JPH0451353A - Pattern learning system - Google Patents

Abstract

PURPOSE:To perform learning arranging phases simultaneously and automatically by selecting the output of a false pattern generating means by a switching means when an idle signal decided in advance is inputted for prescribed time. CONSTITUTION:When no external input is added any more, a switching circuit 14 selects the output of a false pattern generation circuit 13, and adds it on a learning processing circuit 11. A learning pattern data base 12 stores pattern information to be learned and learned pattern information, and the false pattern generation circuit 13 reads out those information, and changes the information by a noise pattern, etc., and generates a false learning pattern. A generated false learning pattern is added on the learning processing circuit 11 via the switching circuit 14, and learning when no input data is added is performed. Thereby, a learning means can be efficiently and sufficiently offered.

Description

[Detailed Description of the Invention] [Summary] The present invention relates to a pattern learning method in which a recognition device is trained to obtain a specific output in response to a specific pattern input. The purpose of the present invention is to provide a pattern learning method that automatically detects and excludes patterns that do not match the pattern, and provides a learning means that performs self-learning and recognizes an input signal when an input signal is added, and a pattern that the learning means should learn. a learning pattern storage means for storing, and a pseudo pattern for generating a pseudo learning pattern by adding a specific variation to the pattern to be learned in the learning pattern storage means 2, in addition to the learning means, and a pseudo pattern for causing the learning means to learn. It is configured to be similar to the generation means.

[Industrial application field]

The present invention relates to a pattern recognition device for outputting specific data such as signal processing, character recognition, graphic recognition, image recognition, voice recognition, robot control, etc., and more specifically, to obtain a specific output in response to a specific pattern input. This invention relates to a pattern learning method for learning a recognition device.

[Prior art] Signal processing, character recognition, figure recognition, image recognition, voice recognition,
In fields such as robot control, a processing system is often adjusted (learned) so that it always outputs a specific output signal sequence (output pattern) in response to a certain input signal sequence (input pattern).

For example, variable adjustment in control systems and weight learning using neural networks are used as such learning processing systems in the various fields mentioned above.

When learning a certain processing system, a set of typical input patterns (hereinafter referred to as learning patterns) is prepared, and after sufficient training is performed until some criterion for determining completion of learning is met, the actual processing is performed. It is used as a system. In some cases, not only the input pattern but also the corresponding expected output pattern may be paired with the input pattern to form a single learning pattern.

In the above-mentioned learning method, there is a phase in which the processing system is trained,
The phases that use the processing system are completely separate.

That is, in the learning phase, as shown in FIG. 16, patterns are sequentially read out from the learning pattern set 40 and added to the learning system 51, and the difference with the output pattern is determined, and the patterns are added to the learning system 51 as self-adjustment. ing. When the learning is completed, the learning system is used as the processing system, and even when an input that slightly changes from the input at the time of learning is added, the desired output pattern is obtained as shown in Figure 17. .

[Problem to be solved by the invention]

In the method of performing post-learning processing described above, the processing system operates as expected for the learned pattern, but
If a pattern that was not presented during learning is input, there is a problem in that an unexpected output pattern is output. That is, there is a problem in that it is not possible to learn new patterns while using the device.

Furthermore, patterns that exist in the natural world (signals, images, sounds, etc.) always contain noise, and it is not possible to prepare and present a complete set of learning patterns during learning. The problem is that it cannot be applied to patterns that include In particular, in response to a significant change, an incorrect output pattern was output as shown in FIG.

Furthermore, in the conventional method described above, the phase in which the processing system is trained and the phase in which it is used are completely separated.
If you want to learn a new pattern, you have to go back to the learning phase again, add a new learning pattern, and start learning again.

An object of the present invention is to provide a pattern learning method that simultaneously and automatically performs learning, and further automatically detects and excludes patterns that should not be learned.

A pseudo learning pattern is generated by adding variation, and the learning means 1 is made to learn in addition to the learning means 1 described above.

A switching means 4 is provided at the input of the learning means 1, and selectively applies the input signal and the pseudo learning pattern to the learning means 1. For example, in this selection, the pseudo learning pattern is selected when the input signal is an idle signal for a predetermined period of time, and when an input is present, the input signal is selected and added to the learning means 1.

[Failure to solve the problem]

FIG. 1 is a block diagram of the principle of the present invention.

The learning means 1 performs self-learning and recognizes an input signal when it is applied.

The learning pattern storage means 2 stores patterns to be learned by the learning means 1. For example, the information to be stored is the pattern to be learned and its expected output data.

