Patent Application
20250232166
The present invention relates to an information processing device.
A neuromorphic device is a device that imitates the human brain using a neural network. A neuromorphic device artificially imitates a relationship between neurons and synapses in the human brain.
For example, a neuromorphic device includes nodes that are hierarchically arranged (neurons in the brain) and transmission means that connect the nodes (synapses in the brain). A neuromorphic device enhances a rate of correct answers to questions by training the transmission means (synapses). Learning is finding knowledge which is likely to be used in the future from information, and a neuromorphic device weights data that it receives.
A recurrent neural network is known as a neural network. A recurrent neural network includes recurrent connections therein and can handle nonlinear time-series data. Nonlinear time-series data is data of which a value changes with the elapse of time, and an example thereof is stock prices. A recurrent neural network can also include a nonlinear activation unit therein. A process in the activation unit can be considered as projection to a nonlinear space. By projecting data to a nonlinear space, a recurrent neural network can extract features of complex signal change of a time-series signal. A recurrent neural network can realize a recurrent process by feeding process results in neurons in a subsequent stage back to neurons in a preceding stage. A recurrent neural network can acquire rules or dominant factors in the background of time-series data by performing such a recursive process.
Reservoir computing is a kind of recurrent neural network including a recurrent connection and a nonlinear activation function (for example, Non-Patent Document 1). Reservoir computing is a neural network which has been developed as a method applied to a liquid state machine.
Reservoir computing includes a reservoir layer. A “layer” mentioned herein is a conceptual layer, and a layer which is a physical structure does not have to be formed. A reservoir layer forms a graph structure including a plurality of nonlinear nodes and recurrent connections between the nodes. In reservoir computing, a reservoir layer imitates connections between neurons in the human brain, and a state is expressed as change of an interference state.
One feature of reservoir computing is that a reservoir layer is not a learning target. Since computer resources required for learning are small, reservoir computing attracts attention as a system for handling time-series signals in the Internet of Things (IoT) or an edge.
U. Ozertem, D. Erdogmus, and I. Santamaria, Detection of nonlinearly distorted signals using mutual information, European Signal Processing Conference. IEEE, 2005
Accuracy of fitting an output value of reservoir computing to training data varies depending on settings of parameters. A method of systematically designing parameters of reservoir computing is not established yet.
The present invention was made in consideration of the aforementioned circumstances, and an objective thereof is to provide an information processing device with a high rate of correct answers.
(1) According to a first aspect, there is provided an information processing device including 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 configured to generate a feature space including information of a first signal input from the input layer. The output layer is connected to the reservoir layer and configured to apply a connection weight to a second signal which is output from the reservoir layer. The evaluation circuit is configured to calculate a distribution of connection weights in the output layer and to evaluate whether the distribution of connection weights is a prescribed distribution. The adjustment circuit is configured to change adjustment parameters for adjusting the first signal when the distribution of connection weights is not the prescribed distribution.
(2) In the information processing device according to the aspect, the prescribed distribution may be a normal distribution.
(3) In the information processing device according to the aspect, the evaluation circuit may evaluate whether the distribution of connection weights is a prescribed distribution when a change in connection weight between before updating and after updating is less than or equal to a prescribed value at the time of updating the connection weights applied to the second signal.
(4) In the information processing device according to the aspect, the adjustment circuit may select an optimal adjustment parameter out of the changed adjustment parameters when the number of times of change of the adjustment parameters reaches a prescribed number.
(5) In the information processing device according to the aspect, the adjustment parameters may be connection weights which are multiplied by an input signal applied to the input layer.
(6) In the information processing device according to the aspect, the adjustment parameters may be filter coefficients of filters for selectively passing frequency components constituting an input signal applied to the input layer.
(7) In the information processing device according to the aspect, a distribution of the adjustment parameters may be a normal distribution.
(8) In the information processing device according to the aspect, a distribution of the adjustment parameters may be a uniform distribution.
