JP3312042B2 - Learning machine - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to character recognition, image recognition,
Such as a recognition device such as voice recognition, a prediction device such as stock price prediction, or a control device such as motion control, machine operation control, etc.
The present invention relates to a learning machine used in a general device that obtains a desired output from a given input.

[0002]

2. Description of the Related Art In a general learning machine including a neural network, when an input is given, parameters are learned so as to obtain a desired output with respect to the input. Conventionally, as this learning method, a back-propagation method as disclosed in, for example, a document “Backpropagation Hideki Aso, Compute Roll 24 Special Issue Neurocomputer (Corona)” is known.

[0003]

By the way, when this kind of learning machine is applied to, for example, character recognition, if a certain character is erroneously recognized as a result of actual character recognition after learning parameters, Re-learning of parameters is desired so that characters can be correctly recognized. However, conventionally, no consideration has been given to parameter re-learning. One of the reasons is that, in the above example, when re-learning is performed so as to correctly recognize a character in which misrecognition has occurred (that is, data x ₀ that outputs an undesired output is data to be newly learned, If re-learning is performed using only the new data to be learned), there is a problem that other characters that have been correctly recognized up to now are erroneously recognized and often have an adverse effect. The inventor of the present application
This problem occurs when re-learning is performed using only the data x ₀ to be newly learned as input data, and the specific gravity of the data x ₀ with respect to the distribution of the entire learning data at the time of initial learning is unknown. It was determined to occur. Therefore, in order to solve this problem by preparing again learning data used for the initial learning, more accurate sample of the entire distribution union of the initial training data and specific data x _0, ie input As data, I first thought of a method to use it for relearning. However, in this method, it is necessary to always retain the initial learning data, and in engineering applications, it is disadvantageous in terms of storage capacity to retain the initial learning data constantly and lacks practicality. There was a problem.

An object of the present invention is to provide a learning machine capable of re-learning which does not adversely affect other data without always holding initial learning data.

[0005]

In order to achieve the above object, according to the first aspect of the present invention, a parameter storage means for holding a parameter representing a probability distribution in an input / output space and an input are provided. Sometimes, according to the parameters held in the parameter storage means, output probability calculation means for calculating the probability distribution of the output, learning the parameters using a set of the input and the desired output corresponding thereto as learning data, the parameter storage means Parameter learning means for setting, input probability calculation means for calculating a probability distribution of the input according to the parameters held in the parameter storage means when the output is given, and input according to the probability distribution calculated by the input probability calculation means The union of the sample and the data to be newly learned and the set of desired outputs for To perform the re-learning of the parameter by parameter learning unit using, it is characterized in having a re-learning means for resetting was subjected to the re-learning parameter in the parameter storage means. Since the input sample generated using the input probability calculation circuit is a good approximation of the initial learning data, re-learning is performed using the union of this input sample and the data x ₀ to be newly learned. By doing so, without always retaining the initial learning data,
In addition, it is possible to perform relearning with little adverse effect on other data.

Further, in the invention according to claim 2, the output probability calculating means and the input probability calculating means are realized by one neural network having a reversible configuration. Thereby, a good input sample can be easily obtained with a simple configuration.

The invention according to claim 3 further comprises learning data number storage means for holding the number of initial learning data, wherein the re-learning means comprises an input sample according to the probability distribution calculated by the input probability calculating means. Out of the number of samples held in the learning data number storage means, and a parameter learning is performed by using a union of the sample and data to be newly learned and a set of desired outputs corresponding thereto as re-learning data. It is characterized in that parameters are re-learned by means. In such a configuration, input samples corresponding to the number of initial learning data can be obtained, so that an input sample that better approximates the learning data at the time of the initial learning can be obtained. Re-learning is performed by using the union of と and a set of desired outputs for them as re-learning data, whereby better re-learning can be performed.

Further, the invention according to claim 4 further comprises learning data number storage means for holding the number of used relearning data after relearning is performed by the relearning means. Of the input samples according to the probability distribution calculated by the input probability calculation means, the same number of samples as the number held in the learning data number storage means are extracted, and the union of the sample and the data to be newly learned and The parameter is re-learned by the parameter learning means by using a set of desired outputs to the re-learning data. Thereby, at the time of re-learning, the initial learning data
The ratio of the learning frequency between the data and the data to be newly learned can be automatically determined, and learning more faithful to the distribution of external data becomes possible.

