JPH06295292A - Learning recognizer - Google Patents

Abstract

(57) [Summary] [Purpose] To provide a learning and recognizing device capable of fast and efficient learning and additional learning. [Configuration] A route selection unit n includes a quantizer for quantizing an input signal, a first route input unit having a plurality of route input terminals, a route output unit having a plurality of route output terminals, and a plurality of routes. The learning recognition unit q has a path output integrator that replaces the output terminal with one path output terminal, and a path weighting section that selects a path between the path input section and the path means according to the quantized output. Is a teacher signal input section for inputting a teacher signal, a weight for weighting an input from a second path input section having a plurality of path input terminals, and an adder for adding outputs from the weights, And a learning device for determining the value of the weight based on the output of a threshold value processor for thresholding the added value and the teacher signal, and a plurality of branches by hierarchically combining a plurality of route selection units. Structure the structure, and arrange the learning and recognition units above it. Which was formed by. The route output integrator integrates / reduces redundant signal paths to improve learning efficiency.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning and recognizing device for acquiring the ability of recognizing and judging an object by learning according to various characteristic data of the object.

[0002]

2. Description of the Related Art Conventionally, as a learning and recognition device for acquiring recognition ability by learning, Nature 323 (1986), pp. 533-
Page 536 (DERumelhart, GEHinton an
dR.J.Williams "Learning Representations by Back-
Propagating Errors ", Nature, vol.323.pp.533-536,
Oct. 9, 1986). FIG. 6 shows the configuration of the conventional learning and recognition device. In FIG. 6, 10 is the output signal calculation unit, and 20 is the output signal calculation unit 1 using the difference between the output of the output calculation unit 10 and the target output value during learning.
A joint weight updating unit that updates the value of the joint weight of 0, and includes a teacher signal generator 100, an error signal calculator 110, and a joint weight update amount calculator 120. This output signal calculation unit 10
As shown in FIG. 7, a hierarchical structure is formed, 30 is a multi-input / one-output signal processing unit, and 40 is an input unit of the output signal calculation unit. FIG. 8 specifically shows the configuration of the multi-input / single-output signal processing unit 30. In FIG. 8, 50 is a signal input section, 60 is a memory for storing weights for weighting a plurality of inputs from the input section 50, and 70 is a multiplication for obtaining the product of the coupling weight of the memory 60 and the input signal value from the input section 50. And 80 is an adder for obtaining the sum of the outputs of the multiplier 70.
Reference numeral 90 is a threshold value processor that limits the output of the adder 80 to a value within a certain range. The input / output characteristics of the threshold value processor 90 are shown in FIG. For example, the input / output characteristics of the threshold value processor 90 that limits the output to the range of (0, 1) can be expressed as in Expression (1).

F (I) = 1 / (1 + exp (−I + θ)) (1) where I is the input of the threshold value processor 90 and θ is the threshold value.

The operation of the conventional learning / recognition device configured as described above will be described below. When an input signal is input to the input unit 40 of the output signal calculation unit 10, each multi-input one-output signal processing unit 30 is stored in the memory 60 and the output of the lower multi-input one-output signal processing unit. The product of the combined weight is obtained by the multiplier 70, the sum of the outputs of the multipliers is obtained by the adder 80, and then the output value of the adder is thresholded by the threshold processor 90. ,
The value is output to the input unit of the upper layer multi-input single-output signal processing unit. That is, the i-th multi-input one-output signal processing unit 30
When the output value o _i of is expressed by a mathematical expression, it can be expressed as Expression (2).

O _i = f (Σ _j w _ij · o _j ) ... (2) where f is a function representing the input / output characteristic of the threshold value processor 90 shown in equation (1), and Σ _j Is the sum of j, and w _ij is the i-th multi-input / output signal processing unit 30 stored in the memory 60 and the j-th multi-input / output signal processing unit 3 in the lower layer.
The coupling weight with 0, o _j , represents the output value of the j-th multi-input one-output signal processing unit 30 in the lower layer.

