JP2776265B2 - Unit recognition unit and learning type recognition judgment 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 learning type recognition judging device for judging and recognizing an object.

[0002]

2. Description of the Related Art A conventional learning-type recognition / judgment device is disclosed in, for example, DE Rummelhart et al.
resentations by back-propagating errors) ", Nature
(Nature) Vol.323 No.9 (1986). FIG. 10 shows a general configuration diagram of a conventional learning-type recognition / determination apparatus, where 201 and 202 are input terminals, 212 is an output terminal, 217 is a learning circuit, 218, 219, and 220 are multiple-input one-output circuits, 221 is an output layer, and 222 is a hidden layer. As shown in FIG. 10, the learning-type recognition / judgment device processes a signal input from input terminals 201 and 202 and outputs the processed signal from an output terminal 212 in a configuration in which multiple input-one output circuits are connected in a hierarchical manner. As described above, among the multi-input and one-output circuits connected in a hierarchical manner, a layer including a multi-input and one-output circuit that outputs an output signal is called an output layer,
A layer composed of other multi-input / one-output circuits is called a hidden layer. The hidden layer may be constituted by a multi-input one-output circuit forming one layer, or may be constituted by a multi-input one-output circuit forming a plurality of layers. FIG. 11 shows a configuration example of a conventional learning-type recognition / determination apparatus having one hidden layer. In FIG. 11, 201 and 202 are input terminals, 203, 204, 205, 206, 207 and 2
08 is a variable weight multiplier, 209, 210, 211 are adders having saturation input / output characteristics, 212 is an output terminal, 213 is a teacher signal generator,
214 is an error calculation unit, 215 is a steepest descent direction determination unit, 216 is a weight change unit, 217 is a learning circuit, 218, 219, and 220 are multiple-input one-output circuits, 221 is an output layer, and 222 is a hidden layer. As shown in FIG. 11, the multi-input one-output circuits 218, 219 and 220 are composed of a variable weight multiplier and an adder having a saturation input / output characteristic. That is, the output signal of the j-th multi-input one-output circuit is represented by (Equation 1).

[0003]

(Equation 1)

[0004] Here, y [i] is an output signal of the i-th multi-input one-output circuit of the preceding layer, and w [i, j]
Is a weight to be multiplied when the output signal of the i-th multi-input one-output circuit of the preceding layer is input to the j-th multi-input one-output circuit. fnc () is a function having a saturation characteristic and is represented by a sigmoid function (Equation 2) or the like.

[0005]

(Equation 2)

FIG. 12 shows a graph of the characteristic function of the adders 209, 210 and 211 having the saturation input / output characteristic, which is represented by fnc (). Learning in the learning-type recognition / judgment device means that the variable weight multipliers 203, 204, 205, 206, and 206 output a desired output signal (hereinafter referred to as a teacher signal) for an input signal.
Change the weight applied by 207 and 208. In order to obtain the change amount of the weight, first, an error E is obtained from the teacher signal and the output signal of the output layer using (Equation 3).

[0007]

(Equation 3)

Where y _p [j] is the output signal of the j-th multi-input one-output circuit of the output layer for the p-th input signal,
t _p [j] is the teacher signal for y _p [j], _p Σ is the sum of all teacher signals, Σ _j is the sum of all multi-input one-output circuits in the output layer, and vector w is the weight w [i, j ] As a component (hereinafter, the vector w is referred to as a weight vector). As shown in equation (2), the error E is represented by the sum of squares of the difference between the teacher signal and the output signal of the output layer, and is a function of the weight vector w. In the learning, the weight is changed to minimize the difference between the teacher signal and the actual output signal, that is, the error. The change amount of the weight is determined by (Equation 4).

[0009]

(Equation 4)

Here, ε is a positive constant called a learning parameter, α is a positive constant called momentum,
∂E / ∂w is the weight w of the error E represented by (Equation 3)
A vector whose component is the derivative by [i, j], which is called the steepest descent direction. Δw ′ is a vector expression of the weight change amount in the previous learning.

