JPH086916A - Method and device for learning recurrent neural network - Google Patents

Abstract

PURPOSE:To attain stable and quick learning by learning only an initial state until the reduction of errors becomes less than a threshold in the case of extending a learning section in a past direction and learning it again, and then learning a weight value and a time constant also. CONSTITUTION:At the time of judging that a learning section is extended in a time-retroactive direction, a learning section control part 41 outputs a control signal to a learning control part 54 and the control part 54 receiving the signal outputs a control signal to a most nose diving direction calculating part 47, a network parameter correction quantity calculating part 48 and a network parameter correcting part 49. Immediately after extending the learning section in the time-retroactive direction, a convergence deciding part 46 decides an error at each time sent from a reverse propagation error calculating part 45 and the control part 41 learns only the initial state by the use of the extended time-sequential data without updating an already learned weight value and a time constant until the quantity of error reduction becomes less than a threshold. After its reduction less than the threshold, the control part 41 learns the weight value and the time constant in addition to the initial state.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention uses a recurrent neural network that allows arbitrary connections between units to perform learning of network weights using learning data in the processing field of pattern recognition and prediction. The present invention relates to a method and an apparatus therefor.

[0002]

2. Description of the Related Art A neural network is a simplified model of a neural network in the human brain, which is a network in which neuron neurons are connected via synapses through which signals pass in only one direction. . Signal transmission between neurons is performed through this synapse, and various information processing can be performed by appropriately adjusting the resistance of the synapse, that is, the weight. In each neuron, the output from the other connected neuron is input with synaptic weighting, the sum is added to the modified nonlinear response function, and the result is output again to the other neuron.

One of the structures of the neural network is
There is a multi-layer network as shown in FIG. This type of network has a layered structure, only inter-layer connections are allowed, no intra-layer or auto-regressive connections exist. This multi-layer network is considered to be suitable for recognition of spatially spreading patterns and information compression.

On the other hand, as shown in FIG. 3, there is a recurrent network which does not impose such a restriction on the network structure and allows arbitrary coupling between units. Strictly speaking, there is no concept of layer in the recurrent network, but here, in order to make correspondence with the multilayer network, the concept of layer is taken in for convenience. In the following, a unit group to which input data is input is called an input layer, a unit group that outputs a network output is called an output layer, and other unit groups are called an intermediate layer.

In recurrent networks, there is a coupling in which the past output of each unit is returned to another unit in the network or to itself. Therefore, the network has dynamics in which the state of each neuron changes depending on time. As described above, since the system has time in the network, it is considered to be suitable for recognizing and predicting a recurrent network time series pattern. Comparing the network shapes of the multilayer network and the recurrent network, the former can be seen as a special case of the latter.

A state equation followed by each neuron of the recurrent neural network is given by the following equations 1 and 2.

[0007]

[Equation 1]

[0008]

[Equation 2]

Here, x _i (t) is the internal state of the unit i at time t, and the output value y _i is determined by nonlinearly converting the internal state. Further, τ _i and X _i are the time constant and the external input of the unit i, respectively, and w _ij is the weight of the connection from the unit j to the unit i. N is the total number of units.

The state of the network at each time from time t ₀ to time t _n can be obtained by the following procedure. First, as an initial state, the internal state of the network and the output value at time t ₀ are set appropriately.
Then, the equations 1 and 2 are solved in the forward direction of time from t ₀ to t _n .

When a recurrent neural network is used for recognizing or predicting a time series pattern, input data, which is time series data, and teacher data, which is also time series data, are paired so that the network outputs a correct output. It is necessary to prepare learning data and use them to learn the weight value, time constant, and initial state of the network in advance. This is usually done using an optimization technique called the backpropagation method. As for this method, for example, learning algorithm to teaches patio temporal patterns to recurrent neural networks, biological cybernetics (A Learning Algorithm to Te
ach Spatiotemporal Patterns to Recurrent Neural Ne
2nd of tworks, Biological Cybernetics 62 (1990)
Discussed at pages 59-263.