Further, the invention according to claim 5 further comprises a learning rate storing means for indicating a learning intensity of the relearning data, wherein the relearning means includes a number of input samples according to the probability distribution calculated by the input probability calculating means. The re-learning data is determined so that the ratio between the number of data to be newly learned and the number of data to be newly learned matches the ratio determined according to the value held in the learning rate storage means, and parameter learning is performed using the re-learning data. It is characterized in that parameters are re-learned by means. Thus, at the time of re-learning, the operator can arbitrarily determine the ratio of the learning frequency between the initial learning data and the data to be newly learned, and only the specific data can be determined. It becomes possible to learn strongly.

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a configuration diagram of a first embodiment of a learning machine according to the present invention. The learning machine of FIG. 1 has an input / output space (x, y)
A parameter storage unit 10 for holding a parameter w representing the above probability distribution, and a hierarchical neural network for outputting a probability distribution of an output y in accordance with the parameter w stored in the parameter storage unit 10 when an input x is given Learning of the parameter w is performed based on learning data including a network NW and a set of an input and a desired output corresponding thereto, and the learned parameter w is stored in the parameter storage unit 1.
It has a parameter learning unit 11 that sets the parameter w to 0 and a relearning unit 12 that relearns the parameter w and sets the relearned parameter w in the parameter storage unit 10.

Here, the hierarchical neural network NW includes an input element 1, a first intermediate layer 2, a normalized cell 3, a second intermediate layer 4, and an output element 5,
This neural network NW is configured reversibly. That is, when the input x is applied to the input element 1, the probability distribution of the output y when the input x is given is obtained from the output element 5, and when the output y is applied to the output element 5, From 1, the probability distribution of the input x when the output y is given can be obtained. Therefore, this neural network NW
When an input x is given, it functions as an output probability calculation circuit that calculates a probability distribution of an output y according to a parameter w held in a parameter storage unit 10, and when an output y is given, the parameter storage unit 10 Also functions as an input probability calculation circuit that calculates the probability distribution of the input x in accordance with the parameter w held in.

During normal operation, the neural network NW functions as an output probability calculation circuit, and when an input x is applied to the input element 1, the parameter storage unit 10
The probability distribution of the output y is calculated in accordance with the parameter w stored in the output element 5 and output from the output element 5. That is, at the time of learning the parameter w (at the time of initial learning), the parameter learning unit 11 learns the parameter w based on learning data composed of a set of an input and a desired output, and obtains a desired output y for the input x. Is updated, the parameter w in the parameter storage unit 10 is updated. When the parameter w is learned by the parameter learning unit 11 and the desired parameter w is set in the parameter storage unit 10, this neural network NW is used for actual character recognition, voice recognition, etc. On the other hand, an output y can be obtained.

However, even if the parameter w is learned and the parameter w considered to be desirable is set in the parameter storage unit 10, characters or the like which are erroneously recognized in character recognition or the like may occur. For example, it may output for a particular input x ₀ may become inadequate.
In such a case, in the learning machine of FIG.
In order to re-learn the parameter w by using a parameter w, an appropriate output is returned even for a specific input x _{0. In} performing the re-learning, the neural network NW may be used as an input probability calculation circuit. It works.

[0014] That is, the in the case of performing the re-learning, as described above, using the union of the initial training data used during learning and x ₀ (data want to learn i.e. newly) which specific data as input data However, holding learning data used at the time of initial learning is disadvantageous in terms of storage capacity. Therefore, in the first embodiment, instead of holding the learning data used in the initial learning,
The re-learning unit 12 generates a teacher signal (desired output with respect to the input) taught as an output during the initial learning, adds these to the output element 5, and outputs a sample of the input vector (input sample) from the input element 1. , Input sample output from input element 1 and specific data (data to be newly learned)
The parameter learning unit 11 re-learns the union with x ₀ and the desired output set for them as re-learning data. As described above, when generating input samples instead of the initial learning data, the neural network NW is made to function as an input probability calculation circuit.

