JPH10198646A - Method and device for successively learning neural network - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable to apply to a lot of learning objects and to suppress the influence of forgetting by preserving the weight learnt in the past and considering the peculiar vector/peculiar value of double stage differentiation matrix of the weight in the past and an error function. SOLUTION: Data inputted at an input part 1 are stored in a data storage part 2. At the data storage part 2, the number of stored data is limited so that the data are stored while being limited to a timewisely nearby number. Next, a learning part 3 inputs the data stored in this storage part 2 and performs learning while using the weight learnt in the past and the peculiar value/peculiar vector stored in a hint storage part 5. A peculiar value/peculiar vector calculating part 4 calculates the peculiar value/peculiar vector of double stage differentiation matrix of error function learnt by the learning part 3. After the ending conditions of learning at the learning part 3 are satisfied, the weight and the peculiar value/peculiar vector calculated by the peculiar value/peculiar vector calculating part 4 are stored in the hint storage part 5.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network type sequential learning method and apparatus for sequentially learning given data.

[0002]

2. Description of the Related Art A conventional neural network type sequential learning machine stores a certain number of temporally neighboring data from input data and performs learning using the data. At this time, a phenomenon occurs in which the data that is no longer used is forgotten from the sequential learning device. As a method for avoiding this phenomenon of being forgotten, it is conceivable to store only the weights after learning and use the weights stored at the next learning to suppress forgetting. It works effectively only on limited learning targets and has not been established as a more versatile methodology.

[0003]

SUMMARY OF THE INVENTION An object of the present invention is to make it possible to apply the suppression of forgetting in a conventional neural network type sequential learning machine to a large number of learning objects by a better approximation of a neural network. It is in.

[0004]

The learning process in the present invention refers to modifying weights in a neural network to match given input / output data. In this case, in the first embodiment of the present invention, the correction of the weight does not use only the error and the inertia term for accelerating the speed of the correction and the weight after the past learning. The method is characterized in that learning is performed by taking account of eigenvectors, and means for calculating and storing eigenvalues / eigenvectors of a second-order differential matrix of weights and error functions after learning is used. Further, the second embodiment of the present invention is characterized in that means for calculating eigenvalues / eigenvectors of a second-order differential matrix of an error function after learning and means for converting the calculated eigenvectors to inertia terms are used. And

[0005]

Embodiments of the present invention will be described below in detail with reference to the drawings.

<First Embodiment> FIG. 1 is a block diagram of a sequential learning apparatus showing a first embodiment of the present invention. In FIG. 1, 1 is an input unit for inputting data to be learned, 2 is a data storage unit for storing data input by the input unit 1, and 3 is a weight and eigenvalue / eigenvector stored in a hint storage unit 5 described later. , A learning unit 4 for learning data stored in the data storage unit 2, and an eigenvalue / eigenvector calculation unit 5 for calculating eigenvalues / eigenvectors of a second order differential matrix of the error function after learning by the learning unit 3. Is a hint storage unit that stores the weight after learning by the learning unit 3 and the eigenvalue / eigenvector calculated by the eigenvalue / eigenvector calculation unit 4, 6 is an output unit that outputs the learning result, and 7 is a control unit that controls the whole. It is.

The outline of the operation of the sequential learning apparatus is as follows. Data input at the input unit 1 is stored in the data storage unit 2. The data storage unit 2 has a limitation on the number of data that can be stored. The learning unit 3 inputs the data stored in the data storage unit 2 and inputs the data into the hint storage unit 5.
The learning is performed using the weights and the eigenvalues / eigenvectors after the learning stored in the past. The eigenvalue / eigenvector calculation unit 4 calculates the eigenvalue / eigenvector of the second order differential matrix of the error function after learning by the learning unit 3. The weight after the learning unit 3 satisfies the condition for terminating the learning and the eigenvalue / eigenvector calculated by the eigenvalue / eigenvector calculation unit 4 are stored in the hint storage unit 5.

Next, the operation of the learning section 3 in this embodiment will be described in detail below taking as an example a case where the previous weight is considered.

The learning section 3 is composed of a multilayer artificial neural network including an input layer 31, an intermediate layer 32, and an output layer 33 as shown in FIG. The inside of the multilayer is composed of a multi-input / multi-output artificial neural unit 30. This artificial neural network learns the relationship between an input (I) and an output (T) given as weights between artificial neural units.