The feature of this method is to correct the weight based on the steepest descent method so as to reduce the error in the finite time interval from t ₀ to t _n as given by the equation (3).

[0013]

(Equation 3)

Here, Y _k (t) is teacher data presented to the unit k at time t. Here, k represents a unit group belonging to the output layer, and N _v represents the number of units belonging to the output layer.

The steepest descent direction used in calculating the weight value, the time constant, and the correction amount in the initial state is given by the following equations 4, 5, and 6.

[0016]

[Equation 4]

[0017]

(Equation 5)

[0018]

(Equation 6)

Here, P _i (t) is the back propagation error of unit i, respectively. The back-propagation error is calculated according to Equation 7.

[0020]

(Equation 7)

Here, δ _ik is the Kronecker delta symbol. In Expression 7, the back propagation error of the neuron belonging to the output layer is in a form in which the error of the output value is added at each time in addition to the back propagation error from other neurons and the sum of the weights. Also, df (x _i ) / dx _i , (y _i (t) −Y _i (t))
Cannot be calculated unless x _i (t) and y _i (t) are determined.
Therefore, after calculating x _i (t) and y _i (t) in the forward direction of time, it is assumed that the back propagation error is 0 at time t _n , and the boundary condition given by Equation 8 is set, The back propagation error is t _n →
Calculate in the opposite direction of the time of t ₀ .

[0022]

[Equation 8]

The procedure for calculating the steepest descent direction is as follows. That is, first, after properly setting the initial state of each unit at time t ₀ , the internal state and output value at each time are calculated according to equations 1 and 2, and the values are stored.
Next, the boundary condition of the back propagation error given by the equation 8 at the time t _n is set. After that, using the internal state and the output value calculated earlier, the backward propagation error at each time is calculated and stored while moving backward in time according to Equation 7. Finally, using the obtained output value at each time and the back propagation error,
Calculate the steepest descent direction given by.

Details of the processing procedure of the back propagation learning will be described with reference to FIG.

First, the learning data 8 in FIG. 4 will be described. This is, for example, learning data for a network having two input layer neurons and one output layer neuron as shown in FIG. As shown in FIG. 5, the learning data is made up of input data and teacher data, and has data numbers of (the number of input layer neurons × the number of time series sample points) and (the number of output layer neurons × the number of time series sample points), respectively.

In step 20, the weight value, the time constant, and the initial value of the initial state are set by using an appropriate random number or the like.

In step 22, after properly setting the initial state of the network, that is, the internal state of the network and the output value, the internal state of each unit is calculated in the forward direction of time using the input data according to Equations 1 and 2. Calculate the output value and save it.

In step 23, the boundary conditions of equation 8 are first set. After that, the back propagation error P _i (t) of each unit is calculated in the backward direction of time using the teacher data according to Formula 7, and the calculated back propagation error P _i (t) is stored.

In step 24, the steepest descent direction is calculated according to equations (4), (5) and (6) using the internal state and output value obtained in step 22 and the back propagation error obtained in step 23.

In step 25, the equation 9 and the equation 1 are obtained according to the steepest descent direction obtained in step 24 and the correction amount of the previous weight.
The correction amount of the current weight is calculated according to 0 and 11.

[0031]

[Equation 9]

[0032]

[Equation 10]

[0033]

[Equation 11]

Here, η is a learning coefficient, α is a moment coefficient, and n is the number of times of learning. The second term on the right side is called a moment term and is an term added empirically to accelerate learning.

In step 26, the following equation 12, equation 13,
According to equation 14, the weight of each unit is modified.

[0036]

[Equation 12]

[0037]

[Equation 13]

[0038]

[Equation 14]

In step 21, the processing from step 22 to step 26 is repeated until the error converges below a certain value.

A problem of the back propagation learning of the recurrent neural network is that the learning does not easily converge when the length of time series data for learning becomes long. This problem will be described in detail.