Next, the processing operation of the learning machine having such a configuration will be described in more detail. During normal processing operation,
The neural network NW functions as an output probability calculating circuit, and the input element 1 receives, for example, M-dimensional input vector data x from an external device (not shown). For example, when applied to character recognition, a feature vector of a character extracted by an external device is used as an input. When the input vector data x is added to the input element 1, the neural network, that is, the output probability calculation circuit calculates the conditional probability distribution of the N-dimensional output vector data y when x is given according to the given input x. I do. Here, the probability distribution satisfying the input / output (x, y) is estimated by the probability density function p (w; x, y) using the parameter w stored in the parameter storage unit 10. At this time, the output probability calculating circuit uses the parameter w held in the parameter storage unit 10 to calculate the conditional probability p (w; y | x) of the output y when the input x is input according to the following equation. Perform a numerical operation.

[0016]

(Equation 1)

[0017] Now, for example, R (x), S and (y) respectively R ^M, and the probability density function on ^{R N, ξ∈R M, η∈R N} ,
For ρ, σ ≧ 0, it is defined as follows.

[0018]

(Equation 2)

In this case, the probability density function p (w; x, y)
As w = (θ _h , ξ _h , ρ _h , η _h , σ _h ) (1 ≦ h ≦
The following equation can be used with H) as a parameter.

[0020]

(Equation 3)

In the following, it is assumed that R and S are represented by a Gaussian distribution as shown in the following equation.

[0022]

(Equation 4)

At this time, the conditional probability p (w; y | x)
Is as follows:

[0024]

(Equation 5)

The probability distribution p (w; y |
In order to generate the output y in accordance with x), the following processing is specifically performed in the neural network NW. First, the first intermediate layer 2 includes parameters ξ _h , ρ _h , θ
_{Using h} and input vector x, output O _h ⁽¹⁾ (x) of each element
Is calculated as in the following equation.

[0026]

(Equation 6)

Next, the normalized cell 3 calculates and outputs the sum O (x) of the outputs O _h ⁽¹⁾ (x) of the respective elements of the first intermediate layer 2 by the following equation.

[0028]

(Equation 7)

Here, α _h (x) = O _h ⁽¹⁾ (x) / O (x)
Satisfies the following equation.

[0030]

(Equation 8)

Therefore, ｛α ₁ (x), α ₂ (x),..., Α
_H (x)} represents probabilities on H discrete sets {1, 2,..., H}. Based on this discrete distribution, the set {1,2,2,
, H}, one element h ₀ is selected, and the output value of the second intermediate layer 4 is set to (00,..., 1,..., 0). That is, only the output value corresponding to one element h ₀ is set to “1”.

The output element 5 is a neural network N
One output y of W is stochastically determined as a sample according to the Gaussian distribution S (η _h0 , σ _h0 ; y). In this way, the output probability calculation circuit can generate the output value y as a sample according to the probability distribution of Expression 5.

In order for the output y to be obtained in accordance with the desired probability distribution, it is necessary to determine the parameter w to an appropriate value in the output probability calculation circuit. This learning is performed by the parameter learning unit 11. Previously Dzu, upon learning, called learning data, input a set of examples of desirable output thereto _{{(x s, y s)} | (1 ≦
s ≦ S)｝ is prepared. For example, when this neural network NW is applied to character recognition, the feature vector for the characters “A”, “I”, “U” _,.
And then, also, the output y _s corresponding to them (1,0,
0, ...), (0, 1, 0, ...), (0, 0, 1, ...),
...

After initializing the parameter w by a method such as using a random number, the parameter learning unit 11 performs the maximum likelihood estimation by the steepest ascent method. That is, as in the following equation:
The value of the parameter w is updated according to the following rule of the steepest ascent method so that the value of the log likelihood function L (w) becomes as large as possible.

[0035]

(Equation 9)

When a Gaussian distribution is used, p _s = p
When writing (w; x _s , y _s ), the parameter w is updated sequentially according to the following equation. When the amount of change of the parameter w becomes smaller than a certain threshold, the update is stopped and written into the parameter storage unit 10. .

