response to the input first signals S11 to S1 n into a multidimensional nonlinear space.

[0037] 1st signal S11~S1 n The first signals S11 to S11 propagate between the multiple nodes 21, so that the reservoir layer 20 receives the first signals S11 to S11 n In the reservoir layer 20, the input first signals S11 to S1 n are the second signals S21 to S2 i The second signals S21 to S2 are converted into the following. For example, i is an arbitrary natural number, and may be the same as or different from n. i are the first signals S11 to S1 n For example, the first signal S11 to S1 n changes nonlinearly in the reservoir layer 20, and the second signals S21 to S2 i The first signals S11 to S1 n As the molecules interact within reservoir layer 20, the state of the reservoir layer 20 system changes over time.

[0038] The output layer 30 receives signals from the reservoir layer 20. The second signals S21 to S2 i is input to the output layer 30.

[0039] The output layer 30 includes, for example, a bias application unit 31, a summation circuit 32, and an activation function circuit 33.

[0040] The bias application unit 31 outputs the second signals S21 to S2 i Weights wc1 to wci The second signals S21 to S2 i Each of the connections has weights wc1 to wc i The bias application unit 31 applies the second signals S21 to S2 i Each of the weights wc1 to wc i A multiplication operation is performed by multiplying the .

[0041] Connection weight wc1~wc i Each of the connection weights wc1 to wc may be the same or different. i For example, when plotted with the horizontal axis representing the value of the connection weight and the vertical axis representing the number of connection weights with a specific value, the distribution of connection weights wc1 to wc i The distribution of the connection weights wc1 to wc is determined based on the task. i The distribution of is, for example, a normal distribution, a uniform distribution, a Laplace distribution, or the like.

[0042] Connection weight wc1~wc i fluctuates during the learning stage. On the other hand, the connection weights wc1 to wc i In the calculation (inference) stage, is fixed based on the learning results tailored to the task in the learning stage.

[0043] In one example of implementing the concept of reservoir computing in a physical device (hardware), the coupling weights wc1 to wc i is, for example, the amplitude, frequency, phase, etc. of the propagating wave. The wave may be generated by vibration, for example, an electromagnetic field, a magnetic field, a spin wave, or an elastic wave. In another example where the concept of reservoir computing is implemented in a physical device (hardware), the coupling weights wc1 to wc i is the resistance value of the variable resistor.

[0044] The summation circuit 32 outputs the second signals S21 to S2 i Each of the weights wc1 to wc i The summation circuit 32 multiplies the second signals S21 to S2 i Each of the weights wc1 to wc iAll of the results of multiplication may be summed up, or only some of them may be summed up. The result output from the summing circuit 32 may be one as shown in FIG. 1, or multiple results may be output. For example, when signals are propagated using waves, the waves are joined in the summing circuit 32. For example, when signals are propagated using currents, the wiring is grouped together in the summing circuit 32 and the currents are joined.

[0045] The activation function circuit 33 performs a calculation by substituting the product-sum calculation result into the activation function f(x). The activation function circuit 33 converts the product-sum calculation result into a non-linear function. The activation function circuit 33 is not necessary.

[0046] In the learning stage, the output signal S from the output layer 30 out is sent to the comparison circuit 60. In the calculation (inference) stage, the output signal S out is output to the outside as the answer. out For example, when the information processing device 100 is adapted to a multi-class classification problem, which is a general application of machine learning, the output layer 30 generates a plurality of output signals S corresponding to each class. out Output.

[0047] The comparison circuit 60 outputs an output signal S out The comparison circuit 60 compares the output signal S out The mutual information between the training data t and the training data t is compared. The mutual information is a quantity that represents a measure of the interdependence between two random variables. The comparison circuit 60 transmits the comparison result to the instruction circuit 70.