The information processing device according to the aspect has a high rate of correct answers.
Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. In the drawings referred to in the following description, featured constituents may be conveniently enlarged for the purpose of easy understanding of features, and dimensions, proportions, and the like of the constituents may be different from actual ones. Materials, dimensions, and the like exemplified in the following description are only examples, and the present invention is not limited thereto and can be appropriately modified within a range in which advantages of the present invention are achieved.
The information processing device 100 includes 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 learning for enhancing a rate of correct answers to a task and an arithmetic operation (inference) of outputting a response to a task on the basis of a result of learning. The evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60, and the instruction circuit 70 are used in a learning stage and are not necessary for an arithmetic operation (inference) stage.
Input signals Sin1 to Sinn are input to the input layer 10. The number of input signals Sin1 to Sinn is not particularly limited, and n is a natural number.
The input layer 10 is, for example, a single-layered or multi-layered perceptron. The input signals S1n to Sinn input to the input layer 10 are converted to first signals S11 to S1n. The number of first signals S11 to S1n is not particularly limited, and n is a natural number. The number of first signals S11 to S1n may be equal to or different from the number of input signals Sin1 to Sinn. Signal conversion from the input signals Sin1 to Sinn to the first signals S11 to S1n is adjusted using adjustment parameters.
The input layer 10 includes, for example, a filter 11 and a bias applying unit 12. For example, the filter 11 separates the input signals Sin1 to Sinn into a signal and noise. A ratio between the signal and the noise is determined, for example, by filter coefficients f1 to fn. The filter coefficients f1 to fn are one type of adjustment parameters. The filter 11 is provided, for example, for each of the input signals Sin1 to Sinn.
The filter coefficients f1 to fn for the input signals Sin1 to Sinn may be the same or different from each other. The filter coefficients f1 to fn may represent, for example, a distribution in which values of filter coefficients are plotted on the horizontal axis and the number of filter coefficients of a specific value is plotted on the vertical axis. The distribution of the filter coefficients f1 to fn is, for example, a normal distribution, a uniform distribution, or a Laplacian distribution.
The bias applying unit 12 applies connection weights way to wan to the input signals Sin1 to Sinn. The connection weights way to wan are applied to the input signals S1n to Sinn, respectively. The bias applying unit 12 performs a product operation of multiplying the input signals Sin1 to Sinn by the connection weights way to wan. The connection weights wa1 to wan are a kind of adjustment parameters.
The connection weights wa1 to wan may be the same or different from each other. The connection weights wa1 to wan may represent, for example, a distribution in which values of the connection weights are plotted on the horizontal axis and the number of connection weights of a specific value is plotted on the vertical axis. The distribution of the connection weights way to wan is, for example, a normal distribution, a uniform distribution, or a Laplacian distribution.
In an example in which the concept of reservoir computing is implemented in a physical device (hardware), the connection weights way to wan are, for example, an amplitude, a frequency, or a phase of propagating waves. Waves have only to be generated due to vibration and examples thereof include an electromagnetic field, a magnetic field, a spin wave, and an elastic wave. In another example in which the concept of reservoir computing is implemented in a physical device (hardware), the connection weights wa1 to wan are, for example, resistance values of variable resistors. A variable resistor is, for example, a variable-resistance element called a memristor. For example, a magnetoresistance effect element of a magnetic domain wall movement type in which a resistance value changed at a position of a magnetic domain wall is an example of a memristor.
The filter coefficients f1 to fn and the connection weights wa1 to wan vary in a learning stage. On the other hand, the filter coefficients f1 to fn and the connection weights wa1 to wan are fixed on the basis of a learning result based on tasks in the learning stage in an arithmetic operation (inference) stage.
The reservoir layer 20 includes a plurality of nodes 21. The number of nodes 21 is not particularly limited. As the number of nodes 21 increases, the expressive power of the reservoir layer 20 increases. For example, the number of nodes 21 is set to i. Here, i is an arbitrary natural number.