The recurrent neural network is
It is a nonlinear dynamical system with time inside the network.
The learning of the weight value, the time constant, and the initial state corresponds to adjusting the parameters of this nonlinear dynamic system. In a non-linear dynamics system, there is a phenomenon called bifurcation in which the behavior of the system suddenly greatly changes at certain values as the parameters are changed. Therefore, when the learning coefficient is set to a large value, an unstable phenomenon in which the gradually decreasing error suddenly increases due to this branching phenomenon often occurs.

On the other hand, the error depends on the weight value, the time constant, and the initial state, and can be regarded as a function of those parameters. Therefore, learning can be rephrased as searching the minimum point of the error function in the parameter space because the parameter is adjusted so as to reduce the error. On the other hand, learning is to give constraints to the trajectory drawn by the network so as to follow the teacher data, and to adjust the parameters to satisfy those constraints. Therefore, increasing the length of the time series to be learned will search for parameters that satisfy more constraints. The more constrained the trajectory of the network, the more complicated the error function surface becomes,
There will be many peaks and valleys and local minimums, so it will be difficult for learning to converge.

In order to solve this problem, instead of learning long time series data at once from the beginning, learning is first performed using short time series data that is a part of the learning data, and the learning converges. A method was proposed in which each time the learning was performed, the length of the time-series data was made to gradually increase and the learning was performed. This will be described with reference to FIG.

At first, the weight value, the time constant, and the initial state are learned in the learning section of the first stage. When the learning of the section converges, the learning section is expanded, and the learning value of the second step is learned again with the weight value and the time constant learned in the first step as a starting point. Similarly, the learning is advanced while expanding the learning section in the third stage and the fourth stage. When expanding in the time direction in the learning section proceeds always state at the time t _a is the initial state of the network for each of the learning period. Therefore, even if the learning section is expanded, the initial state learned before the expansion can be used as it is.

[0045]

On the other hand, in order to avoid problems such as the local minimum, the learning section does not always expand in the direction of time progress as shown in FIG. 6, and as shown in FIG. May be expanded in the past. But,
In this case, there is a problem in learning the initial state.

This will be described with reference to FIG. When the learning in the first stage has converged, the state at time t _d is obtained by learning as the initial state. Next, it is assumed that the learning section is expanded and learning is performed on the second-step learning section. In that case, the state at time t _c is required as the initial state, but since it cannot be obtained by the first stage learning, it is necessary to set an appropriate initial state and perform learning again.

On the other hand, the state of the recurrent neural network is determined by a non-linear differential equation as shown in equations 1 and 2. Therefore, the trajectory drawn by the state of the network largely depends not only on the weight values and time constants that are the parameters of the equation, but also on the initial state. Therefore, even if the weight value and the time constant are the same, the trajectory drawn may differ greatly if the initial state is different. As a result, if the initial state is appropriately initialized when the learning section is expanded to the second stage, even the trajectory of the learning section of the first stage, which has followed the turning angle teacher signal due to learning, is significantly deviated.

In this case, the learning for the second stage learning section is started from the state where the trajectory is largely deviated also for the first stage learning section. As a result, when the learning section is expanded from the first stage to the second stage, it becomes closer to the form of performing learning for the learning section of the second stage from the beginning rather than performing additional learning of the section from t _c to t _d. .

In order to prevent this, at the time when the learning section is expanded to the second stage, the state of t _c is calculated by solving equations 1 and 2 in the direction backward from the state of t _d as a starting point. Can be considered. However, calculating backwards in time is not practical because the trajectory becomes unstable due to the form of the equation.

The present invention is a method for back propagation learning of a recurrent neural network, which realizes stable and high-speed learning even when a learning section is expanded in a backward direction in order to avoid a local minimum. Is to provide a method.

[0051]

The above problem is that when the learning section is expanded in the past direction and learning is performed again, immediately after the learning section is expanded, the amount of error reduction is equal to or less than the threshold value. Up to this point, the weight value and time constant are not updated, only the initial state is learned, and then the weight value and the time constant including the initial state are learned.