[0037]

(Equation 10)

The parameter w determined as described above and finally stored in the parameter storage unit 10 is used by the output probability calculation circuit during an operation such as actual character recognition.

After the parameter w is set by learning, when an operation such as actual character recognition is performed,
Erroneous recognition of characters may occur. In such a case, that is, when the output for the input vector x ₀ is inappropriate, and would like to re-learn to appropriate output is returned. At this time, the re-learning unit 12 causes the neural network NW to function as an input probability calculation circuit, generates teacher signals taught as outputs in the initial learning, and generates samples of input vectors corresponding thereto from the input probability calculation circuit.

That is, the input probability calculation circuit numerically calculates the conditional probability p (w; x | y) of the input vector x when the output vector y is given according to the following equation.

[0041]

[Equation 11]

The probability distribution p (w; x |
In order to generate the input x according to y), the following processing is specifically performed in the neural network NW. That is, first, the second intermediate layer 4 has the parameter η
_{Using h} , σ _h , θ _h and the output vector y, the output r _h ⁽¹⁾ (x) of each element is calculated as follows.

[0043]

(Equation 12)

Next, the normalized cell 3 calculates and outputs the total sum r (x) of the outputs of the respective elements of the second intermediate layer 4 by the following equation.

[0045]

(Equation 13)

Here, β _h (x) = r _h ⁽¹⁾ (x) / r (x)
Satisfies the following equation.

[0047]

[Equation 14]

Therefore, ｛β ₁ (x), β ₂ (x),.
_H (x)} represents probabilities on H discrete sets {1, 2,..., H}. Based on this discrete distribution, the set {1,2,2,
, H}, one element h ₀ is selected, and the output value of the first intermediate layer 2 is set to (0, 0,..., 1,..., 0). That is, only the output value corresponding to one element h ₀ is set to “1”. The input element 1 stochastically determines one output in the reverse direction of the neural network NW as a sample according to a Gaussian distribution R (ξ _h0 , ρ _h0 ; x). In this way,
The input probability calculation circuit operates in the same manner as the output probability calculation circuit, except that the calculation order of each element is reversed.
Can be generated according to the probability distribution of

The re-learning unit 12 uses the union of these input samples and the vector data x ₀ to be newly learned and a set of desired outputs for them as re-learning data.
As in the case of the initial learning, the parameter learning unit 11 re-learns the parameter w.

In such a process, the input sample generated by using the input probability calculation circuit is a good approximation of the initial learning data. Therefore, the input sample and the data x ₀ to be newly learned are obtained. By performing re-learning using the union and a desired set of outputs as the re-learning data, it is possible to perform re-learning with little adverse effect on other data.

Also, of the initial learning data, the input data
As described above, since the input sample is obtained using the input probability calculation circuit instead of the input sample, it is not necessary to always hold the input sample. Also, a desired output (teacher signal) of the initial learning data can be regularly generated by a simple circuit or CPU or the like, and the desired output (teacher signal) is automatically generated by a simple circuit or the like. If created dynamically, there is no need to keep this. More specifically, in the pattern recognition as shown in the above-described example, output elements are prepared by the number of categories to be identified, and a desired output (teacher signal) for a certain input x is provided.
For example, a method of setting a vector in which only the output element corresponding to the category to which the input x belongs to "1", such as (0,..., 1,... 0), is often used. In such a case, when generating re-learning input data (input samples), (1, 0,..., 0), (0,
, 0), (0, 0, 1, 0,..., 0),... To input elements to generate input data. In practice, these are simple circuits. Alternatively, it can be created regularly by a CPU or the like. Therefore, according to this embodiment, the re-learning does not always hold the initial learning data consisting of the pair of the input and the desired output, and has little adverse effect on other data. Can be performed.

[0052] In the above example, the data x ₀ to be newly learned, which is not limited to one, may be a case where there exist a plurality. Also in this case, re-learning that does not adversely affect other data can be performed by exactly the same procedure.