The pseudo pattern generation means 3 is the learning pattern storage means 2.
When a specific input signal, ie, a pattern to be recognized, is added to the pattern to be learned stored in the memory, the switching means 4 selects the input signal and applies it to the learning means 1. The learning means 1 recognizes the input signal based on the information learned so far. At the same time, students also do self-study. On the other hand, when no input signal is applied, that is, when a predetermined idle signal is input for a specific time, the switching means 4 selects the output of the pseudo pattern generating means 3.

The pseudo pattern of the pseudo pattern generating means 3 is the pattern to be learned stored in the learning pattern storage means 2 with specific variations added thereto. Since the learning means 1 learns about this specific variation while in the idle state, even if various input signals are added after learning, the input signal can be accurately recognized.

Furthermore, the pseudo pattern generating means 3 is checked by the first checking means.
It is detected whether or not the pseudo learning pattern generated by the learning pattern is inappropriate as a learning pattern, and when it is inappropriate, the learning pattern is not added to the learning means 1, and the inappropriate learning pattern is not added to the learning means 1. Preventing learning. In addition, when the generated pseudo learning pattern is recognized by learning means 1,
When the obtained output pattern differs from the expected output pattern, learning is stopped to prevent recognition of incorrect data for subsequent learning and recognition.

〔Example〕

Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 2 is a block diagram of the first embodiment of the present invention.

A first embodiment of the present invention includes a learning processing circuit 11 (for example, a neural network or a linear or nonlinear control system) that automatically changes internal parameters in order to output a specific output pattern (result) in response to a learning pattern; Initial learning pattern presented in advance and learning processing circuit 11
a learning pattern database 12 that stores patterns learned in the learning pattern database 12; a pseudo learning pattern generation circuit 13 that generates pseudo learning patterns by adding specific variations to the learning patterns in the learning pattern database 12; , and a switching circuit 14 for switching to a learning pattern from a pseudo pattern generation circuit 13 which is an internal learning pattern.

When an external input is being applied, the switching circuit 14 selects that external input, and the learning processing circuit 11 executes recognition processing.
Output the result. This recognition uses the results learned previously. Further, the output of the recognition is also added to the learning processing circuit 11 again, and is used as information for subsequent recognition.

On the other hand, when no external input is applied, the switching circuit 14 selects the output of the pseudo pattern generation circuit 13 and applies it to the learning processing circuit 11. The learning pattern database 12 stores pattern information to be learned and learned pattern information, and the pseudo pattern generation circuit 13 reads out this information and changes the information with a noise pattern or the like to generate a pseudo learning pattern. . Then, the generated pseudo learning pattern is transferred to the learning processing circuit 1 via the switching circuit 14.
In addition to step 1, have the robot perform unmanned learning.

The operation will be explained in more detail. The learning processing circuit 11 will be referred to as a "learning system" when the processing system is in a state where learning is performed, and as a "processing system" when normal signal processing is performed without learning.

In the initial state, several sets of learning patterns that the user wants to learn are input into the learning pattern database 12. The learning system 11 is in a state where it is not performing any learning.

The initial learning phase is from database 12 to learning system 11.
Input the pattern as it is and perform learning. At this time, the switching circuit 14 is always in a mode in which the learning pattern is input to the learning system 11. The learning system 11 executes learning for the presented learning pattern, and ends the learning based on some learning end condition. For example, the process ends when the required number of patterns have been learned in advance.

After the initial learning is completed, in the next phase, if there is an external input signal, receive the input pattern from the outside,
The processing system 11 that has completed the learning processes the signal and outputs it. At this time, the switching circuit 14 switches the bus to the input side. The switching circuit 11 determines when the 0 state of the input signal continues longer than a predetermined time, and switches as if there is no external input when this state is present.

When the external input pattern is no longer input, the switching circuit 14 switches to the internal path side, and the pseudo pattern generation circuit 13 starts operating. The pseudo pattern generation circuit 13 sequentially extracts learning patterns from the learning pattern database 12, adds minute variations to them based on specific rules, and outputs them as pseudo patterns. If it is necessary to add an expected output pattern to the learning pattern, the expected output pattern is the same as the original learning pattern before the variation is added. This is based on the concept of lohastness, which states that even if there is a slight change in the human cover turn, the output pattern will not change.

The generated pseudo learning pattern is input to the learning system 11 via the switching circuit 14, and the learning system 11 performs self-learning. At the same time, the pseudo learning pattern is additionally registered in the learning pattern data database 12.

Then, the pseudo pattern generation circuit 13 generates a pattern again and continues learning.