A learning circuit 217 of a conventional learning type recognition / judgment device.
Then, the teacher signal generator 213 generates a teacher signal (desired output signal) t _p [j] for the input signal. Error calculator 214
Is obtained from the teacher signal t _p [j] and the output signal y _p [j] of the output layer.
The error E represented by (Equation 3) is calculated. Error calculator 214
The difference signal t _p between the teacher signal and the output signal required for the change of the weight [j] - a y _p [j], and outputs the steepest descent direction decision unit 215. The steepest descent direction determining unit 215 determines the steepest descent direction of the error E in the weight space in which the weight is expressed by a vector based on the weights of the difference signal, output layer output signal, hidden layer output signal, input signal, and output layer. Ask. The steepest descent direction is obtained from (Equation 5).

[0012]

(Equation 5)

The right-hand side of (Equation 5) is a vector representation of the differentiation by the weight of the error E. The steepest descent direction determination unit 215 multiplies the steepest descent direction by the learning parameter, and
Output to The weight change unit 216 obtains the weight change amount according to the equation (3), and calculates the variable weight multipliers 203, 204, 205, 206, and 207.
And the weight to multiply by 208. As described above, the error is reduced by repeatedly obtaining the change amount of the weight by the steepest descent method, and when the error becomes sufficiently small, it is assumed that the output signal is sufficiently close to a desired value, and the learning is terminated. .

[0014]

In the conventional learning-type recognition / judgment apparatus described above, the network configuration cannot be changed from the initially defined or designed fixed state. The learning and recognition abilities are also determined by the initial design, and cannot be adaptively handled according to the input data. In addition, no network design method has been established, and the actual design has to rely on trial and error based on experience and feeling.

The present invention is intended to solve such a problem of the prior art. The present invention not only performs learning for changing the weight of input data, but also adaptively adapts the network according to the input signal. A unit recognition unit capable of automatically changing, constructing, and self-organizing the structure of a task, and suppressing unnecessary changes in the structure of the organization by using existing knowledge for the object of judgment recognition. And a learning-type recognition / judgment device.

[0016]

In order to achieve the above object, the present invention comprises a signal input unit, a quantizer for performing quantization in accordance with an input signal from the signal input unit, and a path selecting unit. A unit storing the unit recognition unit; a structure storage unit configured to store a quantization constant; an internal state storage unit configured to store an internal state of the unit unit; and a structure of the unit recognition unit based on the stored internal state. A duplicating unit having a structure and a function stored in a storing unit and creating a duplicate of the unit recognition unit; and a dividing control unit for controlling two division points of a quantization range of the quantizer of the duplicating unit. Are provided to form a unit recognition unit.

Further, a plurality of the unit recognition units are combined in a multi-layered hierarchical network to constitute a learning type recognition / judgment device.

[0018]

When various characteristic data of an object is inputted to the signal input section of the unit recognition unit of each layer of the apparatus constructed as described above, it is stored in the internal state storage means according to the input data. An internal state changes. When this internal state reaches a certain state, the duplication unit creates a copy of the existing unit recognition unit at an appropriate location based on the network structure and division control unit stored in the internal state storage unit. In addition, a network automatically adapted to input data can be efficiently formed in a self-organizing manner.

[0019]

【Example】

(Embodiment 1) Hereinafter, a first embodiment of a unit recognition unit of the present invention will be described with reference to the drawings.

In FIG. 1, reference numeral 8 denotes a unit. 1
Is a signal input unit for inputting feature data to be recognized, which is input via the signal input terminal 1a, to the quantizer 2. The quantizer 2 quantizes the input feature data, and inputs the quantized value to the path selection unit 3. 3a is the route input terminal, 3
b1 and 3b2 are path output terminals for connecting these terminals to each other when the unit recognition unit 9 is combined to form a network. The path selection unit 3 includes a path input terminal 3a based on the value input from the quantizer 2,
It is configured to change the way of connection with the path output terminal 3b1 or 3b2. The structure storage means 4 stores the quantization range of the quantizer, the number of quantizations, the number of path input terminals and the number of path output terminals of the path selection unit. The internal state storage means 5 stores the average, variance, and total number of input signals so far as the internal state of the unit recognition unit 9. The copy creation means 6 is composed of the internal state storage means 5
When the internal state stored in the storage unit reaches a certain value, the quantization range of the unit recognition unit stored in the structure storage unit 4 is divided into two, and a copy of the unit recognition unit 9 is created. It is. The division control means 7 is configured to control a method of dividing the quantization range into two by the copy creation means 6.