FIG. 2 is a configuration diagram of a second embodiment of the learning machine according to the present invention. In FIG. 2, the same parts as those in FIG. 1 are denoted by the same reference numerals. In this second embodiment,
A learning data number storage unit 13 for storing the number of learning data used at the time of the initial learning is further provided, and the re-learning unit 12 outputs the teacher signal taught as an output at the time of the initial learning to the learning data number storage unit 13. Are generated for the number of pieces of learning data held in, the input samples corresponding thereto are generated in the input probability calculation circuit, and the parameter learning unit 11 performs re-learning.

In such a configuration, among the teacher signals taught as outputs during the initial learning, the number of input samples corresponding to the number of data held in the learning data number storage unit 13 can be obtained. Can be obtained, and a re-training is performed using a union of the input sample and data x ₀ to be newly _trained and a set of desired outputs for the unsampled data as re-training data. Thereby, a better parameter w can be obtained.

FIG. 3 is a block diagram of a third embodiment of the learning machine according to the present invention. In the third embodiment, the relearning unit 1
2, after the re-learning is performed, a learning data number storage unit 14 for storing the number of re-learning data is provided. The number held by the unit 14 is generated, input samples for these are generated by the input probability calculation circuit, and the parameter learning unit 11 performs re-learning.

In such a configuration, when the re-learning is completed, the number of the re-learning data is written in the learning data number storage unit 14. Thus, the value held in the learning data number storage unit 14 becomes the total number of data learned so far. Thereby, the re-learning unit 12 generates the teacher signal taught as an output at the time of the initial learning by the total number of data learned so far, the union of the input sample for these and the data x ₀ to be newly learned, and the Re-learning is performed using a desired set of outputs as re-learning data. As described above, in the third embodiment, when the neural network is re-learned after the learning has been performed, the learning frequency ratio between the data that has been learned so far and the data that is to be newly learned is automatically determined. Can be determined, and learning that is more faithful to the distribution of external data becomes possible.

FIG. 4 is a block diagram of a fourth embodiment of the learning machine according to the present invention. In the fourth embodiment, a learning rate storage unit 15 for holding a learning intensity G determined by an operator's instruction or the like is provided, and the re-learning unit 12 generates a teacher signal taught as an output during initial learning. However, at this time, the number of teacher signals to be generated is determined so that the ratio between the data x ₀ to be newly learned and the generated teacher signal is G: 1, and input samples for these teacher signals are input to the input probability calculation circuit. The parameter is generated, and re-learning is performed by the parameter learning unit 11.

In such a configuration, when re-learning the neural network once the learning has been performed, the operator or the like arbitrarily determines the ratio of the learning frequency between the initial learning data and the data to be newly learned. Thus, it is possible to strongly learn only specific data, that is, only new data to be learned.

[0059]

As described above, according to the first aspect of the present invention, when performing re-learning, the re-learning means includes:
An output is provided to the input probability calculation means to calculate an input probability distribution according to the parameters held in the parameter storage means. In this case, the input sample generated by the input probability calculation means is initially initialized. Since the learning data of is a good approximation, by performing re-learning using the union of this input sample and the data to be newly learned, the initial learning data is not always held.
Re-learning that does not adversely affect other data can be performed. Further, since the output probability calculating means and the input probability calculating means are realized by one neural network having a reversible configuration, a good input sample can be easily obtained with a simple configuration. Obtainable.

According to the third aspect of the present invention, there is further provided a learning data number storing means for holding the number of initial learning data, wherein the re-learning means follows the probability distribution calculated by the input probability calculating means. Among the input samples, the same number of samples as the number held in the learning data number storage means are extracted, and a union of the sample and data to be newly learned and a set of desired outputs for them are used as relearning data. Since parameter re-learning is performed by the parameter learning means, an input sample that better approximates the learning data at the time of initial learning can be obtained, and better re-learning is performed using this input sample. be able to.

According to the fourth aspect of the present invention, after the re-learning is performed by the re-learning means, there is further provided a learning data number storage means for holding the number of used re-learning data. Means for extracting, from among input samples according to the probability distribution calculated by the input probability calculation means, as many samples as the number held in the learning data number storage means, and a union of the samples and data to be newly learned In addition, the parameter learning means is used to re-learn the parameters by using a set of desired outputs and re-learning data, so that when re-learning, the initial learning data and the data to be newly learned are obtained. The ratio of the learning frequency with the data can be automatically determined, and learning can be performed more faithfully in the distribution of external data.

