JP3052520B2 - Pattern classification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern classification apparatus to which a neural network used for pattern classification such as voice recognition is applied.

[0002]

2. Description of the Related Art In recent years, a pattern classification apparatus using a neural network has been widely applied to speech recognition and the like. A pattern classification device to which a neural network is applied is roughly classified into two types. One is the literature (PDP model, DE Ramelhart and two others, translated by Shunichi Amari, 198
9 years), and another is called a multi-layer perceptron. Another is a literature (self-organization and associative memory, Se
lf-Organization and Associative Memory, T.Kohonen
Author, Springer-Verlag, 1987). The present invention relates to a pattern classification device to which a neural network is applied, and improves learning vector quantization.

[0003] A conventional pattern classification apparatus will be described below. FIG. 3 is a block diagram of a conventional pattern classification device configured based on learning vector quantization.

In FIG. 3, reference numerals 1-1 to 1-4 denote reference vector circuits for calculating a distance between a reference vector stored therein and an input signal from an input terminal 100. Reference numeral 2 denotes each reference vector circuit. A comparison circuit that outputs a minimum distance number signal of reference vector circuits 1-1 to 1-4 that receives distance signals from 1 to 1-4 and provides a minimum distance value, and 4 is a category corresponding to each reference vector circuit. A readout circuit 3 which receives the minimum distance number signal from the comparison circuit 2, reads out the category specified by the minimum distance number signal from the category storage circuit 4, and outputs it as a category output; 5 is a read circuit 3
Is a judgment circuit which compares the category output from the teaching signal from the teacher signal from the teacher signal input terminal 102 and outputs a match / mismatch as a judgment output, and 6 is the minimum distance number signal from the comparison circuit 2 and the judgment from the judgment circuit 5 A learning signal generation circuit that receives a circuit signal and generates a learning signal that corrects the reference vector stored inside the reference vector circuits 1-1 to 1-4 using a given algorithm.

[0005] The operation of the pattern classification device configured as described above will be described below with reference to FIG. First, the initial learning of this type of pattern classification device will be described, and then the normal learning will be described.

[0006] The initial learning may be considered equivalent to the initial setting of the pattern classification device. That is, when the number of reference vector circuits of the pattern classification device is N, the initial state of the pattern classification device is set with the first N inputs. Specifically, the n-th (n
.Ltoreq.N) is a reference vector stored inside the n-th reference vector circuit 1-n. Category. The initial learning is performed for the first N data. Note that a write circuit that performs this procedure is not shown in FIG. 3 because it only complicates the drawing.

In normal learning, first, an input signal X is input from an input terminal 100 and all the reference vector circuits 1-1 to 1-1 are input.
The input signal X is sent to 1-4. Reference vector circuit 1-
n is the input signal X and the internally stored reference vector

[0008]

[Outside 1]

Distance from

[0010]

(Equation 1)

Is calculated and output. The comparison circuit 2 receives the distance signal D (n) from each of the reference vector circuits 1-n, and outputs a minimum distance number signal n _min that gives the number of the reference vector that gives the minimum value. The readout circuit 3 receives the minimum distance number signal n _min from the comparison circuit 2 and reads out and outputs the category signal C (min) of the reference vector circuit 1-n _min corresponding to the minimum distance number signal n _min from the category storage circuit 4. I do. The category signal C (min) is output from the output terminal 101 and sent to the determination circuit 5 at the same time. The determination circuit 5 outputs the category signal C
(Min) from the teacher signal input terminal 102 to the teacher signal T
When T = C (min), the judgment signal J = (match) When T ≠ C (min), the judgment signal J = (mismatch) is output. The learning signal generation circuit 6
The minimum distance number signal n _min from the comparison circuit 2 is determined by the determination circuit 5
From the reference vector circuit 1-n _min corresponding to the minimum distance number signal n _min

[0012]

[Outside 2]

[0013]

[0014]

[Outside 3]

And outputs a learning signal L that changes
By repeating this procedure, the reference vector circuit 1
Reference vector in -1 to 1-4

[0016]

[Outside 4]

Will reflect the input distribution, and the output C (min) of the pattern classification device will give an appropriate categorization.