That is, a step of setting a time series section shorter than the entire length of the time series prepared in advance as learning data as an initial learning section of the time series data, and a boundary condition for the set time series section, Using the step of setting the initial state and the time series data of the set section, solve the state equation of the network in the direction of time, so that it approaches the teacher data that is the target trajectory,
At the step of learning the initial state, the weight value, and the time constant, and at the time when the learning using the time series section is completed, the length of the time series section used for learning for the next learning is traced back in time. A step of expanding, a step of setting a new initial state for the expanded time series section, and a step of setting a new initial state for the expanded time series section based on the newly set initial state Solving the state equation of the network and performing learning again until the amount of error reduction of the set initial state is less than or equal to the threshold value so as to approach the teacher data in the expanded time series interval,
After learning only the initial state, the initial state, weight value,
A recurrent type characterized by including a step of learning a time constant and a step of repeating the above until the length of the time series section becomes the length of the entire time series of the training data prepared in advance. It is solved by a neural network learning method.

[0053]

According to the above method, even if it is necessary to expand the learning section in the direction going back in time and set a new initial state, only the initial state is maintained for a fixed number of learning times for the expanded time series section. By learning, a new initial state of the expanded learning section can be obtained. This initial state is
The weight values and time constants learned in the time series section before expansion are set again so as to be closest to the time series data after expansion. Therefore, the trajectory drawn by the network does not deviate significantly from the teacher data because the initial state is set inappropriately. As a result, it is possible to prevent the trajectory from being greatly deviated, and the learning of the time series section before expansion is wasted, and it is possible to realize high-speed learning.

[0054]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention will be described below with reference to FIGS. 1, 3, 4, 5, 7, and 8.

FIG. 1 is a processing flowchart of the embodiment, and the features of the present invention are underlined.

FIG. 3 is an explanatory diagram showing an example of the recurrent neural network. Strictly speaking, there is no concept of layers in the recurrent network, but here, for convenience, the unit group to which input data is input is the input layer, the unit group that outputs the network output is the output layer, and the other unit groups are intermediate. Call it a layer.

FIG. 4 shows the weight value of the recurrent network,
It is a flow chart of processing of back propagation learning to a time constant and an initial state.

FIG. 5 shows learning data 8 for performing weight learning.
It is an explanatory view showing the details of. This is learning data for a network having two input layer neurons and one output layer neuron as shown in FIG. 3, for example. The learning data is composed of input data and teacher data, and has a number of data equal to (the number of input layer neurons × the number of time series data samples) and (the number of output layer neurons × the number of time series data samples).

Next, the operation of this embodiment will be described with reference to FIGS. 1, 4 and 8. In FIG. 1, steps 4 and 5 characterize the present invention.

In step 1, an appropriate random number is used to start learning, and a weight value of the network, a time constant, and an initial value of an initial state are set.

In step 2, the initial learning section of the time series data to be learned first is set.

At step 4, it is judged whether the learning section has been expanded in the backward direction. When the initial learning section is set, or when the learning section is expanded in the direction in which time advances, the process directly proceeds to step 7. If the learning section is expanded in the direction going back in time, the process proceeds to step 5.

In step 5, the initial state is learned using the already learned weight value and time constant, and the expanded time series data. Details of step 5 will be described with reference to FIG.

In step 31, after the initial state of the network is set, the output value of each unit is calculated in the direction of time progress according to equations 1 and 2 by using the input data whose learning section has been expanded, and the calculated value is calculated. save.

In step 32, the boundary conditions of equation 8 are first set. After that, the back propagation error P _i (t) of each unit is calculated in the backward direction of time using the teacher data according to Formula 7, and the calculated back propagation error P _i (t) is stored.

In step 33, using the output value obtained in step 30 and the back propagation error obtained in step 31,
The steepest descent direction with respect to the initial state is calculated according to Equation 6.