The division control means 7 shown in FIGS. 1 to 4 is specifically constituted by the embodiment shown in FIG. Division control means 7 in FIG.
Is a midpoint calculator 7a for calculating the midpoint of the quantization range of the unit recognition unit 9 to be the target of the duplication unit 6, and a division range storage unit 7c storing the upper and lower limits of the divisible range of the quantization range.
1, a comparator 7d for comparing the result of the midpoint computing unit with the upper and lower limits of the divisible range, and the comparator determines that the midpoint obtained by the midpoint computing unit is between the upper and lower limits of the dividable range. , The division point output unit 7b that outputs the intermediate point as two division points of the quantization range to the copy creation means.

The operation of the unit recognition unit configured as described above will be described below. In order to actually make a copy, it is necessary to prepare an empty unit recognition unit that is not used at all and has no connection with other unit recognition units. From the signal input terminal 1a of the signal input unit 1 of the unit recognition unit serving as the basis of the duplication, a signal having a value of, for example, 1 to 10 is input, and the quantization range of the quantizer is 1 to 10.
Is set to The internal state storage means 5 calculates and stores the average, variance, and number of inputs of the sequentially input data signal each time the data signal is input. Then, when the product of the number of inputs and the variance exceeds a certain value, the midpoint calculator calculates 1 to 1
Find the midpoint 5 of the quantization range up to 0. The upper limit of the divisible range stored in the divided range storage means is, for example, 6 or more,
If the lower limit is 4 or less, the division point output unit outputs the middle point 5 as a division point in the range of quantization to the copy creation means. The dividable range is used to suppress the excessive duplication of the unit recognition unit by using the existing knowledge of the input data about the approximate range of the input data. It can be set freely by the user of the type recognition determination device. When the division point in the range of quantization is output from the division point output unit, the replica creation unit outputs the quantization range of the quantizer stored in the structure storage unit, the number of quantizations, and the path selection unit. With reference to the number of path input terminals and the number of path output terminals, the information of the unit recognition unit is copied to an empty unit recognition unit that is not used, and the information including the connection with other unit recognition units is included. Create an exact duplicate of. The range of each quantization at the time of duplication is based on the instruction from the division point output unit,
Divide the quantization range from ~ 10 to the middle point 5 as a boundary, the quantization range of the original unit recognition unit's quantizer to 1 ~ 5, and the quantization of the duplicate unit recognition unit's quantizer. Range from 6 to 10
, And a duplicate is created so that the function is shared between the original unit and the duplicate unit.

The above-described unit recognition unit of the present invention can be configured as follows in addition to the above. FIG. 2 shows a configuration of the route selection unit 3 of FIG. 1 including a route input unit 3a having one route input terminal 3a1, a route output unit 3b having two route output terminals 3b1 and 3b2, and a switch 3c. FIG. The switch 3c, based on the value input from the quantizer 2, switches the way of connection between the path input terminal 3a1 of the path input unit 3a and the path output terminal 3b1 or 3b2 of the path output unit 3b, FIG. 3 is a block diagram showing a configuration of the route selection unit 3 of FIG. 1 with a route input unit 3a having one route input terminal 3a1.
And a path force unit 3b having two path output terminals 3b1 and 3b2
FIG. 9 shows an embodiment configured with a path load section 3c. The loads 3c1 and 3c2 are weights to be added to the route output signals output to the route output terminals 3b1 and 3b2 of the route output unit 3b, and the loader 3c0 calculates these loads according to the value output from the quantizer 2. Change. The loads 3c1 and 3c2 weight the route signal input from the route input unit, and output the route output unit 3b
Outputs the weighted path signal to path output terminals 3b1 and 3b.
Output to b2. FIG. 4 shows a configuration in which the path input unit 3a of FIG. 1 is configured by an adder 3a0 that adds input signals from a plurality of path input terminals, and the path output unit 3b is a threshold that performs threshold processing on the path signal. This shows an embodiment constituted by the value processor 3b0. The adder 3a0 uses 8 from 3a1 to 3a8.
Route signals input from the route input terminals
Input to the route selector 3c. The path selector 3c includes the quantizer 2
Determines how to output the added path signal to the path output terminal in accordance with the output value of the path signal. Route selector
3c has already been described with reference to FIGS. 2 and 3, and any of the configurations of FIGS. 2 and 3 can be used as the route selector of this embodiment. In this case, the route output unit has only one route output terminal, and therefore, the route selector 3c
Simply changes the weight for weighting the path signal when the configuration of FIG. 3 is used.