Now, the data stored in the data storage unit 2 is

[0011]

[Outside 1]

The weight and the eigenvalue / eigenvector stored in the hint storage unit 5 are

[0013]

[Outside 2]

It is assumed that In addition, the amount of weight correction that is corrected by the conventional back propagation method

[0015]

[Outside 3]

Then, the correction amount of the weight corrected in each artificial neural unit in the present embodiment

[0017]

[Outside 4]

Is represented by the following equation (1).

[0019]

(Equation 1)

Λ and α are constants. The error back-propagation method is described in, for example, the document "DE Rumelhart, G.
E. FIG. Hinton and R.S. J Williams (1986). “Le
arninginternal representations by error propagatio
n ”, in ParallelDistributed Processing: Explor
ations in the Microstructure of Cognition, Vol.
1, D. E. FIG. Rumelhard and J.M. L. McClelland (ed
s) MIT Press Cambridge MA. "It is described in.

The input / output relationship of the artificial neural network is

[0022]

(Equation 2)

In this case, the output of the artificial neural network after weight correction and the error function E of the given data are expressed by the following equation (3).
It is represented by

[0024]

(Equation 3)

The learning in the learning section 3 is performed by repeatedly performing the equation (1), and ends when the equation (3) and the weight | dW / dt | satisfy certain conditions.

Thereafter, the eigenvalue / eigenvector calculation unit 4 obtains a second-order differential matrix (Hessian matrix) based on the weights of the error function equation (3), and obtains eigenvalues and eigenvectors of the matrix. After that, weights and eigenvalues, eigenvectors

[0027]

[Outside 5]

Are stored in the hint storage unit 5. After the learning is completed, data whose output is desired to be known is input to the artificial neural network, and the output is examined. Further, when learning data is input, the above learning is repeated.

FIG. 3 shows a flowchart of a series of processing in this embodiment. In FIG. 3, 31 surrounded by a thick line
Processes 0 and 320 are the features of this embodiment.

FIG. 6 shows the performance when the present embodiment is applied to the learning of the Sin function, which had no effect in the prior art. From FIG. 6, it can be seen that in the present embodiment, the overall error is reduced earlier in the present embodiment than in the conventional method. In FIG. 6, in the present invention, since the oscillation of the entire error decreases and decreases rapidly, it is shown that the past data is well remembered, and the decrease in forgetting of the conventional method is suppressed. You can see that it is. Also, in this example, learning is performed only on data to be stored, which is simply data that is close in time, so that the effect of forgetting can be effectively achieved without using many calculations to select the data to be stored. It can be seen that can be reduced.

<Embodiment 2> FIG. 4 is a block diagram of a sequential learning apparatus showing a second embodiment of the present invention. In this embodiment,
The eigenvalue of the second derivative matrix of the error function after learning by the learning unit 3
Eigenvalue / eigenvector calculator 4 for calculating eigenvectors
And a conversion unit 8 for converting the eigenvalue / eigenvector calculated by the calculation unit 4 into an inertia term, and providing the inertia term to the learning unit 3.

The outline of the operation of the present sequential learning apparatus is as follows. Data input at the input unit 1 is stored in the data storage unit 2. As in the first embodiment of FIG. 1, the data storage unit 2 limits the number of data that can be stored, so that the data is limited to a number that is close in time. After the learning unit 3 learns the data stored in the data storage unit 2, the eigenvalue / eigenvector calculation unit 4 calculates the eigenvalue / eigenvector of the second derivative matrix (Hessian matrix) of the error function. The conversion unit 8
The eigenvector is transformed into an inertia term according to the eigenvalue.

Next, the operation of the learning section 3 in the present embodiment will be described in detail below, taking as an example a case where previous learning data is considered.

As shown in FIG. 2, the learning section 3 is composed of a multilayer artificial neural network composed of an input layer, an intermediate layer, and an output layer. Each layer has a multi-input / multi-output artificial neural unit. It consists of. The artificial neural network learns a relationship between an input (I) and an output (T) given as a weight between the artificial neural units.

Now, the data stored in the data storage unit 2 is

[0036]

[Outside 6]

Assume that: The input / output relationship of the artificial neural network

[0038]

(Equation 4)

In this case, the error function E between the output of the artificial neural network and the given data is represented by the following equation (5).