When the concept of reservoir computing is implemented in a physical device (hardware), each node 21 is replaced with, for example, a physical device. A physical device is, for example, a device that can convert an input signal to vibration, an electromagnetic field, a magnetic field, or a spin wave. Each node 21 is, for example, an MEMS microphone. An MEMS microphone can convert vibration of a vibration membrane to an electrical signal. Each node 21 may be, for example, a spin-torque oscillator (STO). A spin-torque oscillator can convert an electrical signal to a high-frequency signal. Each node 21 may be a Schmitt trigger circuit including a hysteresis circuit of which an output sate changes with hysteresis with change in potential of an input signal, an operational amplifier having different nonlinear response characteristics, or the like. Each node 21 may be a memristor.
A signal from each node 21 interacts with signals from neighboring nodes 21. For example, connection weights wbm are set between the nodes 21. The number of connection weights wbm corresponds to the number of combinations of connection between the nodes 21. Here, m is, for example, an arbitrary natural number. The connection weights wbm between the nodes 21 are fixed in principle and do not vary through learning. The connection weights wbm between the nodes 21 are arbitrary and may be the same or different from each other. Some of the connection weights wbm between a plurality of nodes 21 may vary through learning.
In an example in which the concept of reservoir computing is implemented in a physical device (hardware), the connection weights wbm are, for example, an amplitude, a frequency, or a phase of propagating waves. The connection weights wbm may be resistance values of variable resistors.
The first signals S11 to S1n are input to the reservoir layer 20. The first signals S11 to S1n interact with each other while propagating between a plurality of nodes 21 in the reservoir layer 20. Interaction between the first signals S11 to S1n means that a signal propagating to a node 21 affects a signal propagating to another node 21. For example, the connection weights wbm are applied to the first signals S11 to S1n at the time of propagation between the nodes 21, and the first signals S11 to S1n changes. The reservoir layer 20 projects the input first signals S11 to S1n to a multi-dimensional nonlinear space.
When the first signals S11 to S1n propagate between a plurality of nodes 21, the reservoir layer 20 generates a feature space including information of the first signals S11 to S1n input to the reservoir layer 20. In the reservoir layer 20, the input first signals S11 to S1n are converted to second signals S21 to S2i. Here, i is, for example, an arbitrary natural number and may be the same as or different from n. The second signals S21 to S2i hold at least a part of information included in the first signals S11 to S1n in changed forms. For example, the first signals S11 to S1n change nonlinearly in the reservoir layer 20 and become the second signals S21 to S2i. By allowing the first signals S11 to S1n to interact with each other in the reservoir layer 20, a system state of the reservoir layer 20 changes with the elapse of time.
Signals from the reservoir layer 20 are sent to the output layer 30. The second signals S21 to S2i output from the reservoir layer 20 are input to the output layer 30.
The output layer 30 includes, for example, a bias applying unit 31, a sum operation circuit 32, and an activation function circuit 33.
The bias applying unit 31 applies connection weights wc1 to wei to the second signals S21 to S2i. The connection weights wc1 to wci are applied to the second signals S21 to S2i, respectively. The bias applying unit 31 performs a product operation of multiplying the second signals S21 to S2i by the connection weights wc1 to wci.
The connection weights wc1 to wci may be the same or different from each other. The connection weights wc1 to wci represent a distribution in which values of the connection weights are plotted on the horizontal axis and the number of connection weights of a specific value is plotted on the vertical axis. The distribution of the connection weights wc1 to wci is determined on the basis of a task. The distribution of the connection weights wc1 to wci is, for example, a normal distribution, a uniform distribution, or a Laplacian distribution.
The connection weights wc1 to wci vary in the learning stage. On the other hand, the connection weights wc1 to wci are fixed on the basis of the results of learning based on a task in the learning stage in the arithmetic operation (inference) stage.