In step 34, the correction amount in the initial state this time is calculated according to the equation 11 by the steepest descent direction obtained in step 32 and the correction amount in the previous time. In Equation 11, η
Is a learning coefficient, α is a moment coefficient, and n is the number of times of learning. The second term on the right side is called a moment term and is an term added empirically to accelerate learning.

In step 35, the initial state is corrected according to the equation (14).

In step 30, steps 31 to 35 are repeated until the amount of error reduction becomes equal to or less than a certain threshold value.

Through the processing of steps 30 to 35 described above, the initial state immediately after the learning section is expanded is learned, and a new initial state is set.

In step 7, the weight value, time constant, and initial state of the network are learned with the initial state learned in step 5 as the starting point for the expanded learning section. Details of step 7 will be described using steps 22 to 26 in FIG.

In step 22, after the initial state of the network is set, the internal state and output value of each unit are calculated in accordance with the equations 1 and 2 using the input data, and the output values are stored.

In step 23, first, the boundary conditions of equation 8 are set. After that, the back propagation error P _i (t) of each unit is calculated in the backward direction of time using the teacher data according to Equation 7, and the calculated back propagation error P _i (t) is stored.

In step 24, using the output value obtained in step 22 and the back propagation error obtained in step 23,
The weight value, the time constant, and the steepest descent direction for the initial state are calculated according to the equations 4, 5, and 6, respectively.

In step 25, the current weight value, time constant, and initial state correction amount are calculated in accordance with equations 9, 10, and 11 based on the steepest descent direction obtained in step 24 and the previous correction amount.

In step 26, the number 12, the number 13, the number 1
4, the weight value, the time constant, and the initial state are corrected.

The weight value, the time constant, and the initial state are corrected by the processing of steps 22 to 26 described above.

In step 6, the error of the units belonging to all the output layers at each time converges to be equal to or less than a preset value, or the mean square error of the output layer units at each time is set to a preset value. Until it converges to
The process of step 7 is repeated.

In step 3, the learning section is expanded by a certain length, and steps 4 to 7 are repeated until the length of the time-series data prepared in advance is reached.

With such a method, learning is started from a short learning section, and the learning section is gradually widened in a direction going back in time. Therefore, stable and high-speed learning is possible as compared with learning a long learning section from the beginning. .

The learning method of the recurrent neural network is not limited to that of the present embodiment using the back propagation error which requires a process that goes backward in time.
For example, while performing forward calculation, at the same time a method that can learn in real time by modifying the weight based on the steepest descent method,
A method using the variational principle and a non-linear optimization method such as a conjugate gradient method or a quasi-Newton method may be adopted. Regarding the first two methods, for example, a study on each learning rule of the recurrent neural network, the shape of the learning surface, the Institute of Electronics, Information and Communication Engineers D-II Vol.j74-DI
I No. 12 (December 1991) pp. 1776 to 1787.

The method of expanding the learning section is not limited to that in this embodiment, and may be exponentially lengthened with the number of times of expansion, for example. Further, the learning convergence condition for moving to the next learning section is not limited to that of the present embodiment, and the learning error of the increment of the section to be expanded next from the current learning section is minimized. Alternatively, the time point or the mean square error may be equal to or less than a certain threshold value.

In addition, the model of the recurrent network is not limited to that of this embodiment, and the synapse characteristic may have not only the weight value but also the delay characteristic.

Next, the operation of the second embodiment will be described with reference to FIGS. 1 and 9.

FIG. 9 shows the entire functional structure of the apparatus of the second embodiment. In particular, the functional blocks showing the features of the present invention are underlined.

In step 1, the network parameter initialization unit 43 sets the weight value of the network, the time constant, and the initial value of the initial state by using an appropriate random number to start learning, and stores them in the network parameter storage unit 42. To do.

In step 2, the learning section control unit 41 outputs a control signal to the network state calculation unit 44, the back propagation error calculation unit 45, and the steepest descent direction calculation unit 47 to set the learning section of the time-series data for the first learning. Set.