Further, the division control means may be configured as shown in FIGS. FIG.
A middle point calculator 7a for finding the middle point of the quantization range of the unit recognition unit 9 to be the target of the duplication unit 6, a division point storage unit 7c2 storing the divisible points of the quantization range, and A distance calculator 7f for calculating the distance between the result of the point calculator and the division point storage means, a minimum detector 7e for calculating the minimum value of each distance obtained by the distance calculator, and a dividable having a minimum value. It is constituted by a division point output unit 7b which outputs the points to the duplication creating means 6 as two division points within the range of quantization.

FIG. 7 shows a division point output unit shown in FIG. 6 as a permissible error storage means 7b1 for storing a permissible division error, and a comparator for comparing the minimum value obtained by the minimum value detector with the permissible division error.
7b shows an embodiment comprising a split point output unit 7b2 which outputs a dividable point having a minimum value to the duplication means as two split points in a quantization range when the minimum value is smaller than the division allowable error. It is a thing.

The operation of the unit recognition unit configured as described above will be described below with reference to FIG. FIG.
Alternatively, the operation of the unit recognition unit of FIG. 7 is the same as that of FIG. 5 except for the division control means. In the same manner as in FIG. 5, when the input data is sequentially input and the product of the number of inputs and the variance stored in the internal state storage means exceeds a certain value, the midpoint calculator calculates the quantization of 1 to 10. Find the midpoint 5 of the range The distance calculator determines the distance between the result of the midpoint calculator and the dividable point stored in the division point storage means. Usually, there are a plurality of dividable points. The minimum value detector obtains the minimum value of each distance obtained by the distance calculator and the dividable point when the distance becomes the minimum value. The comparator compares the divisional tolerance stored in the tolerance storage means with the minimum value obtained by the operation distance unit, and when the minimum value is smaller than the divisional tolerance, the division point output unit outputs the minimum value detector. Is output to the duplicating means as two division points in the quantization range. Fig. 5
As in the case of the dividable range described with reference to the above, this is for suppressing the excessive duplication of the unit recognition unit by using the existing knowledge of the input data. Can be set freely by the user. When the division point of the range of quantization is output from the division output unit, the copy creation unit outputs the quantization range of the quantizer stored in the structure storage unit, the number of quantization, and the Referring to the number of path input terminals and the number of path output terminals, copy the information of this unit recognition unit to an empty unit recognition unit that is not being used, and connect it to itself including connection with other unit recognition units. Make an exact duplicate. For example, when the two division points of the quantization range output from the division point output unit are 3, the quantization range of the original unit unit quantizer is set to 1 to 3, and the quantization unit Set the quantizer quantization range to 4-10, and share the functions between the original unit and the duplicate unit.
Make a duplicate.

In the above embodiment, when the path input terminal is 1
Although the description has been given of the case where a plurality of path input terminals are provided, the same operation can be performed when a plurality of path input terminals are provided.

(Embodiment 2) Next, an embodiment of a learning type recognition judging device according to the second invention will be described with reference to the drawings.

FIG. 8 shows a network in which unit recognition units according to the present invention are interconnected in a multi-layer hierarchy. The unit recognition units n11 to n12 , n21 to n24 , and n31 to n38 constituting the first layer, the second layer, and the third layer use, for example, the unit recognition unit shown in FIG. As described above, the path selection unit 3 is connected to one path input terminal 3
path input unit 3a having a1 and two path output terminals 3b1 and 3b
It is composed of a route output unit 3b having b2 and a route selection unit. Also, the unit recognition unit constituting the fourth layer
n41 to n43 use, for example, the unit recognition unit shown in FIG. 4, and as described above, the route input unit 3a
Adder 3a that adds input signals from multiple path input terminals
0, and the path output unit 3b is configured by a threshold value processor 3b0 that performs threshold processing on the path signal. The learning type recognition judging device shown in FIG. 8 classifies and recognizes three types based on three types of feature data composed of two pieces.