In an example in which the concept of reservoir computing is implemented in a physical device (hardware), the connection weights wc1 to wci are, for example, an amplitude, a frequency, or a phase of propagating waves. Waves have only to be generated due to vibration and examples thereof include an electromagnetic field, a magnetic field, a spin wave, and an elastic wave. In another example in which the concept of reservoir computing is implemented in a physical device (hardware), the connection weights wc1 to wci are, for example, resistance values of variable resistors.
The sum operation circuit 32 sums results of multiplication of the second signals S21 to S2i by the connection weights wc1 to wci. The sum operation circuit 32 may sum all the results of multiplication of the second signals S21 to S2i by the connection weights wc1 to wci or may sum some thereof. Results output from the sum operation circuit 32 may be one as illustrated in
The activation function circuit 33 substitutes the product-sum operation results into an activation function f(x) to perform calculation. The activation function circuit 33 converts the product-sum operation results nonlinearly. The activation function circuit 33 may be skipped.
In the learning stage, an output signal Sout from the output layer 30 is sent to the comparison circuit 60. In the arithmetic operation (inference) stage, the output signal Sout from the output layer 30 is externally output as a response. The output signal Sout is not limited to one signal. For example, when the information processing device 100 copes with a multi-class classification problem which is an application of general machine learning, the output layer 30 outputs a plurality of output signals Sout corresponding to the classes.
The comparison circuit 60 compares the output signal Sout with training data t. For example, the comparison circuit 60 compares the output signal Sout with the training data t to obtain the mutual information. The mutual information is a quantity of mutual dependency between two probability variables. The comparison circuit 60 transmits the comparison result to the instruction circuit 70.
The instruction circuit 70 sends an instruction to the bias applying unit 31 on the basis of the comparison circuit 60. The bias applying unit 31 updates the connection weights wc1 to wci on the basis of the instruction from the instruction circuit 70. When the connection weights wc1 to wci change, the output signal Sout from the output layer 30 changes. The instruction circuit 70 feeds back information to the bias applying unit 31 such that the mutual information between the output signal Sout and the training data t is increased (maximized). The connection weights wc1 to wci change on the basis of the fed-back data.
The evaluation circuit 40 calculates the distribution of the connection weights wc1 to wci in the output layer 30 and evaluates whether the distribution of the connection weights wc1 to wci is a prescribed distribution. In the learning stage, the evaluation circuit 40 may perform the evaluation whenever the connection weights wc1 to wci are updated or may perform the evaluation on the basis of a predetermined rule. The evaluation circuit 40 may calculate differences between the connection weights wc1 to wci before and after updating.
The distribution of the connection weights wc1 to wci is calculated, for example, by applying a reference signal to the bias applying unit 31. For example, a switch may be provided in the middle way of the second signals S21 to S2i to the bias applying unit 31 and be able to convert the second signals S21 to S2i and the reference signal. By inputting the reference signal with a fixed value instead of the second signals S21 to S2i, the connection weights wc1 to wci can be extracted. The reference signal is output from, for example, the evaluation circuit 40. For example, when the connection weights wc1 to wci are resistance values of memristors, the resistance values (the connection weights wc1 to wci) of the memristors may be calculated by applying a reference current to the memristors.
The adjustment circuit 50 adjusts the first signals Sli to S1n on the basis of the evaluation result from the evaluation circuit 40. The adjustment circuit 50 changes the adjustment parameters of the input layer 10, for example, when the distribution of the connection weights wc1 to wci is not a prescribed distribution. Examples of the adjustment parameters are the filter coefficients f1 to fn and the connection weights wa1 to wan. When the adjustment parameters are changed, the rule of signal conversion from the input signals S1n to Sinn to the first signals S11 to S1n changes, and the first signals S11 to S1n change. The adjustment circuit 50 may count the number of times of change of the adjustment parameters.
The evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60, and the instruction circuit 70 operate in the learning stage, but do not operate in the arithmetic operation (inference) stage.