In step 4, the learning section control unit 41 determines whether or not the learning section has been expanded in the direction going back in time. When the initial learning section is set, or when the learning section is expanded in the direction in which time advances, the process directly proceeds to step 7.
If the learning section is expanded in the backward direction, step 5
Proceed to.

In step 5, the initial state is learned using the weight values and time constants already learned and the expanded time series data. Therefore, the learning section control unit 41 outputs a control signal to the learning control unit 54. Upon receipt of the control signal, the learning control unit 54 outputs a control signal to the steepest descent direction calculation unit 47, the network parameter correction amount calculation unit 48, and the network parameter correction unit 49 to learn the initial state. Details of step 5 will be described with reference to FIG.

In step 31, after the network state calculation unit 44 sets the initial state of the network, the input data corresponding to the current learning section is read from the learning data storage unit 40, and the forward direction of time is calculated according to the equations 1 and 2. Then, the internal state and output value of each unit are calculated and stored in the network state storage unit 50.

In step 32, the back propagation error calculation section 45
First sets the boundary condition of Eq. After that, the teacher data corresponding to the current learning section is read from the learning data storage unit 40, the back propagation error P _i (t) of each unit is calculated in the backward direction of time according to Equation 7, and the back propagation error P _i (t) is calculated. 5
Store in 1. Then, the error of the output value at each time of the units belonging to the output layer is sent to the convergence determination unit 46.

In step 33, the steepest descent method direction calculation unit 47 first reads the time constant from the network parameter storage unit 50, and reads the back propagation error at the initial time from the back propagation error storage unit 51. Next, the steepest descent method direction calculation unit 47 calculates the steepest descent direction according to Equation 6 using these data, and stores it in the steepest descent direction storage unit 52.

In step 34, first, the correction amount calculation unit 48
However, the steepest descent direction data is read from the steepest descent direction storage unit 52, and the correction amount in the previous initial state is read from the correction amount storage unit 53. Next, the correction amount calculation unit 48 calculates the correction amount in the current initial state using these data according to the equation 11, and stores it in the correction amount storage unit 53.

In step 35, the network parameter correction unit 49 first reads the initial state data from the network parameter storage unit 42 and the initial state correction amount from the network parameter correction amount storage unit 53. Next, the correction unit 49 corrects the initial state of each unit according to Equation 14, and stores new initial state data in the network parameter storage unit 42.

In step 30, the convergence determination section 46 determines the error at each time sent from the back propagation error calculation section 45 in step 32, and the processing from step 31 to step 35 is performed until it converges to a certain value or less. repeat.

By the above processing from step 30 to step 35, learning of the initial state immediately after the expansion of the learning section is performed, and a new initial state is set.

In step 7, the initial state of the network, the weight value, and the time constant are learned for the current learning section.
For details of step 7, see step 22 to step 2 in FIG.
This will be described using 6.

In step 22, after the network state calculation unit 44 sets the initial state of the network, the input data corresponding to the current learning section is read from the learning data storage unit 40, and the forward direction of time is calculated according to the equations 1 and 2. Then, the internal state and output value of each unit are calculated and stored in the network state storage unit 50.

In step 23, the back propagation error calculation section 45
First sets the boundary condition of Eq. After that, the teacher data corresponding to the current learning section is read from the learning data storage unit 40, the back propagation error P _i (t) of each unit is calculated in the backward direction of time according to Equation 7, and the back propagation error P _i (t) is calculated. 5
Store in 1. Then, the error of the output value at each time of the units belonging to the output layer is sent to the convergence determination unit 46.

In step 24, the steepest descent method direction calculation unit 47 first reads the internal state and the output value from the network state storage unit 50, and reads the back propagation error from the back propagation error storage unit 51. Next, the steepest descent method direction calculation unit 47 calculates the weight value, the time constant, and the steepest descent direction in the initial state, respectively, using these data in accordance with Formula 4, Formula 5, and Formula 6, and stores them in the steepest descent direction storage unit. Store in 52.