For the signal input from the input unit 40 of the output signal calculation unit 10, the teacher signal generation unit 100 generates a desired output signal for the input signal as a teacher signal, and the error signal calculation unit 110 outputs the output signal. Signal calculator 10
The error between the actual output value output from the and the teacher signal is output. The error E is calculated as in equation (3).

E = {Σ _i (t _i − o _i ) ² } / 2 (3) where o _i is the output value of the i-th multi-input one-output signal processing unit 30 in the uppermost layer, t _i represents the target output value for the output. The connection weight updating unit 120 calculates the update amount Δw _ij of the connection weight stored in the memory 60 of the output signal calculation unit 10 based on the error E output from the error signal calculation unit 110, and updates the connection weight. Do. Update amount Δ
w _ij is _{calculated as} in equation (4).

Δw _ij = −ε · ∂E / ∂w _ij + α · Δw ′ _ij (4) Here, ε and α are positive constants used for learning, and Δ
w ′ _ij is the connection weight update amount in the previous learning.

As described above, by repeatedly changing the coupling load, the output of the output signal calculation unit 10 when a certain signal is input to the output signal calculation unit 10 gradually approaches the target value, and the error E Is reduced. The learning ends when the error E reaches a value that is sufficiently small for practical use. After the learning is completed, the learning recognition device can output a desired value for the input signal and perform the recognition operation.

[0010]

However, in the above-mentioned conventional configuration, all the connection weights need to be changed at the time of learning, so that the amount of calculation required for learning is large and the learning takes time. In addition, when additional learning is performed so as to obtain a desired output for an unknown input signal after completion of learning, it is necessary to perform new learning again using the previously learned input signal, which causes a problem that it takes time. Had.

In order to solve the problems of the conventional learning and recognizing device, it is an object of the present invention to provide a learning and recognizing device which can perform learning at high speed and efficiently and can perform additional learning for unknown input signals at high speed. To do.

[0012]

According to the present invention, for example, a path selection unit includes a signal input section, a quantizer for quantizing an input signal from the signal input section, and a single or a plurality of path input terminals. A path input section having, a path output section having a single or a plurality of path output terminals, a path output integrator for replacing a plurality of path output terminals with one path output terminal, and a quantization output of a quantizer A learning signal recognition unit, a learning signal recognition unit, a path input unit having a plurality of path input terminals, and a weight for weighting an input from the path input unit. An adder that adds the outputs from the weights, a thresholder that thresholds the output of the adder, a threshold learner that learns the thresholds of the thresholder, and a weight value And a learner that determines A combination of units hierarchically configured multiple branched structure, a learning and recognition device formed by arranging the learning and recognition units thereon.

[0013]

According to the present invention, in the above structure, when the feature data of the object is input as an input signal to the signal input unit of the route selection unit of each layer of the learning and recognition apparatus having the hierarchical branch structure, the route of each layer is input. The selection unit determines the coupling weight with the upper layer according to the output of the quantizer, and transmits the signal to the upper layer. Here, the path output integrator integrates redundant signal paths and reduces the number of signal paths, which enables faster learning and recognition and saves memory. Learning is a weight included in the learning recognition unit in the top layer. High-speed learning is possible because learning is performed by changing only.

Further, the threshold value processor normalizes the output value of the learning recognition unit so that the output value of the final layer has the upper limit value, so that the overflow due to the learning of the connection weight and the output of the learning recognition unit is prevented. Can be prevented.

Further, the threshold learning device can improve the learning efficiency by learning the appropriate threshold and reduce the number of times of repeating the learning.

Furthermore, when referring to the output results of a plurality of learning and recognizing devices, the outputs of the learning and recognizing devices can be compared as they are by making the upper limit values of the outputs coincide with each other. By using them at the same time, the speed can be increased.