Each of the evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60, and the instruction circuit 70 includes, for example, a processor such as a CPU and a memory. Each of the evaluation circuit 40, the adjustment circuit 50, the comparison circuit 60, and the instruction circuit 70 operates by causing the processor to execute a program. The processor instructs the constituent circuits to operate, and the memory stores the program or past results.
Some or all of the operations of the circuits may be realized by hardware such as an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a field-programmable gate array (FPGA). The program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disc, a ROM, a CD-ROM, or a semiconductor storage device (for example, a solid-state drive (SSD)) or a storage device such as a hard disk or a semiconductor storage device built in a computer system. The program may be transmitted via an electrical communication line.
The information processing device 100 outputs an output signal Sout when input signals Sin1 to Sinn are input thereto. The information processing device 100 compares training data t and the output signal Sout while changing the connection weights wc1 to wci of the output layer 30 and the adjustment parameters of the input layer 10. When a quantity of mutual information between the training data t and the output signal Sout increases sufficiently (when the rate of correct answers to a task increases sufficiently), learning ends.
First, in a learning process, a first process S1 is performed. In the first process S1, when the comparison result from the comparison circuit 60 is not sufficient (when the rate of correct answers to the task is not sufficient), the connection weights wc1 to wci of the output layer 30 are updated. The connection weights wc1 to wci of the output layer 30 are changed on the basis of an instruction from the instruction circuit 70 to the bias applying unit 31. For example, the first process S1 is performed a plurality of times, and the connection weights wc1 to wci are updated every time.
Subsequently, a second process S2 is performed. In the second process S2. changes D of the connection weights wc1 to wci between before and after updating are measured. The changes D of the connection weights wc1 to wci are measured, for example, by the evaluation circuit 40.
In the second process S2, it is determined whether the changes D of the connection weights wc1 to wci between before and after updating are equal to or less than a threshold value Δ. When the changes D of the connection weights wc1 to wci between before and after updating are equal to or less than the threshold value Δ, the output signal Sout from the output layer 30 converges. When the changes D of the connection weights wc1 to wci between before and after updating are equal to or less than the threshold value A, a third process S3 is performed. Each of the connection weights wc1 to wci and the threshold value Δ may be set separately. When the changes D of the connection weights wc1 to wci are greater than the threshold value Δ and the output signal Sout from the output layer 30 diverges, the process flow returns to the first process S1, and the connection weights wc1 to wci are updated.
In the third process S3, a distribution of the connection weights wc1 to wci in the output layer 30 is calculated. The distribution of the connection weights wc1 to wci is calculated by the evaluation circuit 40. The distribution of the connection weights wc1 to wci is calculated by plotting the values of the connection weights on the horizontal axis and plotting the number of connection weights at a specific value on the vertical axis. The values of the connection weights wc1 to wci are calculated, for example, by inputting the reference signal as described above.
Then, it is evaluated whether the distribution of the connection weights wc1 to wci corresponds to a prescribed distribution. The prescribed distribution is, for example, a normal distribution, a uniform distribution, or a Laplacian distribution. The prescribed distribution varies according to a task given to the information processing device 100. The prescribed distribution is determined at a time point at which the task is determined and is stored in the evaluation circuit 40.
For example, when the prescribed distribution is a normal distribution, for example, it is determined, for example, on the basis of a Kolmogorov-Smirnov test or a Shapiro-Wilk test. When a P value of the distribution of the connection weights wc1 to wci is less than 0.05, it can be determined that the distribution of the connection weights wc1 to wci is not a normal distribution.
When the distribution of the connection weights wc1 to wci corresponds to a prescribed distribution, adjustment parameters of the input layer 10 are determined in a fourth process S4. When the information processing device 100 exhibits a sufficient rate of correct answers after the adjustment parameters of the input layer 10 have been determined, learning ends. When the information processing device 100 does not exhibit a sufficient rate of correct answers after the adjustment parameters of the input layer 10 have been determined, the process flow returns to the first process S1, and the connection weights wc1 to wci of the output layer 30 are updated.