In step 25, first, the network parameter correction amount calculation unit 48 reads the steepest descent direction data from the steepest descent direction storage unit 52 and the previous correction amount from the network parameter correction amount storage unit 53. Next, the network parameter correction amount calculation unit 48 calculates the weight value, the time constant, and the correction amount in the initial state of this time, respectively, in accordance with Formulas 9, 10, and 11 using these data,
It is stored in the network parameter correction amount storage unit 53.

In step 26, first, the network parameter correction unit 49 causes the network parameter storage unit 4
The weight value, the time constant, and the initial state data are read from 2 respectively, and the weight value, the time constant, and the constant are read from the correction amount storage unit 53, respectively.
Read the correction amount in the initial state. Next, the network parameter correction unit 49 corrects the weight of each unit according to Formula 12, Formula 13, and Formula 14, and stores a new value in the network parameter storage unit 42.

In step 6, the convergence determination section 46 determines the error at each time sent from the back propagation error calculation section 45 in step 23, and the processing from step 22 to step 26 is performed until it converges to a certain value or less. repeat.

In step 3, the learning section control unit 41 outputs a control signal to the network state calculation unit 44, the back propagation error calculation unit 45, and the steepest descent direction calculation unit 47 to expand the learning section by a certain length. Then, steps 4 to 7 are repeated until the length of the time-series data to be learned prepared is reached.

[0105]

According to the present invention, even if it is necessary to expand the learning section in the direction going back in time and newly set the initial state, the reduction amount of the error is reduced with respect to the expanded time series section. By learning only the initial state until it becomes a certain threshold value or less, a new initial state used for re-learning the weight value and the time constant can be obtained. This initial state is
The optimum value is set again so as to be closest to the time series data after expansion using the weight value and the time constant learned in the time series section before expansion. Therefore, the trajectory drawn by the network does not deviate significantly from the teacher data because the initial state is set inappropriately. As a result, it is possible to prevent the trajectory from being greatly deviated, and the learning of the time series section before expansion is wasted, and it is possible to realize high-speed learning.

[Brief description of drawings]

FIG. 1 is a flowchart of a process according to a first embodiment of this invention.

FIG. 2 is an explanatory diagram showing an example of a multilayer neural network.

FIG. 3 is an explanatory diagram showing an example of a recurrent neural network.

FIG. 4 is a weight value of a recurrent neural network,
The processing flowchart of back propagation learning with respect to a time constant and an initial state.

FIG. 5 is an explanatory diagram showing details of learning data.

FIG. 6 is an explanatory diagram showing an example of expansion of a learning section in a direction in which time advances.

FIG. 7 is an explanatory diagram showing an example of expansion of a learning section in a direction going back in time.

FIG. 8 is a flowchart of learning processing in an initial state.

FIG. 9 is a block diagram of an apparatus according to a second embodiment of the present invention.

[Explanation of symbols]

1 ... weight value, time constant, initial setting of initial state, 2 ... initial setting of learning section, 3 ... expansion of learning section, 5 ... learning of initial state of network, 7 ... correction of weight value, time constant, initial state , 8 ... Learning data.

Front page continuation (72) Inventor Yumiko Kitagawa 2-4-1, Ichibancho, Aoba-ku, Sendai City, Miyagi Prefecture Hitachi Tohoku Software Co., Ltd.

Claims (5)