[0048] The instruction circuit 70 sends an instruction to the bias application unit 31 based on the comparison circuit 60. Based on the instruction from the instruction circuit 70, the bias application unit 31 adjusts the coupling weights wc1 to wc i Update the connection weights wc1~wc i When changes occur, the output signal S out The indication circuit 70 outputs an output signal S out The information is fed back to the bias application unit 31 so that the mutual information between the training data t and the weights wc1 to wci changes based on the fed back data.

[0049] The evaluation circuit 40 evaluates the connection weights wc1 to wc i Calculate the distribution of the connection weights wc1 to wc i The evaluation circuit 40 evaluates whether the distribution of the connection weights wc1 to wc i The evaluation circuit 40 may perform the above evaluation every time the weights wc1 to wc are updated, or may perform the above evaluation based on a predetermined rule. i The difference may be calculated.

[0050] Connection weight wc1~wc i The distribution of the second signals S21 to S2 i A switch is provided between the first and second signals S21 to S2 i The second signals S21 to S2 may be switchable between the first and second signals S21 to S22 and the reference signal. i Instead of the weights wc1 to wc i The reference signal is output from, for example, the evaluation circuit 40. Also, for example, the coupling weights wc1 to wc i If is the resistance value of the memristor, then applying a reference current to each memristor will determine the resistance value of each memristor (connection weights wc1 to wc i ) may be obtained.

[0051] The adjustment circuit 50 adjusts the first signals S11 to S1 n The adjustment circuit 50 adjusts, for example, the coupling weights wc1 to wc i If the distribution of f1 to f2 is not a specified distribution, the adjustment parameters of the input layer 10 are changed. The adjustment parameters are, for example, the filter coefficients f1 to f2. n , connection weights wa1~wa n When the adjustment parameters are changed, the input signals Sin1 to Sin n From the first signal S11 to S1 nThe rule for converting the signal to the first signal S11-S1 n The adjustment circuit 50 may also count the number of times the adjustment parameters are changed.

[0052] The evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60 and the indication circuit 70 operate in the learning phase and do not operate in the calculation (inference) phase.

[0053] The evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60, and the instruction circuit 70 each have a processor, such as a CPU, and a memory. The evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60, and the instruction circuit 70 each operate by the processor executing a program. The processor instructs each circuit to operate, and the memory records the programs and past results.

[0054] All or part of the operation of these circuits may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array). The above programs may be recorded on a computer-readable recording medium. Examples of computer-readable recording media include portable media such as flexible disks, magneto-optical disks, ROMs, CD-ROMs, and semiconductor storage devices (e.g., SSDs: Solid State Drives), and storage devices such as hard disks and semiconductor storage devices built into computer systems. The above programs may be transmitted via electric communication lines.

[0055] 3 is a flow diagram of the information processing device 100 according to the first embodiment. The information processing device 100 operates in the learning stage according to the flow diagram shown in FIG.

[0056] The information processing device 100 receives input signals Sin1 to Sin n When input, the output signal S outThe information processing device 100 outputs the connection weights wc1 to wc i , while changing the adjustment parameters of the input layer 10, the training data t and the output signal S out Compare the training data t and the output signal S out Learning ends when the mutual information with the trained model becomes sufficiently large (the correct answer rate for the task becomes sufficiently high).

[0057] First, in the learning process, the first step S1 is performed. In the first step S1, when the comparison result in the comparison circuit 60 is insufficient (when the correct answer rate for the task is insufficient), the connection weights wc1 to wc i Update the connection weights wc1 to wc i is changed based on an instruction from the instruction circuit 70 to the bias application unit 31. For example, the first step S1 is performed multiple times, and the connection weights wc1 to wc i will be updated accordingly.

[0058] Next, the second step S2 is performed. In the second step S2, the connection weights wc1 to wc i Measure the change D of the connection weights wc1 to wc i The amount of change D is measured, for example, by an evaluation circuit 40.