On the other hand, when the distribution of the connection weights wc1 to wci does not correspond to a prescribed distribution, the process flow proceeds to a fifth process S5. In the fifth process S5, it is determined whether the number of times of change of the adjustment parameters is equal to or greater than a prescribed number. The adjustment circuit 50 counts the number of times of change of the adjustment parameters. The prescribed frequency is appropriately determined depending on a task, accuracy of correct answers, a calculation load, and the like of the information processing device 100. The prescribed frequency is stored in the adjustment circuit 50.
When the number of times of change of the adjustment parameters is less than the prescribed number, the process flow proceeds to a sixth process S6, and 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 fn. The adjustment circuit 50 may change, for example, the connection weights wa1 to wan. The adjustment circuit 50 may change both the filter coefficients f1 to fn and the connection weights way to wan.
For example, the adjustment circuit 50 sets a distribution of the adjustment parameters of the input layer 10 to a normal distribution. When the adjustment parameters are changed by the adjustment circuit 50 a plurality of times, an average or a variance is changed on the basis of the normal distribution.
For example, the adjustment parameters are adjusted at the first time by the adjustment circuit 50, the distribution of the adjustment parameters is set to a normal distribution with an average of 0 and a variance of 1. Then, when the adjustment parameters are adjusted at the second time or later by the adjustment circuit 50, the distribution of adjustment parameters is changed such that the variance increases. For example, when the adjustment parameters are changed at the n-th time, the variance is increased by 10% of the variance value at the (n−1)-th time.
The adjustment circuit 50 may set the distribution of adjustment parameters of the input layer 10 to a uniform distribution with a fixed random number width. When the adjustment parameters are changed by the adjustment circuit 50 a plurality of times, the random number width is changed.
For example, when the adjustment parameters are changed by the adjustment circuit 50 at the first time, the distribution of adjustment parameters is set to a uniform distribution with a random number width of ±0.1. When the adjustment parameters are adjusted at the second time or later by the adjustment circuit 50, the distribution of adjustment parameters is changed such that the random number width increases. For example, when the adjustment parameters are changed at the n-th time, the distribution is set to a uniform distribution with a random number width ±0.1×n.
On the other hand, when the number of times of change of the adjustment parameters is equal to or greater than the prescribed number, the adjustment parameters of the input layer 10 are not changed any more. When the number of times of change of the adjustment parameters is equal to or greater than the prescribed number, the process flow proceeds to a seventh process S7. In the seventh process S7, optimal adjustment parameters are selected as the adjustment parameters of the input layer 10 out of the changed adjustment parameters. The adjustment parameters when the rate of correct answers is the highest in the change history are selected as the optimal adjustment parameters.
Through this process flow, the information processing device 100 determines the connection weights wc1 to wci of the output layer 30 and the adjustment parameters of the input layer 10 and ends the learning. In the arithmetic operation (inference) stage, the information processing device 100 performs an arithmetic operation using the connection weights wc1 to wci of the output layer 30 and the adjustment parameters of the input layer 10 which are determined in the learning stage.
An example of the process flow performed by the information processing device 100 has been described above, but the process flow is not limited to this example. For example, when the distribution of the connection weights wc1 to wci is evaluated whenever the connection weights wc1 to wci are updated, the second process S2 may not be performed. When an upper limit of the number of times of change of the adjustment parameters is not set, the fifth process S5 and the seventh process S7 may not be performed.
The information processing device 100 according to this embodiment adjusts the adjustment parameters of the input layer 10 on the basis of the distribution of the connection weights wc1 to wci of the output layer 30. The information processing device 100 can further enhance the rate of correct answers to a task by also adjusting the adjustment parameters of the input layer 10.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2021/038325 | 10/15/2021 | WO |