[0018]

However, in the above-mentioned conventional configuration, since the distribution of the reference vector becomes proportional to the input frequency, the normally distributed data which often appears is stored in the reference vector circuit. The distribution of the obtained reference vector also has a normal distribution. In general, what is important in pattern classification is the boundary between different categories, and the center of the category is not always important. Therefore, when an attempt is made to configure a pattern classification device with the same number of reference vector circuits, it is desirable that more reference vectors be distributed on the boundary surface than on the center of the distribution. On the other hand, in the conventional pattern classification device, since many reference vectors are gathered at the center of the distribution of each category, the pattern classification device has an inefficient internal representation. There was a problem that storage capacity was required.

The present invention solves the above-mentioned conventional problems. By reducing the distribution density of reference vectors in places where the value of input is low in terms of pattern classification and high in input frequency, the accuracy of the storage is small and the accuracy is small. An object of the present invention is to provide a pattern classification device capable of performing high pattern classification.

[0020]

In order to achieve the above object, the pattern classification apparatus of the present invention is provided with a means for correcting a distance from a reference vector held therein according to an input frequency, and the corrected distance is minimized. The pattern classification is performed by correcting the reference vector according to the match / mismatch between the category corresponding to the reference vector and the teacher signal.

Therefore, a reference vector circuit for calculating a distance between an internally stored reference vector and an input signal, and an input frequency correction circuit for correcting a distance signal from the reference vector circuit according to input frequency information held therein A comparison circuit that receives a correction distance signal from each input frequency correction circuit and outputs a minimum correction distance number signal of a reference vector circuit that provides a minimum correction distance value, and a category storage that stores a category corresponding to each reference vector circuit A circuit, a read circuit for receiving a minimum correction distance number signal from the comparison circuit, reading a category specified by the minimum correction distance number signal from the category storage circuit and outputting it as a category output, a category output from the read circuit and a teacher signal And a judgment circuit that outputs a match / mismatch as a judgment output and a minimum correction from the comparison circuit Generation of a learning signal for receiving a separation number signal and a judgment circuit signal from a judgment circuit, and generating a learning signal for correcting the reference vector stored inside the reference vector circuit and the input frequency information of the input frequency correction circuit by a given algorithm. And a circuit.

[0022]

With this configuration, the distance signal from the reference vector circuit is converted into a corrected distance signal reflecting the input frequency by the input frequency correction circuit, and is classified into categories according to its magnitude. It is possible to realize a pattern classification device that efficiently distributes reference vectors so that many reference vectors are not distributed.

[0023]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS First, the principle of the present invention will be described with reference to the drawings. FIG. 2 is a conceptual diagram for explaining the principle of the present invention when the input is two-dimensional.

Referring to FIG.

[0025]

[Outside 5]

Is a two-dimensional reference vector,

[0027]

[Outside 6]

Is a three-dimensional reference vector, R1, R2
Are the input X and the two-dimensional reference vector, respectively.

[0029]

[Outside 7]

Where S1 and S2 are the input X and the three-dimensional reference vector, respectively.

[0031]

[Outside 8]

Is the distance between In the conventional pattern classification device, the dimension of the reference vector is configured to match the dimension of the input. Therefore, in FIG.

[0033]

(Equation 2)

Therefore,

[0035]

[Outside 9]

Is the reference vector closest to the input.
On the other hand, the three-dimensional reference vector

[0037]

[Outside 10]

Considering the following,

[0039]

(Equation 3)

Therefore,

[0041]

[Outside 11]

Is the vector closest to the input. As described above, by setting the space of the reference vector to be larger than the dimension of the input vector, the distance can be changed. Now a three-dimensional reference vector

[0043]

[Outside 12]

The z components P1z and P2z of the input frequency K1,
These are given by (Equation 4) and (Equation 5) by K2.