[0059] In the second step S2, the connection weights wc1 to wc i It is judged whether the change amount D of each of the connection weights wc1 to wc i If the change amount D of each of these is equal to or less than the threshold value Δ, the output signal S out The connection weights wc1 to wc before and after the update have converged. i If the change amount D of the connection weights wc1 to wc is equal to or smaller than the threshold value Δ, the process proceeds to the third step S3. i Each threshold Δ may be set separately. i is greater than the threshold Δ, and the output signal S out If diverges, return to the first step S1 and recalculate the connection weights wc1 to wc i Update.

[0060] In the third step S3, the connection weights wc1 to wc i Find the distribution of the connection weights wc1 to wc i The distribution of the connection weights wc1 to wc i The distribution of each of the connection weights wc1 to wcc is obtained by plotting the value of the connection weight on the horizontal axis and the number of connection weights with a specific value on the vertical axis. i The value of can be obtained, for example, by inputting a reference signal as described above.

[0061] And the connection weights wc1 to wc i It is evaluated whether the distribution of corresponds to a predetermined distribution. The predetermined distribution is, for example, a normal distribution, a uniform distribution, a Laplace distribution, etc. The predetermined distribution differs depending on the task given to the information processing device 100. The predetermined distribution is determined when the task is determined, and is stored in the evaluation circuit 40.

[0062] For example, if the default distribution is normal, it is determined based on, for example, the Kolmogorov-Smirnov test or the Shapiro-Wilk test. i If the P value of the distribution is less than 0.05, the connection weights wc1 to wc i It can be determined that the distribution of is not normal.

[0063] Connection weight wc1~wc i If the distribution of wc1 to wc2 of the output layer 30 corresponds to the predetermined distribution, the process proceeds to the fourth step S4, where adjustment parameters for the input layer 10 are determined. If the information processing device 100 shows a sufficient accuracy rate after the adjustment parameters for the input layer 10 are determined, learning is terminated. If the information processing device 100 does not show a sufficient accuracy rate even after the adjustment parameters for the input layer 10 are determined, the process returns to the first step S1 again, where the connection weights wc1 to wc2 of the output layer 30 are adjusted. i Update.

[0064] In contrast, the connection weights wc1 to wc iIf the distribution does not correspond to the predetermined distribution, the process proceeds to a fifth step S5. In the fifth step S5, it is determined whether the number of times the adjustment parameter has been changed is equal to or greater than a specified number. The adjustment circuit 50 counts the number of times the adjustment parameter has been changed. The specified number is appropriately determined depending on the task of the information processing device 100, the accuracy of the correct answer, the calculation load, etc. The specified number is stored in the adjustment circuit 50.

[0065] If the number of times the adjustment parameters have been changed is less than the specified number of times, the process proceeds to a sixth step S6, where the adjustment parameters of the input layer 10 are changed. The adjustment parameters are changed by the adjustment circuit 50. The adjustment circuit 50 changes, for example, the filter coefficients f1 to f n The adjustment circuit 50 changes the connection weights wa1 to wa n The adjustment circuit 50 may change the filter coefficients f1 to f n and the connection weights wa1~wa n You may change both.

[0066] The adjustment circuit 50, for example, sets the distribution of the adjustment parameters of the input layer 10 to a normal distribution. When the adjustment parameters are changed multiple times by the adjustment circuit 50, the mean or variance is changed based on the normal distribution.

[0067] For example, when the adjustment circuit 50 adjusts the adjustment parameters for the first time, the distribution of the adjustment parameters is set to a normal distribution with a mean of 0 and a variance of 1. When the adjustment circuit 50 adjusts the adjustment parameters for the second time or later, the distribution of the adjustment parameters is changed so that the variance becomes larger. For example, when the adjustment circuit 50 adjusts the adjustment parameters for the nth time, the variance value is increased by 10% compared to the variance value for the (n-1)th time.

[0068] Furthermore, the adjustment circuit 50 may, for example, set the distribution of the adjustment parameters of the input layer 10 to a uniform distribution having a certain random number width. When the adjustment parameters are changed multiple times by the adjustment circuit 50, the random number width is changed.

