JPH09259101A - Category sorting method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To sort the vector categories even when the umber of categories is unknown. SOLUTION: When a vector X is inputted at the time of learning and the vector X belongs to one of existing categories C1 ...Ci , the mean M and a covariance matrix Σ are calculated and corrected (5). If the vector X does not belong to any of categories C1 ...Ci , a new category Ci+1 is decided to show M=X and Σ=Σ0 (4). After the learning is over, 'n-variate probability density function f (X)' is calculated based on the M and Σ for the decided category and against the input vector X. Then the category where the value of f (x) is maximized is defined as a category where the vector X belongs (8).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a category classification method, and more particularly to a method of classifying an n-dimensional vector quantity input after learning using a neural network into any of a plurality of categories.

[0002]

2. Description of the Related Art As a conventional category classification method, for example, there is a composite self-organization type pattern classification system described in JP-A-04-288663.

This conventional pattern classification system is composed of self-organizing type classifying means and comprehensive judging means. The self-organizing type classifying means stores the weighting coefficients learned in advance in the learning mode, receives the feature signal vector F _i as an input, and receives the M elements r _i1 , ..., M corresponding to the M categories to be classified. Generate an output signal vector R _i consisting of r _iM . Here, the element r _ij indicates how far the input pattern signal vector S is from the category C _j and the feature signal vector F _i . In the comprehensive judgment means, among the _K output signal vectors R ₁ , ..., _RK , the elements r _1i , ..., R _{Ki of} the output signal vector corresponding to the category C _j.
, And the total output values corresponding to the categories C ₁ , ..., C _M are obtained. Then, the category name that gives the smallest value in the total output value is determined as the classification result, and this is output as the classification response signal P.

[0004]

In the category classification method by the above-mentioned conventional system, it is necessary to determine the number of categories to be classified in advance and then learn the weighting coefficient to be stored in the self-organizing type classification means. It has the drawback that it is difficult to use if the number is unknown. Also,
When a category is added after learning once, there is a problem that the weighting coefficient has to be learned again.

An object of the present invention is to provide a general category classification method in which the above-mentioned drawbacks are eliminated.

[0006]

According to the category classification method of the present invention, a constant constituting a function having an input vector as an independent variable is determined by learning for each category, and a dependent variable of the function for each category with respect to the vector. In the category classification method that determines the one to which the vector of the category belongs by comparing the sizes of, if the constant is undecided for the category to which the vector input at the time of learning, the constant in the category is determined according to the vector, If another vector belonging to the category to which the vector input at the time of learning is already input and the constant is already determined for the category, the constant is changed according to the input vector and the other vector. To do.

According to the category classification method of the present invention, the mean and covariance matrix are determined for each category by learning, and the category having the maximum value of the n-variable probability density function obtained for each category with respect to the input vector is determined as the vector. If the mean and covariance matrix for the category to which the vector input at the time of learning belongs is not determined, the mean and the covariance matrix for the category are determined according to the vector, and the category to which the vector input at the time of learning belongs If another vector belonging to is already input and the mean and covariance matrix is already determined for the category, the mean and covariance matrix is changed according to the input vector and the other vector. To do.

The category classification method of the present invention determines the mean and the root mean square by learning for each category,
For a vector whose input components have no correlation with each other, the category having the maximum value of the n-variable probability density function obtained by using only the diagonal components of the covariance matrix for each category is set as the category to which the vector belongs, and learning If the mean and the root mean square for the category to which the vector to be inputted at times is undecided, the mean and the mean square are determined for the category according to the vector, and the mean and the mean square are related to the category to which the vector to be inputted during learning belongs. Is determined from the input vector, and the predetermined average and the square mean, and the average and the square mean for the category are changed. And

According to the category classification method of the present invention, an average is determined by learning for each category, the category having the smallest distance from the average of the input vector is set as the category to which the vector belongs, and the category to which the vector input at the time of learning belongs If the average is undecided about, determine the average in the category according to the vector,
If the average is already determined for the category to which the vector input at the time of learning belongs, the average for the category is changed by the average obtained from the input vector and the predetermined average. To do.

[0010]

Next, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a diagram showing the configuration of one neuron having a neural network used in the embodiment of the present invention.

When an n-dimensional vector X is input, the neuron outputs Y = f (X) as an output. Here, the function f is an n-variable probability density function,

[0013]

Where M is the average of n-dimensional vectors and Σ is n
Represents a xn covariance matrix. ┃Σ┃ is the determinant of Σ,
Σ ⁻¹ is the inverse matrix of Σ.

The mean M and the covariance matrix Σ are determined by learning.

FIG. 2 is a diagram showing the configuration of the neural network used in the embodiment of the present invention.

Each individual neuron forming the network corresponds to one category, and when there are a plurality of categories, the network is composed of a plurality of neurons. In FIG. 2, the vector X is the category C ₁ , C.
₂ , ..., C _i are input to the neurons, and scalar values Y ₁ , Y ₂ , ..., Y _i ... Are output from the respective neurons.