[0017]

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

FIG. 1 shows a first embodiment of the learning and recognition device of the present invention. In the present embodiment, two sets of two types of characteristic data (first characteristic data, second characteristic data) are associated with one category out of three categories. It has a branched structure, and is configured to have three layers in one branched structure. Route selection unit n11 for the first layer and the second layer
~ N12, n21 to n30, and the uppermost layer is composed of learning recognition units q31 to q33. The feature data to be judged is input to the signal input unit of the route selection unit of each hierarchy in each branch structure, and the teacher signal is input to the teacher signal input unit. From the second layer, the quantized signal is input to the quantized signal input unit of the uppermost layer, and the value of the load of the uppermost layer is input to the load input unit of the second layer.

FIG. 2 shows an embodiment of a route selection unit used in the learning and recognition device of the present invention. Here, 11 is a signal input unit, which inputs to the quantizer 2 the feature data signal to be recognized, which is input via the signal input terminal 11a. The quantizer 2 quantizes the input signal and outputs the quantized result to the quantized signal output terminal 2a and the path weighting unit 3b.
Output to. 3a is a route input unit, which is a route input terminal 3a0.
The route signal input via is input to the route load unit 3b. Reference numerals 3e1 to 3e5 denote route output terminals which connect these terminals to each other when the route selection units are hierarchically combined.

The route load unit 3b connects the loads 3b01 to 3b05 between the route input unit 3a and the route output unit 3e.
And the value of the load is a load 3b0 according to the quantization result.
1 to 3b05 and the load setting unit 3b0. In the path load unit 3b, the load setting unit 3b0 sets the load value to the load corresponding to each quantization level based on the quantization result of the input signal. Load 3b01-3
b05 weights the route signal input from the route input unit 3a, and the route output unit 3e outputs the weighted route signal to the route output terminals 3e1 to 3e5. 3f is a route output terminal integrator, which is a route output unit 3e and a route output terminal 3
The connection between e1 to e5 is switched.

3f01 to 3f05 are input terminals of the path output terminal integrator 3f, and 3f11 to 3f15 are output terminals thereof. In the initial state, the input terminals 3f01 to 3f05 and the output terminals 3f11 to 3f15 are connected to each other in a one-to-one manner, but a plurality of input terminals are connected to one output terminal to reduce the number of paths, thereby reducing the learning / recognition time. It can be reduced.
In the figure, the connection from 3f02 to 3f12 is 3f02 to 3
It shows a state where the mode is switched to f11. At this time, even if the path selected by the quantizer is 3f01, 3f0
Also in the case of 2, a signal is output from the output terminal 3e1 of the route selection unit. For the paths that are not used as a result of the integration, the weights are not changed in the learning recognition units located in the upper layer.

The figure shows a case where the route output terminals are integrated by using the load value of the uppermost layer. The terminals will be integrated after learning several times. Using the load value of the uppermost layer input to the route output integrator 3f via the load input unit 3h, the evaluation values D _ij (i
Evaluation values for the th and jth ports, where i and j
_Are defined as different terminals), and the two terminals that _minimize D _ij are integrated. As the evaluation value, for example, one defined by the equation (5) can be used.

[0023] _{D ij = Σ n (w in} -w jn) 2 ··· (5) where w _in, w _jn is out of the load that is included in the n-th of learning and recognition unit of the uppermost layer, i-th, respectively , J represents the value of the weight that belongs to the same path as the j-th terminal, and Σ _n represents the sum of all learning and recognition units.

FIG. 3 shows an embodiment of a learning and recognition unit used in the uppermost layer of the learning and recognition apparatus of the present invention. Reference numeral 1 is a teacher signal input unit, 3a is a route input unit, 3b is a route weighting unit, and weights 3b01-3 for weighting the output value of the route input unit 3a.
b050 and a learning device 3b1 for changing the connection weight. 3c is an adder for adding the outputs of the weighted weights, 3d is a threshold value processor for thresholding the output of the adder 3c, and 3d1 is for setting the threshold value of the adder 3c by learning. The threshold value learning device 5 is a quantized signal input unit, which inputs quantized signals of all route selection units to the learning device 3b1 via the quantized signal input terminals 5a01 to 5a10, and 3g is a weight value output terminal. The value of is output and used to integrate the route output terminals in route selection units.