[0069] For example, when the adjustment circuit 50 adjusts the adjustment parameters for the first time, the distribution of the adjustment parameters is set to a uniform distribution with a random number width of ±0.1. When the adjustment circuit 50 adjusts the adjustment parameters for the second time or later, the distribution of the adjustment parameters is changed so that the random number width becomes larger. For example, when the adjustment circuit 50 adjusts the adjustment parameters for the nth time, the random number width is set to a uniform distribution of ±0.1×n.

[0070] On the other hand, if the number of times the adjustment parameters have been changed is equal to or greater than the specified number, no further changes are made to the adjustment parameters of the input layer 10. If the number of times the adjustment parameters have been changed is equal to or greater than the specified number, the process proceeds to the seventh step S7. In the seventh step S7, an optimal adjustment parameter is selected from among the adjustment parameters that have been changed so far as the adjustment parameter of the input layer 10. The adjustment parameter with the highest accuracy rate in the change history so far is set as the optimal adjustment parameter.

[0071] In the above procedure, the information processing device 100 calculates the connection weights wc1 to wc i In the calculation (inference) stage, the information processing device 100 determines the adjustment parameters of the input layer 10 and ends the learning. i , calculations are performed using the adjustment parameters of the input layer 10.

[0072] Here, an example of the processing flow of the information processing device 100 is shown, but the processing flow is not limited to this example. i Each time the connection weights wc1 to wc i In the case where the distribution of is evaluated, the second step S2 does not have to be performed. In addition, in the case where the upper limit of the number of times the adjustment parameter is changed is not set, the fifth step S5 and the seventh step S7 do not have to be performed.

[0073] The information processing device 100 according to this embodiment uses the connection weights wc1 to wc i Based on the distribution of , the adjustment parameters of the input layer 10 are adjusted. By adjusting the adjustment parameters of the input layer 10 as well, the information processing device 100 can further increase the rate of correct answers to the task. [Explanation of symbols]

[0074] 10...input layer, 11...filter, 12, 31...bias application unit, 20...reservoir layer, 21...node, 30...output layer, 32...sum operation circuit, 33...activation function circuit, 40...evaluation circuit, 50...adjustment circuit, 60...comparison circuit, 70...indication circuit, 100...information processing device, f1 to f n …Filter coefficients, Sin1 to Sin n …Input signal, S11~S1 n …1st signal, S21~S2 i …Second signal, wa1~wa n ,wc1~wc i …Connection weights

Claims

1. An input layer; a reservoir layer connected to the input layer and configured to generate a feature space including information of a first signal input from the input layer; an output layer connected to the reservoir layer and configured to apply a coupling weight to a second signal output from the reservoir layer; an evaluation circuit for determining a distribution of connection weights in the output layer and evaluating whether the distribution of connection weights is a specified distribution; an adjustment circuit that changes an adjustment parameter that adjusts the first signal when the distribution of the connection weights is not the specified distribution.

2. The information processing device according to claim 1 , wherein the specified distribution is a normal distribution.

3. 3. The information processing device according to claim 1, wherein the evaluation circuit, when updating the connection weight to be applied to the second signal, evaluates whether a distribution of the connection weights is a specified distribution if a change in the connection weight before and after the update is equal to or less than a threshold.

4. 4. The information processing device according to claim 1, wherein the adjustment circuit selects an optimal adjustment parameter from among the adjustment parameters that have been changed so far when the number of times the adjustment parameters have been changed reaches a prescribed number.

5. The information processing device according to claim 1 , wherein the adjustment parameters are connection weights by which an input signal applied to the input layer is multiplied.

6. 6. The information processing device according to claim 1, wherein the adjustment parameters are filter coefficients of a filter that selectively passes frequency components that constitute an input signal applied to the input layer.

7. The information processing device according to claim 1 , wherein the distribution of the adjustment parameters is a normal distribution.

8. The information processing device according to claim 1 , wherein the distribution of the adjustment parameters is a uniform distribution.

Images