According to the fifth aspect of the present invention, there is further provided a learning rate storing means for indicating a learning intensity of the relearning data, wherein the relearning means has an input sample according to the probability distribution calculated by the input probability calculating means. The re-learning data is determined by using the re-learning data so that the ratio between the number of data and the number of data to be newly learned matches the ratio determined according to the value held in the learning rate storage means. Since the parameters are re-learned by the parameter learning means,
At the time of re-learning, the operator or the like can arbitrarily determine the ratio of the learning frequency between the initial learning data and the data to be newly learned, whereby only specific data can be determined. It becomes possible to learn strongly.

[Brief description of the drawings]

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

FIG. 2 is a configuration diagram of a second embodiment of the learning machine according to the present invention.

FIG. 3 is a configuration diagram of a third embodiment of the learning machine according to the present invention.

FIG. 4 is a configuration diagram of a fourth embodiment of the learning machine according to the present invention.

[Explanation of symbols]

 Reference Signs List 1 input element 2 first intermediate layer 3 normalized cell 4 second intermediate layer 5 output element 10 parameter storage unit 11 parameter learning unit 12 relearning unit 13, 14 learning data number storage unit 15 learning rate storage Section NW neural network

──────────────────────────────────────────────────続 き Continuation of the front page (56) References Sumio Watanabe and Kenji Fukumizu, "Unified Theory of Neural Networks and Proposal of New Model", IEICE Technical Report, Japan, The Institute of Electronics, Information and Communication Engineers, Japan -Published, March 18, 1992, Vol. 91, No. 529 (NC91-98 to 131), pp. 179-186 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06N 1/00-7/08 G06F 9/44 G06F 19/00 G06K 9/00 -9/82 G10L 15/00-17/00 G06T 7/00-7/60 JST file (JOIS) CSDB (Japan Patent Office)

Claims (5)

(57) [Claims] 1. Parameter storage means for storing a parameter representing a probability distribution in an input / output space, and output probability calculating an output probability distribution in accordance with parameters stored in the parameter storage means when an input is given. A calculating means, a parameter learning means for learning a parameter using a set of an input and a desired output corresponding thereto as learning data, and setting the parameter in the parameter storage means; Input probability calculation means for calculating the probability distribution of the input according to the parameters which are present, a union of the input sample according to the probability distribution calculated by the input probability calculation means and the data to be newly learned, and a desired output set for them. Using the parameter learning means to re-learn the parameters using as learning data, A re-learning means for resetting re-learned parameters in the parameter storage means. 2. The learning machine according to claim 1, wherein said output probability calculating means and said input probability calculating means are realized by one neural network having a reversible configuration. 3. The learning machine according to claim 1, further comprising learning data number storage means for holding the number of initial learning data, wherein said re-learning means follows the probability distribution calculated by the input probability calculation means. Among the input samples, the same number of samples as the number held in the learning data number storage unit are extracted, and a union of the sample and data to be newly learned and a set of desired outputs corresponding thereto are used as relearning data. A parameter learning means for causing the parameter learning means to relearn parameters. 4. The learning machine according to claim 1, further comprising learning data number storage means for holding the number of used relearning data after the relearning is performed by said relearning means, The learning means extracts, from among input samples according to the probability distribution calculated by the input probability calculation means, as many samples as the number held in the learning data number storage means,
A learning machine characterized in that the parameter learning means performs parameter re-learning by using a union of the sample and data to be newly learned and a desired output set for the union set as re-learning data. . 5. The learning machine according to claim 1, further comprising a learning rate storage unit that represents a learning intensity of the relearning data, wherein the relearning unit is configured to input samples according to the probability distribution calculated by the input probability calculating unit. The re-learning data is determined by using the re-learning data so that the ratio between the number of data and the number of data to be newly learned matches the ratio determined according to the value held in the learning rate storage means. A learning machine characterized in that the parameter learning means re-learns parameters.