The operation of the learning-type recognition / judgment device configured as described above will be described below.

First, the learning operation will be described. First, "1" is given as a path signal to the path input terminals of the unit recognition units n11 and n12 of the first layer. Also, a first series of feature data of the object to be recognized is input to a signal input terminal 1a to the quantizer of these units. (In the case of this figure, two first feature data are input to two unit recognition units, respectively.) These first feature data are input to n1
The quantizers 1 and n12 quantize, and based on the quantized values, the paths p11 and p12 are selected by the path selection unit shown in FIG. 5, and the unit recognition units n22 and n23 in the second layer.
The path signal "1" is sent to the path input terminal of the. To a signal input terminal to the quantizer of these units, a second series of feature data of the object to be recognized is input. (In the case of this figure, two second feature data are input to two unit recognition units n22 and n23, respectively.) These second feature data are quantized into n22 and n23. The devices are quantized, and paths p21 and p22 are selected by the path selection unit shown in FIG. 5 based on the quantized values, and are routed to the path input terminals of the unit recognition units n34 and n36 on the fourth layer. Signal "1" is sent. To the signal input terminals to the quantizers of these units, a teacher input signal indicating which of the three items the recognition target belongs to, that is, a signal indicating which of n41 to n43 is to be selected, is input. . (In the case of this figure, the teacher input signal is input to the two unit recognition units n34 and n36.) The teacher input signal is quantized by the quantizers n34 and n36,
Based on the quantized values, the paths p31 and p32 are
Is selected by the route selection unit shown in FIG.

Next, the operation of the entire apparatus when each unit recognition unit creates a copy will be described.

As described above, in order to actually make a copy, it is necessary to prepare an unused unit recognition unit which has no connection with other unit recognition units at all and is not used. When the product of the number of times of input and the variance of the signal input to the signal input unit of the unit recognition unit serving as the source of the duplication exceeds a certain value, the midpoint computing unit in the division control means shown in FIG.
Find the midpoint of the quantization range. If the obtained midpoint is between the upper limit and the lower limit of the dividable range stored in the division range storage means, the division point output unit outputs the midpoint to the copy creation means as the division point of the quantization range. I do. The divisible range is
By using existing knowledge of the input data about the approximate range of the input data, this is for suppressing excessive duplication of the unit recognition unit. It can be set freely by the user. When the division point in the quantization range is output from the division point output unit, the copy creation unit operates the structure storage unit. When the unit recognition unit in the learning type recognition judging device creates a copy, the fourth unit shown in FIG.
The unit recognition unit of each layer up to the layer needs to be copied including all of the unit recognition units connected to the lower layer of each unit. FIG. 9 shows a state of duplication including all of the unit recognition units connected to the lower layer. The copy creation detecting unit S detects all the unit recognition units that have caused the copy operation, and issues an instruction to make a copy including all the unit recognition units connected to the lower layer of each unit for which a copy is made. For example, as shown in FIG.
The unit recognition unit 711 and the unit recognition unit 712 linked to the lower layer are copied, and the unit recognition unit 72 of the copy and the unit recognition units 721 and 722 of the lower layer are created. When such a duplicating means is used, if an unnecessary unit exists in a lower layer of each unit to be duplicated, the unnecessary unit is also duplicated. Therefore, as learning is repeated, the network becomes larger than necessary. . Since the unit recognition unit according to the present invention can suppress unnecessary duplication of the unit recognition unit and an excessive scale network by the division control means described above, the unit learning unit learns the network for the input data. Besides self-organization, self-organization can be achieved by automatically and efficiently changing the network structure adaptively and efficiently according to input signals, based on knowledge of input data. Further, if the division range within the division control means is set to be wide, the controllability for division is reduced, so that generalization of new data input that has not been learned can be improved.

Next, the recognition operation will be described. In the recognition process, the internal state of the unit recognition unit stored in the internal state storage means does not change, and therefore, the copy creation means does not operate.