The learner 3b1 is composed of an error calculator 3b11 and a weight value calculator 3b12. The error calculator 3b11 outputs the learning recognition unit, the input value from the teacher signal input unit 1, and the quantized signal input unit 5. Based on input, load 3b0
Determine the values 1-3b050. In the learning process, the error calculator 3b11 generates an error signal as shown in equation (6) for the weight belonging to the path selected by the quantized signal of the next lower layer, and the weight value calculator 3b12 outputs the error signal. Using, the load value is set as in equation (7).

E = t−o (6) w ′ _3b0i = η · E + w _3b0i (7) where E is the error signal value, t is the teacher signal value, and o is the learning recognition unit. Output value, _w'3b0i is the i-th load 3b0i
Is the value after the change, and w _3b0i is the value before the change. η represents a constant used in learning.

The relation of the equation (1) is used for the relation between the input and the output of the threshold value processor 3d. Here, I is the input from the adder 3d, θ is the threshold value set by the threshold value learner, and f
(I) is the output of the threshold processor. The threshold learner determines the threshold θ as shown in equation (8).

Θ ′ = ε · E + θ (8) Here, E is the value of the error signal, θ ′ is the newly learned threshold value, θ is the threshold value before learning, and t is the teacher signal. , O represents the output of the learning recognition unit, that is, the output of the threshold value processor 3d, and ε represents the learning constant.

Next, the learning operation of the learning and recognizing device thus configured will be described.

The route selection unit n1 of the first layer of each branch structure
As a route input signal to the route input terminals 1 to n12,
1 ". Also, the first feature data of the recognition object is input to the signal input terminals to the quantizers of these route selection units (in the case of this figure, two first feature data are input). Characteristic data,
Enter in two route selection units each).

Each route selection unit has these first feature data.
The quantizer 2 quantizes the data, and based on the quantized value, the weight setting unit 3b0 sets the weight value corresponding to each data based on the quantized level. In this way, multiple routes are selected in one route selection unit,
A value obtained by multiplying the route input signal by the value of the load is sent to the route input terminals of the route selection units n21 to n30 on the second layer.

The second feature data of the recognition object is input to the signal input terminal of the route selection unit of the second layer (in FIG. 1, two second feature data are respectively n21 to n25). , N
26 to n30). As in the case of the first layer, a plurality of path input signals are selected, and all outputs are sent to the respective path input terminals of all learning and recognition units q31 to q33 in the uppermost layer. Further, the quantized signal is sent from the quantized signal output terminal 2a to the quantized signal input unit 5 of all learning and recognition units in the uppermost layer. In the initial state before learning, the input terminals 3f01 to 3f05 and the output terminals 3f11 to 3f15 of the route output terminal integrator 3f are connected one to one,
After learning several times, the route output terminals 3f are integrated. The route output terminal integrator 3f obtains the evaluation value D defined by the equation (5) for all combinations of the two terminals of the route terminals 3f11 to 3f15, and sets the pair of two terminals that minimizes the value of D to 1 Integrated into one terminal.

At the teacher signal input terminal 1a of the learning recognition unit in the uppermost layer, a teacher input signal indicating which of the three categories the recognition target signal belongs to,
That is, it is 1 for the learning recognition unit in which it is desirable to output the largest value among the learning recognition units q31 to q33.
And 0 for other learning and recognition units.

For example, when q31 is to output the largest output, the route selection units n11 to n12 and n
A signal belonging to the category represented by q31 is input to 21 to n30, 1 is input to the learning recognition unit q31, and 0 is input to the other learning recognition units q32 and q33 as a teacher signal.
At this time, the same value is input to the path input unit 3a of each learning recognition unit q31 to q33 in the uppermost layer, and after the weights 3b01 to 3b050 held for each learning recognition unit are multiplied, the sum is obtained by the adder 3c, The value thresholded by the threshold processor 3d becomes the output value of the learning recognition unit. The error signal calculator 3b11 calculates the difference between the output value of the learning unit and the teacher signal. The new weight value of the path selected by the quantized signal is corrected by the weight value calculator 3b12 by a value proportional to the output of the error signal calculator. Here, the weight is not changed for the path not selected by the quantized signal. By repeating the above operation a plurality of times for all recognition targets to be learned, the learning recognition unit outputs a desired output. When learning is performed according to the equation (5), if the output o of the learning recognition unit and the value of the teacher signal t match, the change amount of the load becomes 0 and no change is made. However, when the learning recognition unit that is desired to output the largest output among the learning recognition units even when they do not match each other outputs the maximum value, the update of the load value can be omitted. As a result, the number of times of learning can be reduced. The threshold value is changed according to the equation (8) similarly to the learning of the weight.