[Claims] 1. A basic unit having a function of weighting and summing outputs from other combined basic units, and performing a non-linear conversion on the summed result and outputting the result, between arbitrary basic units. In the supervised learning method using the learning data composed of the time-series data as the input data to be input to the network and the time-series data as the teacher data, which is the target value of the network output, for the recurrent neural network that allows the connection of As the initial learning section of the time series data, a step of setting a time series section shorter than the entire length of the time series prepared in advance as learning data, and using the time series data of the set section, a target trajectory is obtained. Learning network parameters including boundary states so as to approach the teacher data; When the learning using the time-series interval is completed, a step of expanding the length of the time-series interval used for learning to perform the next learning, and a new boundary state is added to the expanded time-series interval. Setting again, using the expanded time-series interval, re-learning only the boundary states of the network until a preset condition is satisfied, and after learning only the boundary states, the boundary states are also Including the step of learning the parameters that the network has, and the step of repeating the above until the length of the time series section becomes the length of the entire time series of the training data prepared in advance. Recurrent neural network learning method. 2. A base unit having a function of weighting and summing outputs from other combined base units, and performing a non-linear conversion on the summed result and outputting the result, between arbitrary base units. For a recurrent neural network that allows the connection of the As the initial learning section of the time series data, a step of setting a time series section shorter than the entire length of the time series prepared in advance as learning data, and a state equation of the network using the time series data of the set section Is solved in the direction of time, and the initial state is included so that it approaches the teacher data, which is the target trajectory. A step of learning the parameters of the network, and a step of expanding the length of the time series section used for learning to carry out the next learning in a direction going back in time, when the learning using the time series section is completed, A step of setting a new initial state for the expanded time series section; and a step of setting an error reduction amount only for the initial state of the newly set network for the expanded time series section The step of learning again until it becomes less than or equal to the threshold value, the step of learning all the parameters of the network including the initial state after the learning of only the initial state is finished, and the length of the time series section is A recurrent neural network characterized by including the steps of repeating the above until the length of the entire time series of the training data prepared in advance. Network learning method. 3. A basic unit having a function of weighting and summing outputs from other combined basic units, performing a non-linear conversion on the summed result, and outputting the result. In a supervised learning method that uses learning data composed of time-series data as input data to be input to the network and time-series data as teacher data that is the target value of the network output, for a recurrent neural network that allows the connection of As the initial learning section of the time series data, a step of setting a time series section shorter than the entire length of the time series prepared in advance as learning data, and an initial state is set as a boundary condition for the set time series section. And the time series data of the set section, the state equation of the network is calculated. Steps for learning the initial state, weight values, and time constants to solve the problem in the direction of progress and approach the teacher data, which is the target trajectory, and the next learning at the time when learning using the time series interval ends. Expanding the length of the time series section used for learning in order to go back in time, setting a new initial state for the expanded time series section, and the expanded time series section On the other hand, based on the newly set initial state, the state equation of the network is solved in the direction of time progress, and the error is reduced only in the set initial state so that it approaches the teacher data in the expanded time series section. The step of performing learning again until is below a certain threshold, and after learning only the initial state, the initial state, the weight value,
A recurrent type characterized by including a step of learning a time constant and a step of repeating the above until the length of the time series section becomes the length of the entire time series of the training data prepared in advance. Neural network learning method. 4. The recurrent neural network learning method according to claim 3, wherein the backpropagation learning is backpropagation learning with a process in which a back propagation error goes back in time. 5. A basic unit having a function of weighting and summing outputs from other combined basic units, and performing a non-linear conversion on the summed result and outputting the result, between arbitrary basic units. In a supervised learning device using learning data, which consists of time-series data as input data to be input to the network and time-series data as teacher data that is the target value of the network output, for a recurrent neural network that allows the connection of , Means for setting a time series section that is shorter than the entire length of the time series prepared in advance as learning data as the initial learning section of the time series data, and an initial state is set as a boundary condition for the set time series section And the time-series data of the set section to advance the state equation of the network A method for learning the initial state, weight values, and time constants so that they are solved in the direction and approach the teacher data, which is the target trajectory, and for the next learning when the learning using the time-series interval ends. Means for expanding the length of the time-series interval used for learning in a backward direction of time, means for setting a new initial state for the expanded time-series interval, and for the expanded time-series interval , Based on the newly set initial state, solve the state equation of the network in the direction of time progress, and set only the initial state set so that it approaches the teacher data of the expanded time series interval. Means for performing learning again until the amount of decrease is less than or equal to a certain threshold, and after the learning of only the initial state is completed, the initial state, the weight value,
A recurrent type characterized by including means for learning a time constant and means for repeating the above until the length of the time series section becomes the length of the entire time series of the training data prepared in advance. Neural network learning device.