[0045]

(Equation 4)

[0046]

(Equation 5)

[0047] However, P1z> 0, P2z> 0 , K 0 is the initial value that represents the input frequency 0. Therefore, the vector

[0048]

[Outside 13]

The input frequency of the vector is

[0050]

[Outside 14]

If the input frequency is higher than

[0052]

(Equation 6)

Because

[0054]

(Equation 7)

Is as follows. At this time, the square of the distance S1 is

[0056]

(Equation 8)

Is as follows. Similarly, the square of the distance S2 is also given. Now, if the input space is divided by which of these two vectors is closer, the boundary line is

[0058]

(Equation 9)

That is,

[0060]

(Equation 10)

[0061]

[Equation 11]

Is given by The reference vector space is different from the two-dimensional case is the point that has been added (P2z ² -P1z ^2). Input is a vector from (Equation 7)

[0063]

[Outside 15]

The vector compared with the surrounding

[0065]

[Outside 16]

When there are many around, the added term becomes a negative value and the vector

[0067]

[Outside 17]

It can be seen that the range of influence is widened. Also, as an important property, the boundary by this method may be given by a linear equation.

From the above, it has become clear that the power range of the reference vector can be controlled by setting the dimension of the reference vector larger than the input dimension and reflecting the input distribution on the dimension. In particular, by providing components corresponding to the extended dimension as shown in (Equation 4) and (Equation 5), it is possible to widen the power range of the reference vector at a place where the input frequency is high. .

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the embodiment of the present invention.

In FIG. 1, 100 is an input terminal, 101
Is an output terminal, 102 is a teacher signal input terminal, 1-1 to 1-
Reference numeral 4 denotes a reference vector circuit for calculating a distance between a reference vector stored therein and an input signal, and reference numerals 7-1 to 7-4 correct a distance signal from the reference vector circuit according to input frequency information held therein. An input frequency correction circuit 2 receives a correction distance signal from each input frequency correction circuit, and a comparison circuit that outputs a minimum correction distance number signal of a reference vector circuit that gives a minimum correction distance value, and 4 corresponds to each reference vector circuit. A category storage circuit for storing a category, a read circuit for receiving a minimum correction distance number signal from the comparison circuit, reading a category specified by the minimum correction distance signal from the category storage circuit, and outputting the category as a category output, A decision circuit 6 compares the category output from the circuit with the teacher signal and outputs a match / mismatch as a decision output. A learning signal for receiving a correction distance number signal and a judgment circuit signal from the judgment circuit, and generating a learning signal for correcting the reference vector stored inside the reference vector circuit and the input frequency information of the input frequency correction circuit with the given algorithm. It is a generating circuit.

The operation of the pattern classification device having the above configuration will be described below with reference to FIG. First, the initial learning of this type of pattern classification device will be described, and then the normal learning will be described.

The initial learning can be considered to be equivalent to the initial setting of the pattern classification device. That is, assuming that the number of reference vector circuits included in the pattern classification device is N, the initial state of the pattern classification device is set with the first N inputs. Specifically, the n-th (n
≤N) and the teacher signal giving its category correspond to the reference vector stored inside the n-th reference vector circuit 1-n, and the teacher signal corresponds to the n-th reference vector circuit of the category storage circuit 4. The input frequency information of the input frequency correction circuit 7-n is set to an appropriate positive initial value. The initial learning is performed for the first N data. Note that the write circuit for performing this procedure is not shown in FIG. 1 because it only complicates the drawing.

In the normal learning, first, an input signal X is input from the input terminal 100 and all the reference vector circuits 1-1 to 1-1 are input.
The input signal X is sent to 1-4. Reference vector circuit 1-
n is the input signal X and the internally stored reference vector

[0075]

[Outside 18]

Distance from

[0077]

(Equation 12)

Is calculated and output as a distance signal D (n). The input frequency correction circuit 7-n adds the non-negative input frequency information H (n) held therein to the distance signal D (n) from the reference vector circuit 1-n to generate a corrected distance signal.

[0079]

(Equation 13)

Is output. The comparison circuit 2 receives the correction distance signal HD (n) from each input frequency correction circuit 7-n,
A minimum correction distance number signal n _min for giving the number of the reference vector giving the minimum value is output. The read circuit 3
Upon receiving the minimum correction distance number signal n _min from the comparison circuit 2,
The category signal C of the reference vector circuit 1-n _min corresponding to the minimum correction distance number signal n _min from the category storage circuit 4.
(Min) is read out and output. Category signal C (mi
n) is output from the output terminal 101 and sent to the determination circuit 5 at the same time. The determination circuit 5 receives the category signal C (min) from the readout circuit 3 and the teacher signal T from the teacher signal input terminal 102, respectively. If T = C (min), the determination signal J = (coincidence) T ≠ C In the case of (min), the judgment signal J = (mismatch) is output. The learning signal generation circuit 6
The minimum correction distance number signal n _min from the comparator 2 receives respectively a determination signal J from the judgment circuit 5, reference vector minimum correction distance number signal n _min stored in the reference vector circuit 1-n _min, corresponding to