On the other hand, if an external input occurs on the way, switching circuit 1
4 detects the input, interrupts learning, and enters normal signal processing mode.

The above processing procedure automatically separates the learning phase from the normal signal processing phase, allowing efficient learning to be performed.

The operation will be further explained in detail below using a flowchart. FIG. 3 is a processing flowchart of the first embodiment of the present invention. After performing initial learning (SL), it is determined whether there is any external input (S2). This initial learning execution (Sl) is a process of storing initial pattern data in the learning pattern database 12 in FIG. The block diagram of the first embodiment shown in FIG. 2 starts operating by executing the determination (S2) as to whether or not there is the next external input. In determining whether there is an external input (S2), when it is determined that there is an external input (Y), signal processing (S3) is executed. This signal processing is the operation of the recognized processing system, and recognizes patterns and the like. When the signal processing (S3) is completed, the determination (S3
) to execute again.

If there is no external input in the determination (S2) (N), it is determined whether automatic learning is to be executed (S4). In FIG. 2, a circuit that performs processing in a conversational format is not shown, but the processing is similar to that performed by a conversation processing section 15, which will be described later, and has a similar configuration. In the determination of automatic learning execution (S4), when automatic learning is not executed (N), it is determined whether conversation processing is executed (S5) L, and when conversation processing is not executed (N), it is determined again (S2) and executed. .

When the conversation processing is to be executed (Y), the conversation processing (S6) is executed, and the execution is performed again after determining whether there is an external input (S2).

Note that when automatic learning is not executed and conversation processing is not executed, the configuration is similar to that of the conventional system.

On the other hand, when it is determined that automatic learning is to be executed (Y) in the automatic learning execution determination (S4), the pseudo pattern generation circuit 13 executes the process to sequentially read learning patterns from the learning pattern database 12 and transform the read learning patterns. (pseudo pattern) is generated (S7). Then, it is input to the processing system, that is, the learning processing circuit 11 (S8).

The input to the processing system (S8) is a process in which the pseudo pattern is input to the learning processing circuit 1 via the switching circuit 14. When the pseudo learning pattern is added, the learning processing circuit 11 executes learning in the learning system (S9) and registers it in the learning pattern database 12 (SlO). Then, after registration, the process is executed again from the determination (S4).

The above-mentioned pseudo-learning pattern is one in which a minute variation is added to one of the learning patterns within a range that does not deviate from the category to which the pattern belongs.

Learning and recognition are performed through the above operations. Since this learning is performed unattended, a large number of learning patterns can be learned, and the recognition rate can be improved. Also,
Since you study while you are away, you can equivalently shorten your study time.

FIG. 4 is a block diagram of a second embodiment of the present invention.

In the second embodiment shown in FIG. 4, in addition to the configuration of the first embodiment shown in FIG. A check circuit 16 that determines whether or not
．． 17. Pattern file 18 for storing pseudo-learning patterns determined to be inappropriate as learning patterns; history database 19 for storing the history of generation in the pseudo-pattern generation circuit 13'; -U storage for generated pseudo-learning patterns; A pseudo pattern file 20, an external input collection circuit 22, and a collection pattern file 21 are added.

First, when generating a pseudo learning pattern, the number of the learning pattern that is the source of the generation and the number of the generated learning pattern are registered in the history database 19 as a pair.

In addition to the mode in which the generated patterns are directly input to the learning system 11, the pseudo pattern generation circuit 13' also stores them in the pseudo pattern file 10 once, and when a certain number of patterns have been accumulated or when there is no external input It has a mode that outputs the data all at once for checking and learning when it is ready for learning. In this case, the number of patterns to be processed at once must be specified externally (.

The check circuit 16 compares the pseudo learning pattern with the learning pattern before any variation is applied, and determines whether or not it is appropriate. If it is appropriate as a learning pattern, it is input to the processing system 11 via the switching circuit 4.

The processing system 11 performs recognition processing and outputs it. The output signal of the check circuit 17 is compared with the expected output pattern of the learning pattern to determine whether it is appropriate. If the pseudo learning pattern is determined to be inappropriate as a learning pattern by the check circuit 16 or 17, it is stored in the pattern file 18 and no learning is performed. If the learning pattern is appropriate, a signal is received from the check circuit 17, and the learning system 11 learns the learning pattern.

This learning pattern is then additionally registered in the learning pattern database 12.