Unit recognition unit from first layer to second layer
n11, n12, n22, and n23, in exactly the same way as the learning operation, the quantizer quantizes the input feature data, switches based on the quantized data, and sequentially selects paths p11, p12, p21, and p22. I do. In the case of the recognition operation, the teacher input signal is not input to the signal input terminals of the unit recognition units n34 and n36 in the third layer. Therefore, the state of the switches at the time of learning is held, and the paths p31 and p31 are changed according to the states of these switches.
32 is selected, and the path signal “1” is sent to the path input terminal of the unit recognition unit n42 of the fourth layer. The adder in the path input unit of this unit adds the path signals input through the paths p31 and p32. The signal "1" is input to the signal input terminal of the signal input unit, the quantizer quantizes the signal, the path selector sets the path output to a possible state, and when the signal "0" is input The path selector switches to a state in which the path is not output.) The added path input signal is sent to the path output unit. The path output device performs threshold processing on this signal and outputs the signal to a path output terminal. Therefore, if the value of the added signal is larger than a certain threshold value, an output is made. In this way, only the route selection and addition are performed, and the recognition target is input based on the input recognition target feature data. Since the classification and recognition judgment of the object can be performed, the processing speed is extremely high. Note that a sigmoid function, a step function, or the like can be used as a function for performing threshold processing.

[0036]

As described above, the learning type recognition / judgment device of the present invention forms a unit unit by at least the signal input unit and the signal output unit, and stores the structure and function of the unit unit. Means, an internal state storage means for storing the internal state of the unit unit, and a structure and function stored in the structure storage means of the unit recognition unit in accordance with the state of the stored internal state. A duplication unit for duplicating a unit, a unit recognition unit provided with division control means for controlling division of a range of quantization when duplicating a unit recognition unit by the duplication unit, and the unit recognition unit having a multilayer structure. The internal state of each unit recognition unit changes in accordance with input data, and In operation, when making a copy of each unit recognition unit, the division control means performs control based on the existing knowledge of the input data, so that the network structure is automatically adaptively adjusted according to the input signal. It is possible to self-organize by changing the range, by taking a wider dividable range stored in the division range storage means, or by making the dividable points stored in the division point storage means finer, The generalization of the new data input that has not been learned can be improved. Also, various characteristic data of the object is input to the signal input section of each unit recognition unit constituting the multi-layer hierarchical network, and the connection path between the unit recognition units is determined according to the output of the quantizer. In the switching and the lowermost layer, it is only necessary to determine the selection path to the lowermost layer based on the input of the teacher signal, so that the learning process can be performed at a very high speed. Also, in the recognition process, various characteristic data of the object are input to the signal input section of each unit recognition unit constituting the multi-layer hierarchical network, and the unit is changed according to the output of the quantizer. The recognition result is obtained only by switching the connection path between the recognition units and in the preceding layer of the lowest layer, determining the selection path to the lowest layer based on the connection path set in the learning process, Therefore, the recognition processing can be performed very quickly based on the learning result.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of a unit recognition unit according to the present invention.

FIG. 2 is a diagram showing a route selection unit in FIG.
Block diagram showing an example composed of a path output unit and a switch

3 is a block diagram showing an example in which the switch in FIG. 2 is replaced with a path load unit.

4 is a block diagram showing an example in which the route input unit of the route selection unit in FIG. 1 is configured by an adder and the route output unit is configured by a threshold value processor;

FIG. 5 is a block diagram showing an example in which the division control unit in FIG. 1 is composed of a division point output unit, a comparator, a division range storage unit, and a midpoint computing unit.

FIG. 6 is a block diagram showing an example in which the division control unit in FIG. 1 is composed of a division point output unit, a minimum value detector, a distance calculator, a division range storage unit, and a middle point calculator.

FIG. 7 is a block diagram showing an example in which the division point output unit of FIG. 6 is composed of an allowable error storage unit, a comparator, and a division point output unit.

FIG. 8 is a block diagram showing an embodiment of a learning type recognition / judgment device according to the present invention.

FIG. 9 is a diagram showing a copying operation including the network.

FIG. 10 is a block diagram showing a general configuration of a conventional learning-type recognition / judgment device.

FIG. 11 is a block diagram of a conventional learning-type recognition / judgment device.