[0081]

[Outside 19]

Is

[0083]

[Outside 20]

The input frequency correction information H (n _min ) in the input frequency correction circuit 7-n _min is

[0085]

[Outside 21]

A learning signal L is output as follows. If the input frequency correction information H (n) in the input frequency correction circuit 7-n has already become 0, the learning signal L can be output only to the reference vector circuit 1-n.

By repeating this procedure, the reference vectors in the reference vector circuits 1-1 to 1-4 are obtained.

[0088]

[Outside 22]

The input frequency information H in the input frequency correction circuits 7-1 to 7-4 is appropriately distributed in performing the pattern classification, and the output C (min) of the pattern classification device is obtained.
Will give the appropriate categorization.

Even if the output from the comparison circuit 2 is directly input to the input frequency correction circuits 7-1 to 7-4, the subscript n is held inside the input frequency correction circuit 7-n.
Only when n = n _min

[0091]

[Outside 23]

The same effect can be obtained by configuring the input frequency correction circuit 7 as follows. Also,
Input frequency correction information H in the input frequency correction circuit 7-n _min
(N _min )

[0093]

[Outside 24]

[0094] Alternatively, Further, the algorithm for moving the reference vector may be other than the above.

[0095]

As described above, according to the present invention, the distance between the input signal calculated by the reference vector circuit and the reference vector is corrected in accordance with the input frequency information held therein, and the input for outputting the corrected distance signal is provided. A frequency correction circuit is provided, a category corresponding to a reference vector circuit that gives a minimum correction distance value is read out from each correction distance signal, and this category is compared with a teacher signal. Pattern classification is performed by correcting the vector and the input frequency information in the input frequency correction circuit. Therefore, the distance signal from the reference vector circuit is converted into a corrected distance signal reflecting the input frequency by the input frequency correction circuit, and is classified into categories according to its magnitude. It is possible to configure a pattern classification device in which reference vectors are efficiently distributed so as not to be distributed. As a result, it is possible to realize a pattern classification device that accurately forms a category boundary surface even in a small scale.

[Brief description of the drawings]

FIG. 1 is a block diagram of a pattern classification device according to an embodiment of the present invention.

FIG. 2 is a conceptual diagram for explaining the principle of the present invention when the input is two-dimensional;

FIG. 3 is a block diagram of a conventional pattern classification device.

[Explanation of symbols]

 1-1 to 1-4 Reference vector circuit 2 Comparison circuit 3 Readout circuit 4 Category storage circuit 5 Judgment circuit 6 Learning signal generation circuit 7-1 to 7-4 Input frequency correction circuit 100 Input terminal 101 Output terminal 102 Teacher signal input terminal

Continuation of the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G06T 7/00 G06F 15/18 560 G10L 15/10 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A pattern for calculating a distance from each reference vector to an input signal and correcting the reference vector based on a match / mismatch between a category corresponding to the reference vector giving the minimum distance and a teacher signal, thereby performing pattern classification. A pattern classification device, comprising: a unit that corrects at least a distance from a reference vector held inside according to an input frequency. 2. A reference vector circuit for calculating a distance between an internally stored reference vector and an input signal, and an input frequency correction circuit for correcting a distance signal from the reference vector circuit according to input frequency information held therein. A comparison circuit that outputs a minimum correction distance number signal of a reference vector circuit that receives a correction distance signal from each input frequency correction circuit and provides a minimum correction distance value, and a minimum correction distance number signal from the comparison circuit. A readout circuit for reading out a category corresponding to the reference vector circuit and outputting it as a category output;
A judgment circuit that compares the category output from the readout circuit with the teacher signal and outputs a match / mismatch as a judgment output, and receives a minimum correction distance number signal from the comparison circuit and a judgment circuit signal from the judgment circuit to generate a reference vector circuit. A pattern classification device comprising: a reference signal stored therein and a learning signal generation circuit for generating a learning signal for correcting input frequency information of an input frequency correction circuit.