The conversation processing unit 15 determines whether the learning patterns generated by the pseudo learning pattern generation are truly appropriate, and whether the learning patterns stored in the learning pattern database 12 are truly appropriate, or whether the learning patterns stored in the pattern file I8 are Ask the user whether the pseudo-learning pattern is truly inappropriate. The user looks at the pseudo-learning pattern (and its expected output pattern) displayed on the screen and makes a decision. If it is determined whether the generated learning pattern is truly appropriate or not, the pattern is discarded and the next pseudo pattern is displayed. That is, if there is an inappropriate learning pattern in the learning pattern database 12,
The user instructs the conversation processing unit 19 to that effect, and the conversation processing unit 19 deletes the pattern from the database 12.

On the other hand, if there is a suitable learning pattern in the pattern file 18, the user informs the conversation processing unit 1 of this fact.
9, and the conversation processing unit 19 additionally registers the pattern in the learning pattern database 12.

If inappropriate learning patterns are included in the learning pattern database 12, the learning patterns generated by varying the patterns will also be inappropriate, and all such sequences must be deleted from the database. To do this, search the history database 19 using the conversation process described above,
Find the learning pattern number that was judged to be inappropriate, and store all the learning patterns that occurred from there in the learning database 12.
Delete from. Note that the history data of the deleted learning pattern is also marked for deletion so that it will not be referenced again.

The external input collection circuit 22 receives an external input, converts it into an output pattern in the learning processing circuit 11, and then searches the learning pattern database 12 to see if there is a learning pattern that approximates the external input pattern. If such a learning pattern is found in this search and the expected output pattern and the output pattern are significantly different, the external input pattern and output pattern are stored in the collected pattern file 21. That is, the collected pattern file 21 stores patterns that produce outputs that are inconsistent with the learning results so far, among the patterns that are input from the outside.

The conversation processing unit 15 sequentially displays patterns and their output patterns on the screen starting from the collected pattern file 21, and inquires of the user whether or not the output results are correct.

If the output pattern is inappropriate for the input pattern, the user rewrites the output pattern to the originally expected output pattern and converts it to the learning pattern/database 1.
2 to additionally register. In this way, the external input pattern 9 is used for learning correction.

FIG. 5 is a processing flowchart of the second embodiment of the present invention. The initial learning execution (Sll) is similar to the initial learning execution (Sl) shown in FIG. This initial learning execution (S1
1. ), the determination of whether there is an external input (S12) is performed, and when it is determined that there is an external input (Y), the signal processing (S12) is performed.
13), the input pattern is recognized. Then, a determination (S1, 4) is performed as to whether the pattern obtained by the signal processing (S13) is to be collected, and if pattern collection is not to be performed (N), the determination (S12) is again performed. When it is determined in pattern collection determination (S14) that pattern collection is to be performed (Y), external input collection is performed (S15). Then, the external input collection circuit 22 in FIG. 4 operates and collects patterns into the collection pattern file 21.

Then, the determination is performed again (S12).

When it is determined in the determination (S12) that there is no external input (
For N), determine whether it is automatic learning execution 31.6),
When automatic learning is not performed (N), conversation processing is executed (S1
7). Further, when it is determined that the conversation processing is not to be executed (N), the execution is performed again based on the determination that there is an external input (S12).

If the conversation processing is to be executed (Y) in the determination (S17), the conversation processing (S18) is executed and the determination is made again (S12). In addition, processing (312,513S16.S17,51
8) is the process of the flowchart shown in No. 3 (S2.S3
．． S4. S5. This is the same process as S6).

On the other hand, when it is determined in automatic learning execution (S16) that automatic learning is to be executed (Y), a pseudo learning pattern is first generated (319). This pseudo learning pattern generation is similar to the process (S7) in FIG. 2 described above. After pseudo pattern generation (S19), input pattern check (S2
0) is executed.

This input pattern check is performed by the check circuit 16.

This check circuit 16 determines whether the pattern is inappropriate (S21) L, and if it is an inappropriate pattern (Y
) stores the pseudo-i learning pattern in the pseudo-pattern file 20 (S22).

After the process (S22), automatic learning execution is determined (S22).
Repeat from l6) again. Further, if the pattern is not inappropriate in the determination (S20) (N), the history database is registered (S23). This history database registration process (S23) is a process that the pseudo pattern generation circuit 13' stores in the history database 19. If the pattern is changed by the check circuit 16 and is not inappropriate, the process (323) is executed to register the history database, and then input to the learning processing circuit 11 via the switching circuit 14, and input to the processing system (S24). The learning circuit 11 then executes the processing system. The data output by this processing system is sent to the check circuit 17, which checks the output pattern (S25) and determines whether it is an inappropriate pattern (S25).
26), and if the pattern is inappropriate (Y), the inappropriate pattern is stored in the pattern file 18 (S27).