Next, the recognition operation of the learning recognition device shown in FIG. 1 will be described. Just like the learning operation, "1" is first given as the route signal to the route input terminals of the route selection units n11 to n12 in the first layer of each branch structure. Further, the first feature data of the recognition object is input to the signal input terminal to the quantizer of these route selection units (in the case of this figure, two first feature data are , Two route selection units each).

In each route selection unit, the first feature data is quantized by the quantizer 2, and based on the quantized value, the load setting unit 3b0 quantizes the load value corresponding to the feature data. Set based on the conversion level position. in this way,
At the time of recognition, one route is selected, and a value obtained by multiplying the route signal and the load value is sent to the route input terminals of the route selection units n21 to n30 on the second layer.

The second feature data of the object to be recognized is input to the signal input terminal of the route selection unit of the second layer (in the case of this figure, two second feature data, N21 ~
Input to n25 and n26 to n30). Unlike the case of the first layer, all output values are sent to all learning and recognition units q31 to q33 in the uppermost layer.

The learning recognition unit in the uppermost layer multiplies the input value from the path input section by the connection weight, obtains the sum by the adder 3c, and outputs the thresholded value. At this time, the category represented by the learning recognition unit that outputs the maximum value among the plurality of learning recognition units is the recognition result. Further, the weight and the threshold value are not changed at the time of recognition. In this way, the learning recognition device can perform classification and recognition judgment of the recognition target object based on the input recognition target object feature data.

As described above, the learning and recognition device according to the present invention constitutes a multi-layered hierarchical structure having a branch structure,
Depending on the output of the quantizer of each route selection unit, the coupling route between the route selection units is switched, and the signal route up to the uppermost layer is selected. Here, the path output integrator integrates redundant signal paths and reduces the number of signal paths, which enables faster learning and recognition and saves memory. Learning is a weight included in the learning recognition unit in the top layer. High-speed learning is possible because learning is performed by changing only.

Further, by learning an appropriate threshold value by the threshold value learning device, the learning efficiency can be improved and the number of times of repeating learning can be reduced.

FIG. 4 shows a second embodiment of the learning recognition unit used in the uppermost layer of the learning recognition apparatus of the present invention, which has a teacher signal converter. Reference numeral 1 is a teacher signal input unit, 3a is a route input unit, and 3b is a route weighting unit, which is composed of weights 3b01 to 3b050 for weighting the output value of the route input unit 3a and a learner 3b1 for changing the coupling weight. 3c is an adder for adding outputs of weighted weights, 5 is a quantized signal input unit, 5a01 to 5a10 are quantized signal input terminals, 3c
b1 is a learning device and 4 is a teacher signal converter. The configuration excluding the teacher signal converter 4 is the same as that of the learning recognition unit described in FIG.

The teacher signal converter 4 has a function of changing the strength of the teacher signal according to the category to which the input signal belongs, by using the outputs of all learning and recognition units and the teacher signal. That is, the teacher signal converter 4 converts a binary teacher signal of 0 and 1 given as a teacher signal into an intermediate value according to the category to which the signal belongs, and generates a multivalued teacher signal. For the conversion of the teacher signal, the output of the learning recognition unit and the misrecognition characteristic obtained from the teacher signal are used. For example,
Signals belonging to category A are not misrecognized by other categories, but signals belonging to other categories are classified as category A.
When it is erroneously recognized, the teacher signal corresponding to the input signal belonging to category A is converted into a value smaller than the teacher signals corresponding to other categories. When the learning after conversion is advanced, the output value of the learning recognition unit in charge of the category A is suppressed, and the learning recognition unit in charge of the category A rarely outputs a larger value than the other learning recognition units. As a result, if the amount of change in the teacher signal by the teacher signal converter is not large, the learning recognition unit of category A outputs the maximum value among all learning recognition units when a signal belonging to category A is input. The learning recognition unit of category A outputs the maximum value less frequently when the signal other than the category A is input. Conversely, if a signal belonging to category A is erroneously recognized as another category, but a signal belonging to another category is not erroneously recognized as category A, the teacher signal corresponding to the input signal belonging to category A is classified as another category. Convert to a value larger than the corresponding teacher signal.