In category classification, learning is first performed. At the time of learning, when a certain vector X is input, and this vector X belongs to any one of preset i categories C ₁ , ... C _i , the average corresponding to the category is input. The M and covariance matrix Σ are recalculated and corrected from all the stored inputs that have belonged to that category up to that point. On the other hand, when the vector X does not belong to any category, a new neuron having mean M = X and covariance matrix Σ = Σ ₀ (common variance matrix determined by the input vector X) is newly assigned to the new category C _{i + 1} . Correspond and add.

After learning, for a certain input vector X,
For all categories that are set from the beginning and are added after that, using the mean M and the covariance matrix Σ obtained by learning in each category, the n-variable probability density function f (X ). And f
The category having the maximum value of (X) is set as the category to which the input vector X belongs, and the category classification for the vector X is performed.

If there is no correlation for each element of the input vector X, the covariance matrix Σ is a diagonal matrix.

[0021]

[0022] and can be simplified, μ _1, ..., diagonal sigma ₁₁ of mu _n and covariance matrix sigma, ..., it can be calculated knowing the sigma _nn.

That is, at the time of learning, only the mean M and the diagonal components of the covariance matrix Σ are added or updated, and after learning, f (X) is calculated by the equation (6), so the calculation amount is Be reduced.

When there is no correlation for each element of the vector X, at the time of learning, in which category the input X (1) = (x _{1 (1)} , ..., X _{n (1)} ) ^T (7) exists. If it does not belong, mean M (1) = (μ _{1 (1)} , ..., μ _{n (1)} ) ^T = X (1) (8) root mean square K (1) = (ν _{1 (1)} , ..., ν _{n (1)} ) ^T = (x _{1 (1)} ² , ..., X _{n (1)} ² ) ^T (9) is stored as information.

The input vector X (k) = (x _{1 (k)} , ..., X _{n (k)} ) ^T (10) belongs to any one of the categories, and the input belonging to that category up to that time is k. −1, and when the current vector X (k) is the kth input, the average M (k−1) = (μ _{1 (k−1)} , ..., μ _n stored at that time _(k-1) ) ^T (11) root mean square K (k-1) = (ν _{1 (k-1)} , ..., ν _{n (k-1)} ) ^T (12), the average M (k) = (Μ _{1 (k)} ,…, μ _{n (k)} ) ^T (13) and the root mean square K (k) = (ν _{1 (k)} ,…, ν _{n (k)} ) ^T (14) are respectively

[0026]

Is updated.

At this time, the dispersion can be calculated as V (k) = (ν _{1 (k)} −μ _{1 (k)} ² , ..., V _{n (k)} −μ _{n (k)} ² ) ^T (17). V _{1 (k)} -μ _{1 (k)} ² , ... V _{n (k)} -μ _{n (k)} ²
Is σ ₁₁ , ... σ _{nn, the} output value Y can be calculated by the equation (6) for the input vector X at the time of classification.

That is, if the vector X is set so that there is no correlation for each element, at the time of learning, even if it is not necessary to store all the inputs belonging to each category until then, all the categories are successively stored. So average and 2
It suffices to update and store the multiplicative average. By doing so, the amount of information to be stored is greatly reduced.

Instead of (1), Y = f (X) as an output value for an n-dimensional input vector X, where the function f is a function representing a distance.

[0031]

A neuron that is However, X = (x ₁ , ..., X _n ) ^T M = (μ ₁ , ..., μ _n ) ^T , and M represents the average of n-dimensional vectors.

The average M is determined by learning.

First, at the time of learning, a certain vector X
, And this vector X has i categories C ₁ ,
, C _i belongs to any one of the categories, the average M corresponding to the category is stored from all the inputs that have been stored in the category until then,
Recalculate and make corrections. On the other hand, if the input X does not belong to any category, a new neuron having an average M = X, that is, a category C _{i + 1} is newly added.

After learning, for a certain input vector X,
For all existing categories, the average M learned in each category is used to obtain the distance f (X) from the average from the equation (18). Then, the category having the smallest value of f (x) is set as the category to which the input vector X belongs, and category classification is performed on the input vector X.

At the time of learning, if the input vector X (1) = (x _{1 (1)} , ..., X _{n (1)} ) ^T (19) does not belong to any category, the average M ( 1) = (μ _{1 (1)} , ..., μ _{n (1)} ) ^T = X (1) (20) is stored as information.

The input vector X (k) = (X _{1 (k)} , ..., X _{n (k)} ) ^T (21) belongs to any one of the categories, and the inputs that have belonged to that category by then are k. -1 and the current input is the k-th, the average stored at that time M (k-1) = (μ1 _(k-1) , ..., μn _(k-1) ) ^{From T} (22), the average M (k) = (μ _{1 (k)} , ..., μ _{n (k)} ) ^T (23) is

[0038]

And is updated.

That is, at the time of learning, for all categories, only the average is updated and stored sequentially without storing all the inputs belonging to each category until then. Good. By doing this,
The amount of information to be stored is greatly reduced.