Then, the determination is performed again (S16). Further, in determining whether the pattern is inappropriate (S26), if the pattern is not inappropriate (N), learning is executed in the learning system (S28). Then, registration in the learning pattern database (S29) and determination (S16) are performed.

FIG. 7 is a data flow chart of the second embodiment.

In this data flowchart, the learning pattern is No.
There are four two-dimensional data numbers 1 to No. 4, and an expected output signal T is added.

The learning processing circuit 11 has an input layer with a number of neurons (3) and a hidden lt
A three-layer neural network I-work is assumed, where i has a number of neurons (3) and the output layer has a number of neurons (1). This neural network receives two signals x1x2 as input, changes the signals internally, and outputs a signal Y from the output layer. Regarding the characteristics and learning methods of this type of neural network, please refer to "Parallel Distributed Processing" by David Rumelhart et al., published by MIT Press in 1986 (David Rumelhart et al., "P.
arallel Distributed Proc
essingExplorations in th
e Microstructure of C.

ngniti on-” 1986 MTT Press).

The input layer and hidden layer of the neural network according to the second embodiment of the present invention each have a neuron u that always outputs 1.
3. Has u6. The strength of the connection between each neuron (hereinafter referred to as weight) is written as WiJ. Here, 1 is the neuron number to which the connection is made, and j is the neuron number from which the connection is made. The input/output relationship of each neuron is assumed to follow a sigmoid function.

FIG. 14 is a neuron input/output relationship diagram.

Let the human input signal of a specific neuron i be 1nput, and its output signal be output\1nput, = ΣW, j・output; t
01 mountain) uLI= 1 / (1+ e x p (
−input+)). In addition, the output signal output
is a sigmoid function (see Figure 15).

First, perform initial learning. The initial weight data of the neural network are random values shown in FIG. 8th
The data in the learning pattern database shown in the figure undergoes learning using the four sets of learning patterns shown in the table, resulting in the weight state of the neural network shown in FIG. 11. When a learning pattern is input, the corresponding expected output pattern is output. The neural network here uses the hack propagation learning method.

Regarding this learning method, the above-mentioned literature is used as a technique.

This completes the initial learning and allows normal signal processing. Next, determine whether this system is idle. Detection of the presence or absence of an input signal is determined based on whether the state of input signal 0 continues for a long time. That is, when the input signals (x+, X2) are all 0, N
OR (negation of logical sum) outputs 1, and when the signal reaches a certain number of times in the clock counter (1-1), the multiplexer path is switched to the internal signal side. If it is in an idling state, then a pseudo learning pattern is generated.

The pseudo learning pattern is generated by superimposing random noise ω on the learning pattern. If the input pattern is P-dimensional, a P-dimensional random value vector ω is added to the learning pattern. The expected output signal T corresponding to the generated pseudo-pattern is assumed to be the same as that of the input pattern (T). In this example, first, the learning pattern No. 1 is taken out,
Add small fluctuations to it.

The criterion for determining whether a pseudo pattern is appropriate as a learning pattern is that the difference pattern between the original pattern and the generated pattern is taken and its norm is less than or equal to a certain darkness value. In other words, the pattern that was generated is L = (xll, X12°...1Xl11),
The generated pattern is
2p), the learning pattern is one that satisfies the following criteria.

Let us now assume that the threshold value ε+ is 0.01. If the pseudo pattern generated for No. 1 is (0, 01, 0, 01), it satisfies the above conditions and is therefore regarded as a learning pattern.

In addition, to compare output patterns and determine whether they are appropriate as learning patterns, compare the output pattern obtained by actually inputting the input pattern to the processing system (here, Y) and the expected output pattern T. done by. That is, a learning pattern is one that satisfies the following criteria.

≦Threshold ε2 ・・・・・・Condition (2) Now threshold ε2=
It is set to 0.01. Pseudo pattern (0801, 0,01)
When input into the neural network, the output pattern Y is 0.01. Since T is 0, the second condition is satisfied.

This pattern is therefore considered a learning pattern and is input into the neural network again to begin self-learning.

When this pattern is re-learned, the internal weight data will change slightly as shown in the weight data chart after learning the pseudo-learning pattern No. 45 in Fig. 12, No. 1 to No.
The expected output pattern will be output for all of 5. Further, the pseudo pattern is additionally registered as a learning pattern in the learning pattern data file, resulting in the state shown in FIG. At the same time, the history data shown in the history database data diagram of FIG. 9 is created and stored. In this case, the pattern No. 5 was generated from N091, so
Write data 1→5 in the data chart shown in FIG. This completes the processing for one learning pattern.