[0040]

(Equation 5)

The weight correction amount corrected by the conventional error back propagation method or the like is expressed by the following equation (6).

[0042]

(Equation 6)

The learning in the learning section 3 is performed by repeatedly performing the equation (6), and ends when the equation (5) and | dW | satisfy certain conditions. Thereafter, the eigenvalue / eigenvector calculation unit 4 obtains a Hessian matrix based on the weight of the error function expression (5), and calculates the eigenvalue and eigenvector of the matrix.

[0044]

[Outside 7]

Is obtained. After this, the conversion unit 8 converts the sum of the eigenvectors whose absolute value of the eigenvalue is smaller than a certain constant e into an inertia term as follows.

[0046]

(Equation 7)

After the learning is completed, data whose output is desired to be known is input to the artificial neural network, and the output is examined. Further, when learning data is input, the above learning is repeated.

FIG. 5 shows a flow chart of a series of processing in this embodiment. In FIG. 5, 51 surrounded by a thick line
Processes 0 and 520 are features of the present embodiment.

FIG. 7 shows the performance when the present embodiment is applied to the learning of the Sin function, which had no effect in the prior art. As in FIG. 6, it can be seen from FIG. 7 that the overall error is reduced earlier in this embodiment than in the conventional method.

As described above, in the description of each embodiment, the data stored in the data storage unit 2 is limited to data that is close in time, but it is also possible to include a process of selecting according to a certain rule.

[0051]

As described above, according to the present invention,
By learning past weights and considering the eigenvectors and eigenvalues of the second order differential matrix of the error function and the past weights, when learning data to be sequentially input, more learning targets And the effect of forgetting can be suppressed. Specifically, the present invention can be effectively used for devices such as online learning and prediction in which data is sequentially input, and has a wide range of applications such as weather, traffic, and stock price prediction.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a sequential learning type device according to a first embodiment of the present invention.

FIG. 2 is a diagram illustrating a configuration example of a multilayer artificial neural network of a learning unit.

FIG. 3 is an overall processing flowchart of the first embodiment of the present invention.

FIG. 4 is a configuration diagram of a sequential learning device according to a second embodiment of the present invention.

FIG. 5 is an overall processing flowchart of a second embodiment of the present invention.

FIG. 6 is a processing example according to the first embodiment of the present invention.

FIG. 7 is a processing example according to a second embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 1 input unit 2 data storage unit 3 learning unit 4 hint storage unit 5 eigenvalue / eigenvector calculation unit 6 output unit 7 control unit 8 conversion unit

Claims (4)

[Claims] 1. To match given input / output data,
A neural network-type sequential learning method for correcting weights in a neural network, comprising calculating eigenvalues / eigenvectors of a second-order differential matrix of an error function after learning, and holding the weights after learning and the eigenvalues / eigenvectors. And learning the given data using the held weights and eigenvalues / eigenvectors. 2. Input means for inputting a set of input and output of data to be learned, data storage means for storing input data of the input means, and hint storage means for storing weights after learning and eigenvalues / eigenvectors. Learning means for learning data stored in the data storage means using the weights and eigenvalues / eigenvectors stored in the hint storage means, and storing the learned weights in the hint storage means; An eigenvalue / eigenvector calculating means for calculating an eigenvalue / eigenvector of a second order differential matrix of the error function after learning by the learning means, and storing the eigenvalue / eigenvector in the hint storage means; A neural network type sequential learning apparatus, comprising: a control unit for controlling each of the units. 3. Matching given input / output data,
A neural network-type sequential learning method for correcting weights in a neural network, comprising calculating eigenvalues / eigenvectors of a second-order differential matrix of an error function after learning, converting the eigenvalues / eigenvectors into inertial terms, A neural network type sequential learning method characterized by learning given data using an inertia term. 4. Input means for inputting a set of input and output of data to be learned, data storage means for storing input data of the input means, and learning means for learning data stored in the data storage means. Eigenvalue / eigenvector calculation means for calculating eigenvalues / eigenvectors of a second-order differential matrix of the error function after learning by the learning means, and converting the eigenvalue / eigenvector calculated by the eigenvalue / eigenvector calculation means into an inertia term, A neural network type sequential learning apparatus, comprising: conversion means for giving to the learning means; output means for outputting a result learned by the learning means; and control means for controlling each of the means.