In any case, by storing the information required until the end of the previous learning, additional learning from that information can be easily performed.

[Embodiment] FIG. 3 is a flow chart showing a category classification method according to an embodiment of the present invention. This category classification method is executed by an information processing device having an arithmetic processing device, an input / output device, a storage device and the like (not shown). The case where the figure is classified into a plurality of characters will be described below as an example.

In step 1, the image input device and the graphic processing device input a vector X including information about a graphic. In step 2, a person judges the character indicated by the figure, that is, the category-C, and designates it with a keyboard or the like.

In step 3, it is checked whether the category-C is stored in the storage device as already set. If not stored, in step 4, the average M and the root mean square K are obtained from the above equations (8) and (9),
The category C, the vector X, the average M, and the common variance matrix Σ are stored in the storage device in association with each other.

If the category-C designated in step 3 is already stored in the storage device,
The average M and the common variance matrix Σ calculated by the arithmetic processing unit from all the vectors X stored in the storage device corresponding to the category-C and the vector X input in step 1 correspond to the category-C. Then, the vector X input in step 1 is additionally stored together with the vector X already stored corresponding to the category-C. When the learning is further continued in step 6 of steps 4 and 5, the process returns to step 1.

A plurality of categories C are set in advance,
Steps 1 to 6 are repeatedly executed until the mean value M and the common variance matrix Σ are obtained for all of these categories C, but a new category other than the category C that has been set depending on the vector X input during this period. Can also be set.

When the learning is completed in step 6, the vector X for the image for character recognition is input in the next step 7, and then the input vector X is input for all the categories C stored in the storage device. Mean M and common variance matrix Σ stored in the storage device corresponding to each category
The value Y = f (x) of the n-variable probability density function is calculated by the above-described equation (1) in the arithmetic processing device using, and the category C having the maximum value Y, for example, the character to which the input vector X belongs Output (step 8). Step 6 when performing category classification or learning for the new vector X
Return to (Step 9).

[0048]

As described above, according to the category classification method of the present invention, the constants of the functions such as the mean and covariance matrix of the n-variate probability density function for the category to which the input vector belongs are undecided. According to the default, learning is performed by newly adding or updating constants that are information characterizing categories, and after learning, for all categories that exist for a certain input vector,
The constants defined for each category are used to determine the category to which the input vector belongs, and category classification is performed on the input vector, so there is no need to determine the number of categories to be classified in advance, and the number of categories is unknown. Also in the case, there is an effect that the input vector can be classified into a category without paying particular attention.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of one neuron used in an embodiment of the present invention.

FIG. 2 is a diagram showing a configuration of a neural network used in the embodiment of the present invention.

FIG. 3 is a flowchart showing a category classification method according to an embodiment of the present invention.

[Explanation of symbols]

X vector x ₁ , x ₂ , x _n element of vector X Y output values Y ₁ , Y ₂ , Y _i output values C ₁ , C ₂ , C _i categories

Claims (4)

[Claims] 1. A constant that constitutes a function having an input vector as an independent variable is determined for each category by learning, and the magnitude of a dependent variable with respect to the vector of the function for each category is compared to compare the vector of the category. In the category classification method that determines what belongs to
If the constant is undecided for the category to which the vector to be input at the time of learning is determined, the constant in the category is determined according to the vector, and another vector belonging to the category to which the vector to be input at the time of learning has already been input If the constant is already defined, the constant is changed according to the input vector and the other vector. 2. A mean and covariance matrix are determined by learning for each category, and the category having the maximum value of the n-variable probability density function obtained for each category with respect to the input vector is set as the category to which the vector belongs, and learning is performed. If the mean and covariance matrix is undecided for the category to which the vector to be input at times is determined, the mean and the covariance matrix in the category are determined according to the vector, and other vectors belonging to the category to which the vector to be input during learning are A category classification method, characterized in that if the mean and covariance matrix have been already input for the category, the mean and covariance matrix is changed according to the input vector and other vectors. 3. An n-variate probability obtained by determining a mean and a square mean by learning for each category and using only diagonal components of a covariance matrix for each category for a vector whose input components have no correlation with each other. The category having the maximum value of the density function is set as the category to which the vector belongs, and if the mean and the root mean square of the category to which the vector input at the time of learning belong are undecided, the mean and 2 in the category are calculated according to the vector.
If the mean and the root mean of the category to which the vector input at the time of learning belongs is defined, the mean and the root mean obtained from the input vector and the predetermined mean and the root mean are obtained. A method of classifying a category, characterized in that the obtained mean and the mean square of the category are changed. 4. An average is determined for each category by learning, and a category having a minimum distance from the average of an input vector is set as a category to which the vector belongs, and the average is undecided for the category to which the vector input at learning belongs. If there is, the average in the category is determined according to the vector, and if the average is already determined for the category to which the vector input during learning belongs, the average obtained from the input vector and the predetermined average obtained A method for classifying a category, wherein the average for the category is changed.