Subsequently, pseudo patterns are generated in the same manner for patterns No. 2. By adding random fluctuations, (
1, 20, 0, 30) is generated. In this case, since condition (1) is not satisfied, it is not recognized as a learning pattern. Therefore, this pattern is stored in the pattern file 18. Furthermore, pattern file 1
8 has the same structure as the learning pattern example shown in FIG.

On the other hand, when an external input occurs, the switching circuit 4 immediately
changes the path to the external input side, and the external input is now input to the signal processing system. The external input collection circuit 22 is activated when there is an external input, and searches whether a pattern similar to the external input pattern exists in the learning pattern database 12.

That is, a learning pattern X that satisfies the following conditions for external input X=(χ1.X2). Don't look for it.

≦Threshold ε3 ・・・・・・Condition (3) Furthermore, the output pattern Y at that time and the learning pattern XQ
The expected output pattern T. are compared under the following condition (4), and a pair of external human power X and output Y that satisfies this is stored in the collection pattern file 11.

Condition (3) determines which of the learned patterns is similar to the external input, and condition (4) determines whether the external input pattern is similar to the learned pattern, but the output result is different from the learned pattern. Distinguish between discrepancies. As long as there is external input,
This determination process is repeated.

This will be explained using a more specific numerical example. Darkness value is ε3=0.2
0. Let ε4=0.05. Assume that the learning system stops in the state of the weight data chart after learning the pseudo learning pattern No. 5 in FIG. 12, and an external input is added.

At this time, the state of the learning pattern data chain 12 is the 8th state.
It is assumed that five patterns No. 1 to No. 5 are registered. As external human power X, (x
Suppose that I, X2 ) - (0, 80, 0, 90) is input. At this time, from the input/output relationship shown in FIGS. 14 and 15, the signal processing system 11 outputs Y-0, 2 as the output pattern.
Outputs 7. When searching for the learning pattern closest to this external human power X from data base 2, No. 4 is found.

So, if we judge condition (3), (1,0-0,80)2
+(1,0-0,90)2=0.05≦0.20, and the condition was not met. Next, when condition (4) is determined, (0, 2
'7'-0,0)2=0.073≧0.05, so the condition is also not satisfied. Therefore, this external input is stored in the collection pattern file 21. Next external input is X = (0
, 78, 0, 20), No. is close to 2, and if the same judgment as above is made, condition (3) is not satisfied.

(1,00,78)2+(0,0-0,20)2=0
．． 218≧0.20.

Therefore, this external input will not be collected.

If the external input is similarly judged and collected based on conditions (3) and (4), the data chart of the pattern file shown in FIG. 13 will be obtained. Later, using the conversation processing unit 15, the 13th
The user determines whether or not to register by referring to the collection pattern file 21 shown in the figure.

Repeat the above learning phase and external human resources phase.

After an appropriate period of time has elapsed, when the user interrupts the conversation processing section 15, the automatic learning phase is interrupted and the conversation processing section controls data input/output. The conversation processing section 15 is equipped with a display for displaying learning patterns and a keyboard or pointing device for transmitting the user's decisions.

The conversation processing unit 15 displays the pattern in the pattern file and inquires of the user whether the pattern is appropriate as a learning pattern. If the user determines that the pattern is inappropriate, the pattern will be deleted from the file. If the user determines that the pattern is appropriate, the pattern is additionally registered in the learning pattern database. At the same time, historical data is also added.

Conversely, the patterns in the learning pattern database are displayed in order on the display and the same judgment process is performed. If there is an inappropriate learning pattern, that pattern is deleted from the learning pattern database, and at the same time all patterns generated from that pattern are searched from the historical data, and these are also deleted from the database. Adding or deleting pattern data to the database,
Processing such as searching the history database is similar to normal data processing, so it will be described as an obvious technique here.

In addition, the pattern is retrieved from the collection pattern file in which the external input pattern was collected, displayed on the screen, and the user is asked whether the output pattern is appropriate for the input pattern. If the output is not valid, the user rewrites the output pattern of this collected pattern into an expected output pattern, considers it as a learning pattern along with the input pattern, and additionally registers it in the learning pattern database 2.

FIG. 6 is a processing flowchart of the conversation processing section 15.

When the conversation processing unit 15 starts execution, it first determines which database to process (330).