FIG. 12 is a characteristic diagram of a characteristic function of the adder in FIG. 11;

[Explanation of symbols]

 1 Signal input unit 1a Signal input terminal 2 Quantizer 3 Route selection unit 3a Route input terminal 3b1 Route output terminal 3b2 Route output terminal 4 Structure storage unit 5 Internal state storage unit 6 Duplication creation unit 7 Division control unit 7a Midpoint computing unit 7b Division point output unit 7b1 Tolerance storage unit 7b2 Comparator 7b3 Division point output unit 7c1 Division range storage unit 7c2 Division point storage unit 7d Comparator 7e Minimum value detector 7f Distance calculator 8 Unit unit 9 Unit recognition unit

────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Susumu Maruno 1006 Kazuma Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (56) References JP-A-4-142658 (JP, A) (58) Field (Int.Cl. ⁶ , DB name) G06F 15/18 JICST file (JOIS)

Claims (8)

(57) [Claims] A signal input unit, a quantizer that performs quantization in accordance with an input signal from the signal input unit, a single or multiple path input terminals, at least one path output terminal, A unit unit is constituted by a path selection unit for selecting a path according to the output of the quantizer, and the unit unit and the quantization range and the number of quantizations of the quantizer are stored as quantization constants. Structure storing means for storing the average, variance, and input count of the input signal to the unit unit as an internal state; and dividing the quantization range into two based on the stored internal state. The unit structure having the entire structure of the unit unit before division and each quantization range after division is created .
And a division control means for controlling two division points of the quantization range by the duplication means , wherein the division control means calculates a middle point of the quantization range.
And the upper and lower limits of the dividable range of quantization
Division range storage means for storing a value, and the division possible range
Comparison of comparing the upper limit value, lower limit value and the calculated midpoint of
Container, and the midpoint is an upper limit and a lower limit of the dividable range.
, The calculated midpoint is defined as the quantization range.
Divide points to be output to the duplication creation means as two divided points
A unit recognition unit comprising a power unit. 2. A division control means comprising a midpoint arithmetic unit and
A division point notation that stores the dividable point of quantization in one place
Storage means and the distance between the midpoint of the quantization range and the dividable point.
A distance calculator for calculating the distance, and each of the distances calculated by the distance calculator.
A minimum value detector for determining a minimum value of the distance, and the minimum value search
The dividable point having the minimum value obtained by the
Divided point output to the duplication means as two divided points in the range
2. The unit recognition according to claim 1, wherein said unit is replaced with a force part.
Sense unit. 3. A division point output unit includes: an error storage unit that stores a division allowable error; a comparator that compares a minimum value obtained by a minimum value detector with the division allowable error; When the division point is smaller than the division allowable error, the division point having the minimum value obtained by the minimum value detector is configured as a division point output unit that outputs to a duplication unit as two division points in a quantization range, The unit recognition unit according to claim 2 . 4. A route input unit having one or more route input terminals, a route output unit having one or more route output terminals, a route input terminal of the route input unit, and a route of the route output unit. The unit recognition unit according to any one of claims 1 to 3 , wherein a path selector is configured by a switch that switches connection with an output terminal according to an output of the quantizer. 5. A route input unit having one or more route input terminals, a route output unit having one or more route output terminals, a route input terminal of the route input unit, and a route of the route output unit. The unit recognition unit according to claim 4 , wherein the path selector includes a loader that changes the strength of connection with the output terminal according to the output of the quantizer. 6. A path input section comprising an adder for adding input signals from at least a plurality of path input terminals,
An output signal of at least the adder is configured to route output section by thresholding device for thresholding, claim 4
Or the unit recognition unit according to claim 5 . 7. A multi-layered structure in which a plurality of unit recognition units according to any one of claims 1 to 6 are combined in a multi-layer hierarchy, and the lowermost layer is constituted by the unit recognition units according to claim 6. The unit is configured to input a teacher signal to the signal input unit of the unit recognition unit located in the previous layer at the bottom of the structure.
Other studies習型recognition judgment apparatus. 8. A copy detection unit for detecting a copy creation operation of all unit recognition units in a multi-layered hierarchical structure, wherein a lower layer connected to the unit recognition unit is used when a copy of the unit recognition unit is created. 8. The learning-type recognition / determination apparatus according to claim 7 , wherein a copy is created including all of the unit units.