Next, the learning operation of the learning / recognition device in which the uppermost layer is formed by using the learning / recognition unit having the teacher signal converter 4 will be described with reference to FIG. In the learning start state, the teacher signal converter 4 does not convert the teacher signal, and the teacher signal input to the teacher signal input unit is directly input to the learner. At this time, the teacher signal converter 4 records erroneous recognition (records the frequency at which a signal of a certain category is erroneously recognized by another category) using the teacher signal and the output of the learning recognition unit. Other learning operations are the same as in the case of using the first embodiment.

After the learning is performed several times, the teacher signal converter 4 converts the teacher signal based on the record of the misrecognition.
For example, the learning recognition unit t31 does not output a value larger than t32 and t33 with respect to an input signal not in charge of
When t31 or t33 should output the maximum value, t32 or t33 is t
If a value larger than 31 is often output, t31
Is converted to a value less than 1 and larger than 0 (such as 0.8) when it is desirable to output the maximum value, and 0 is converted for other learning recognition units. Do not do. On the contrary, t31 is apt to be erroneously recognized, and t31
If there is little erroneous recognition in the category in charge of, the teacher signal for t31 is converted to a value larger than 1 (such as 1.2), and 0 is not converted for other learning recognition units. . In the configuration having the teacher signal converter as described above, the number of required learning is reduced by changing the teacher signal based on the erroneous recognition result of the learning recognition device so that the teacher signal converter reduces the erroneous recognition. , Enables faster learning.

Since the teacher signal is not used during the recognition operation, the recognition operation is the same as in the case of using the first embodiment.

As described above, the learning and recognition device according to the present invention forms a multi-layer hierarchical structure having a branch structure,
Depending on the output of the quantizer of each route selection unit, the coupling route between the route selection units is switched, and the signal route up to the uppermost layer is selected. Here, the path output integrator integrates redundant signal paths and reduces the number of signal paths, which enables faster learning and recognition and saves memory. Learning is a weight included in the learning recognition unit in the top layer. High-speed learning is possible because learning is performed by changing only.

Further, the learning efficiency can be improved and the number of times of learning can be reduced by the threshold learning device learning an appropriate threshold.

In the configuration including the teacher signal converter, the teacher signal is changed based on the erroneous recognition result of the learning and recognizing device so that the teacher signal converter reduces the erroneous recognition, thereby reducing the required number of times of learning. It enables faster learning.

Each means of the present invention can be realized by software using a computer, or can be realized by using a dedicated hardware circuit having each of these functions.

[0050]

As is apparent from the above description,
The learning recognition device of the present invention, when the feature data of the target object is input as an input signal to the signal input unit of the route selection unit of each layer of the learning recognition device having the hierarchical branch structure as described above, The path selection unit determines the coupling weight with the upper layer according to the output of the quantizer, and transmits the signal to the upper layer. Here, the path output integrator integrates redundant signal paths and reduces the number of signal paths, which enables faster learning and recognition and memory saving, and learning is a path included in the learning recognition unit of the top layer. Since learning is performed by changing only the load between the input unit and the adder, high speed learning is possible.

Further, when the threshold value processor normalizes the output value of the learning recognition unit so that the output value of the final layer has the upper limit value, the overflow due to the learning of the connection weight and the output of the learning recognition unit is caused. Can be prevented.

When the threshold learning device learns an appropriate threshold value, the learning efficiency can be improved and the number of times of learning can be reduced.

When referring to the output results of a plurality of learning and recognizing devices, if the upper limit values of the outputs are matched,
You can compare the output of each learning recognition device as it is,
The speed can be increased by using a plurality of learning and recognition devices at the same time.

Further, in the configuration having the teacher signal converter,
By changing the teacher signal so that the teacher signal converter reduces erroneous recognition based on the erroneous recognition result of the learning recognition device,
The practical effect is great because the number of required learnings is reduced and the learning speed can be increased.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of a learning recognition device of the present invention.

FIG. 2 is a block diagram showing an example of a route selection unit used in the learning recognition device.

FIG. 3 is a block diagram showing a first embodiment of a learning recognition unit used in the learning recognition device.

FIG. 4 is a block diagram showing a second embodiment of a learning recognition unit used in the learning recognition device.

FIG. 5 is a block diagram showing a second embodiment of the learning recognition device.

FIG. 6 is an overall configuration diagram of a conventional learning recognition device.

FIG. 7 is a configuration diagram of an output signal calculation unit of the learning recognition device.

FIG. 8 is a configuration diagram of a multi-input / single-output calculation unit of the output signal calculation unit of the learning recognition device.

FIG. 9 is a graph showing a characteristic function of a threshold value processor of the multi-input / single-output calculation unit of the output signal calculation unit of the learning and recognition device.

[Explanation of symbols]

 1 Teacher Signal Input Section 1a Teacher Signal Input Terminal 2 Quantizer 2a Quantized Signal Output Terminal 3a Path Input Section 3a01-3a050 Path Input Terminal 3b Path Loading Section 3b0 Weight Setting Section 3b01-3b050 Weight 3b1 Learner 3b11 Error Calculator 3b12 Load value calculator 3c Adder 3d Threshold processor 3d1 Threshold learning device 3e Path output unit 3e1 Path output terminal 3f Path output terminal combiner 3f01-3f05 Path output terminal combiner input terminal 3f11-3f15 Path output terminal integrated Output terminal 3h weight input section 3g weight value output terminal 4 teacher signal converter 5 quantized signal input section 5a01 to 5a10 quantized signal input terminal 11 signal input section 11a signal input terminal q, t learning recognition unit n path selection unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Taiji Yuki, 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (5)

[Claims] 1. The route selection unit includes a signal input unit, a quantizer for quantizing an input signal from the signal input unit, and a first route input unit having a single or a plurality of route input terminals. A route output unit having a single or a plurality of route output terminals, a route output integrator that replaces the plurality of route output terminals with a single route output terminal, and the route input according to the quantized output of the quantizer. Section and a path weight section for selecting a path between the path means, and the learning recognition unit is a teacher signal input section for inputting a teacher signal, and a second path having a plurality of path input terminals. An input unit, a weight for weighting an input from the second path input unit, an adder for adding an output from the load, and a direct or indirect manner to the added value and the teacher signal. The value of the load is determined based on And a plurality of the route selection units, the output terminals of the route output integrator and the first route input unit are combined in a hierarchical manner to form a plurality of branch structures, and an upper layer thereof. To
The learning and recognition device, wherein the learning and recognition unit is arranged while combining an output terminal of the path output integrator and the second path input unit. 2. The route output integrator for each route selection unit,
2. The learning and recognition device according to claim 1, wherein the route output terminal is replaced by using the value of the weight included in the upper layer. 3. A learning device, an error signal calculator that calculates an error signal by using the output of the teacher signal and the learning recognition unit, an error signal of the error signal calculator, and a quantized signal of the quantizer. The learning and recognition device according to claim 1 or 2, further comprising: a load calculator that determines the value of the load using. 4. The learning recognition unit has a threshold value processor for thresholding the output of the adder, and a threshold value learner for learning the threshold value of the threshold value processor. The learning and recognition device according to any one of claims 1 to 3, characterized in that: 5. The learning recognition unit has a teacher signal converter for performing intensity conversion of the teacher signal by using the output of the learning recognition unit and the teacher signal.
Any of the described learning and recognition devices.