Note that this process (S30) is determined based on instructions from the operator. When it is determined in the determination (S30) that the process is for pattern file 8, it is first determined whether it is the end (
S31) is executed, and if it is not finished (N), the pattern is read out (S32) and the pattern is further displayed (333). As a result, the pattern is displayed on the display section of the conversation processing section 15, and the operator determines whether the pattern is inappropriate (S34.) L. If the operator indicates that the pattern is inappropriate (Y), the pattern is deleted ( S35). Then, it is determined whether the process is finished or not (S30) and the process is performed again.

Also, when it is determined that it is not inappropriate in the determination (S34) (
At step N), a determination as to whether or not to register is displayed (S36), and when the operator inputs that registration is not to be performed (N), a determination as to whether or not it is finished is subsequently executed (S30). Further, when the user inputs the intention to register (Y), additional registration is made in the database 12' (S37).

Then, it is additionally registered in the history database 19 (338).
. Then, the determination is performed again (S30).

On the other hand, when it is determined in the determination (S30) that it is an instruction from the learning pattern database 12, it is first determined whether it is the end or not (S39), and if it is not the end (N), the pattern is read out in the same way as described above. 540), pattern display (341), and determination of unsuitability (S42). When it is determined in the determination of whether it is inappropriate (S42) that it is not inappropriate (N), the process is executed again from the determination of termination (S39). When it is determined that it is inappropriate (Y), a determination is made as to whether to delete it (S43). When the operator inputs that it is not to be deleted (N), the process is executed again from step S39. When it is deletion (Y), deletion is performed (S44). Then, a pattern corrected from that pattern is searched from the historical data (S45). The history data is stored in the history database 19, and processing (345) is executed to delete the history data as well. Then, the history data of the searched pattern is deleted (S46).

Subsequently, the next generated pattern is searched (S47) and determined (S48) L. If the pattern does not exist (Y), the end determination is repeated again. Also, when it is determined that it exists (N
), repeat from (S45) again.

When it is determined that the file is a pseudo pattern file 20 in the determination (330), it is first determined whether the file is finished (S49)L,
Similarly, it is determined whether the pattern is read (S50), pattern display (S51), and whether it is inappropriate (S52)
Execute. When the operator inputs that it is inappropriate in the determination (S52) (Y), a deletion process (S53) is executed. If it is determined that it is not inappropriate, a determination is made as to whether to register (S54). In addition, when the operator inputs that he/she does not want to register (N), the end is determined again (S49).
Execute more. Also, when registering and when the operator inputs (Y), additional registration is made in the database 12' (355
)L, then additional registration in the history database 19 (S56
)L, perform the determination again (S49). When the instruction in this pseudo pattern file 20 or the pattern file 18 described above is specified, only the input file to be processed is different, and the operation of the processing itself is completely the same.

In addition, in the determination (S30), the collection pattern file 1
When 1 is specified, determination of completion (S57), pattern reading (358), pattern display (S59), and determination of unsuitability (S60) are executed. This process is also similar to the process described above. Then, when it is determined that it is inappropriate (Y) in the determination of whether it is inappropriate (S60), the deletion process (S6
1) is executed, and then it is determined whether the process has ended. When it is determined that the application is not inappropriate (N) in the determination of whether it is inappropriate (360), the next process is to determine whether it is registered (S60).
is executed and the operator determines that it is not registered (N)
Then, the end processing determination (357) is executed. When the operator inputs that it is registration (Y), the expected output signal is written (S62). This is done by the operator. When the output signal is input, it is additionally registered in the database 12' (S63, 64), and further registered in the history data 19 (S64).
Execute from step S57). When the operator inputs the corresponding information in response to the above-mentioned display to the operator, the database 12' history database 1
At 9, storage is performed according to an operator's instruction. Also, if you want to delete it, delete it.

Note that the above-mentioned determination of whether it is the end (S31, S39, S
49, 557), the conversation process ends (Y).

As described above, according to the embodiments of the present invention, automatic learning is performed when no human effort is required, and it is also determined whether the automatically learned learning pattern is inappropriate, and if it is inappropriate, the operator is made to determine through conversation processing. , it is possible to obtain recognition processing with fewer errors.

〔Effect of the invention〕

As described above, according to the present invention, a more efficient and sufficient learning means can be provided in fields such as filter design and pattern recognition for signals, characters, figures, images, sounds, etc. That is, the learning time can be shortened by automatically performing a large amount of learning when the processing system is not performing actual signal processing (idling state).

Furthermore, by automatically generating pseudo-learning patterns from a small number of learning patterns, a processing system that is resistant to noise can be efficiently configured. Furthermore, the accuracy of learning is achieved through pattern checking and conversation processing.

[Brief explanation of the drawing]

Fig. 1 is a principle block diagram of the present invention, Fig. 2 is a block diagram of a first embodiment of the present invention, Fig. 3 is a processing flowchart of the first embodiment of the present invention, and Fig. 4 is a block diagram of the principle of the present invention. A configuration diagram of the second embodiment, FIG. 5 is a processing flowchart of the second embodiment of the present invention, FIG. 6 is a processing flowchart 1 to 1 of the conversation processing section, and FIG. 7 is data of the second embodiment. Flowchart, Figure 8 is a data diagram of the learning pattern database, Figure 9 is a data diagram of the history database, Figure 10 is a diagram of the initial weight data of the neural network, and Figure 11 is the weight data of j after initial learning of the neural network. Figure 12 is a weight data diagram after learning pseudo pattern No. 5, and Figure 13 is a data diagram of the collected pattern file. Fig. 14 is a neuron input/output relationship diagram, Fig. 15 is a sigmoid function diagram, Fig. 16 is an explanatory diagram of the pattern learning phase according to the conventional technique, Fig. 17 is an explanatory diagram of the signal processing phase according to the conventional technique, and Fig. 18 is an explanatory diagram of the signal processing phase according to the conventional technique. FIG. 2 is an explanatory diagram of signal processing of an unknown pattern according to the prior art. DESCRIPTION OF SYMBOLS 1...Learning means, 2...Learning pattern storage means, 3...Pseudo pattern generation means, 4...Switching means.

Claims (1)

[Claims] 1) A learning means (1) that performs self-learning and recognizes an input signal when the input signal is added;
learning pattern storage means (2) for storing patterns to be learned by the learning pattern storage means (2); generating pseudo learning patterns by adding specific variations to the patterns to be learned in the learning pattern storage means (2); 1. A pattern learning method characterized by comprising, in addition to 1), a pseudo pattern generating means (3) for causing the learning means (1) to learn. 2) Learning means (1) that performs self-learning and recognizes input signals when input signals are added;
learning pattern storage means (2) for storing patterns to be learned; and a learning means (1) for generating a pseudo learning pattern by adding a specific variation to the pattern to be learned in the learning pattern storage means 2; In addition, a pseudo pattern generating means (3) for making the learning means (1) learn; A pattern learning method comprising: switching means (4) for selectively applying a signal to the learning means. 3) The switching means (4) selects the pseudo learning pattern and adds it to the learning means (1) when the input signal is a predetermined idle signal for a specific time. The pattern learning method described in 2. 4) Learning means (1) that performs self-learning and recognizes input signals when input signals are added;
learning pattern storage means (2) for storing patterns to be learned; and a learning means (1) for generating a pseudo learning pattern by adding a specific variation to the pattern to be learned in the learning pattern storage means 2; In addition, a pseudo pattern generation means (3) for causing the learning means (1) to learn, and when it is detected that the pseudo learning pattern generated by the pseudo pattern generation means (3) is inappropriate as a learning pattern. , a first checking means that does not add the pattern to the learning means (1). 5) Learning means (1) that performs self-learning and recognizes input signals when input signals are added;
learning pattern storage means (2) for storing patterns to be learned; and a learning means (1) for generating a pseudo learning pattern by adding a specific variation to the pattern to be learned in the learning pattern storage means 2; In addition, a pseudo pattern generating means (3) for causing the learning means (1) to learn, and a pseudo pattern generating means (3) for causing the learning means (1) to
) and a second checking means that prevents learning of the pattern when the obtained output pattern differs from the expected output pattern. 6) The learning pattern storage means (2) displays the stored learning pattern on a screen etc. so that the user can judge whether or not the learning pattern is truly appropriate. 2. The pattern learning method according to claim 1, further comprising a conversation function for updating information. 7) The pseudo pattern generation means (3) temporarily stores the automatically generated pseudo learning patterns in an internal buffer, and outputs them all at once when a certain number of patterns are accumulated or when there is no external input. 2. The pattern learning method according to claim 1, wherein the pseudo learning pattern can be generated independently of the state of external input. 8) When there is an external input, the learning means (1) calculates the input external input pattern by
By comparing the patterns stored in the learning pattern storage means (2) or with conditions set from the outside in advance, the learning patterns are automatically collected as learning patterns and stored in the learning pattern storage means (2). The pattern learning method according to claim 1, characterized in that: