JP2009043122A - Data processing apparatus, data processing method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To appropriately perform a processing considering errors. <P>SOLUTION: A main body learning module 101 learns data D. An error learning module 102 learns errors ΔD between the data D and reconfigured data D<SB>R</SB>resulting from reconfiguring the data D by the main body learning module 101. The main body learning module 101 reconfigures data D upon input of the data D and outputs reconfigured data D<SB>R</SB>, and the error learning module 102 reconfigures error data ΔD between the data D and the reconfigured data D<SB>R</SB>and outputs reconfigured error data ΔD<SB>R</SB>. Then a learning module 100 adds the reconfigured data D<SB>R</SB>and the reconfigured error data ΔD<SB>R</SB>and outputs the addition result. This invention is applicable to, for example, a robot and the like. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a data processing device, a data processing method, and a program, and more particularly, to a data processing device, a data processing method, and a program that perform processing that appropriately considers an error.

  For example, a robot that performs tasks autonomously can be realized by using an input / output system that outputs output data to be output in response to input data.

  Here, the input data and the output data may be any physical quantities as long as they can be observed.

  For example, in an input / output system used for a robot that detects a physical quantity using a sensor and takes an action according to the detection result, the sensor detects the physical quantity and outputs the sensor. Data or the like becomes input data. In order to drive an actuator or the like necessary for the action of the robot, action data or the like supplied to the actuator or the like becomes output data.

  More specifically, for example, when recognizing a voice and outputting a synthesized sound corresponding to the recognition result (a synthesized sound as a response to the voice), voice data output by a microphone (sensor) that collects the voice, The cepstrum and other characteristic parameters extracted from the speech data are input data, and the parameters used for the generation of synthesized sound (speech synthesis) according to the recognition result are output data.

  Further, for example, when recognizing a voice or an image and moving the arm or head of the robot according to the recognition result, the voice data output by the microphone that collects the voice, the feature parameter of the voice data, and the image The image data output by the camera that captures the image, the motion vector detected from the image data, and other features are used as input data, and control data that controls the actuator that drives the arm and head according to the recognition result is output. It becomes data.

  Further, the input data and the output data do not need to be data of different modalities such as the sensor data and the action data described above.

  That is, for example, when a section of data is divided into sections of a predetermined length, each section is referred to as a frame, and a t-th (time t) frame from the beginning is referred to as a frame #t. In the output system, sensor data of frame #t can be used as input data, and sensor data of frame # t + 1 can be used as output data.

  In this case, in the input / output system, sensor data of the next frame # t + 1 as output data having the same modality as the input data is predicted from sensor data of frame #t as input data.

  Further, for example, as input data, a combination of sensor data of frame #t and action data of frame #t, which is a combination of data of different modalities, is adopted, and output data of frame # t + 1 is adopted. A combination of sensor data and action data of frame # t + 1 can be employed.

  In this case, in the input / output system, both sensor data and action data of the next frame # t + 1 are predicted from the sensor data and action data of frame #t.

  By the way, in an input / output system, it is necessary to output the output data which should be output with respect to the input data.

  That is, for example, as described above, when voice is recognized and the arm of the robot is moved according to the recognition result, the voice data is used as input data and control data for controlling an actuator that drives the arm is output. When the data is used, the voice data is voice-recognized (identified) in response to the voice data input, and further, control data for executing the arm movement according to the voice recognition result is generated (obtained). There is a need.

  For example, as described above, when the sensor data and action data of frame #t are used as input data, and the sensor data and action data of frame # t + 1 are used as output data, the sensor of frame #t For the input of data and action data, the sensor data and action data of the frame #t are recognized (identified), and the sensor data and action data (predicted value) of the next frame t + 1 from the recognition result. Must be generated.

  As described above, the input / output system recognizes (identifies) input data and acquires (generates) and outputs output data to be output with respect to the input data (hereinafter referred to as recognition generation as appropriate). ) Is required.

  In order for the input / output system to respond to input of input data, perform recognition generation using the input data, and output output data to be output with respect to the input data, for example, Prepare many sets of input data (data to be) and output data to be output (data to be) for the input data (hereinafter referred to as teaching data as appropriate) It is necessary to learn the input data and the output data, and learn the relationship between the input data and the output data to be output with respect to the input data.

  Accordingly, the applicant of the present application has a node for input data SOM (Self-Organization Map: hereinafter referred to as input SOM as appropriate) and an output data SOM (hereinafter referred to as output SOM as appropriate). ) Is used to learn input data and output data, and to learn the relationship between input data and output data to be output with respect to the input data. A method of performing learning without learning has been previously proposed (see, for example, Patent Document 1).

  In addition, the applicant changes the parameter of the node of the output SOM in the input / output relation model, generates voice data as output data from the parameter after the change, and uses the voice data as input data for the input SOM. Is repeated until the recognition result matches the recognition result at the input SOM of the voice data given from the outside. The method of performing unsupervised learning for learning to perform listening imitation (hereinafter, also referred to as listening imitation learning as appropriate) that outputs speech data that imitates as output data has been proposed (for example, Patent Document 2).

  Here, learning in pattern recognition that learns and recognizes patterns can be divided into supervised learning and unsupervised learning.

  As a learning method of supervised learning, for example, there is a learning method using a learning model such as template matching, NN (Neural Network), HMM (Hidden Markov Model) or the like. In supervised learning, each class learns a category belonging to (belonging to) a class by giving information called a correct answer label to which class learning data (data used for learning) belongs. Data learning is performed.

  That is, in supervised learning, learning data used for learning is prepared for each predetermined category (for example, each phoneme category, each phoneme category, each word category, etc.) and used for learning. A learning model (a learning model for learning learning data of each category) is also prepared for each predetermined class.

  In supervised learning, learning of the learning model of the class represented by the correct label is performed using only the learning data of the category belonging to (belonging to) that class, and as a result, the template for each class is based on the correct answer label. (The learning model of the class represented by the correct answer label) can be obtained.

  Therefore, in pattern recognition using a learning model obtained by supervised learning, the correct label of the template (the template with the highest likelihood) that best matches the data to be recognized is recognized. As a result, it can be output.

  On the other hand, as an unsupervised learning method, there is a learning method using NN, SOM, or the like. Unsupervised learning differs greatly from supervised learning in that a correct answer label is not given.

  That is, unsupervised learning is also referred to as self-organization (self-organized learning), and is performed in a situation where a correct answer label is not given to each pattern of learning data. According to unsupervised learning, for example, classes of learning data having similar patterns (classes to which the category of learning data belongs) are placed close to each other in a space in which the distance between classes can be defined.

  Pattern recognition can be viewed as quantization of a signal space in which data (signals) to be recognized to be recognized by pattern recognition is observed.

JP 2006-162898 A JP 2006-201665 A

  In the input / output relation model that has been proposed earlier, the input SOM is used to learn the input data that makes up the teaching data, and the input data pattern is obtained at each node of the input SOM. The output data constituting the teaching data is learned, and the pattern of the output data is obtained at each node of the output SOM. Furthermore, when the input data constituting the teaching data is given (input) to the input SOM, the node of the input SOM that becomes the winner node and the output data set with the input data (the teaching data together with the input data When the output data to be configured is given to the output SOM, the connection with the node of the output SOM that becomes the winner node is strengthened.

  In the recognition generation, a winner node for the input data is determined from the nodes of the input SOM, and the combination of the recognition process for making the winner node the recognition result of the input data and the winner node as the recognition result is performed. The node of the strongest output SOM is determined as the node used to generate the output data to be output with respect to the input data (hereinafter referred to as the generation node as appropriate), and the output of the pattern acquired by the generation node A generation process for acquiring (generating) an estimated value of data is performed.

  The input / output relationship model is composed of two SOMs, an input SOM and an output SOM, but recognition generation uses not a single input / output relationship model composed of such two SOMs, but a single SOM. It is also possible to do this.

  That is, in recognition generation using one SOM, the winner node is determined as a recognition process for data input, and further, the pattern data acquired by the winner node is acquired ( Generation) is performed as a generation process.

  Here, in recognition generation using an input / output relation model composed of two SOMs, the input data recognized in the recognition process and the output data acquired (generated) in the generation process have the same modality. It may be data or data of different modalities.

  On the other hand, in recognition generation using one SOM, data recognized by the recognition process and data acquired (generated) by the generation process are data of the same modality (data of the same physical quantity). It becomes.

  Now, for the sake of simplicity, taking recognition generation using one SOM as an example, in the recognition processing, a winner node is determined from a finite number of nodes for data input, and generation processing is performed. Then, the data of the pattern acquired by the winner node is output. The pattern data acquired by the winner node is referred to as reconstructed data obtained by reconstructing the data recognized in the recognition process. Can do.

  Furthermore, for the sake of simplicity, assuming that a vector is given as a parameter for each node in the SOM, determining the winner node in the recognition process is equivalent to (vector) quantization, Outputting pattern data acquired by the winner node in the generation process is equivalent to (vector) inverse quantization. Therefore, according to the recognition generation process, a so-called quantization error occurs.

  For example, the quantization error increases the size of the SOM, that is, increases the number of nodes that make up the SOM, or increases the dimension of the SOM topology, that is, the dimension in which the SOM nodes are placed. By increasing the number, it can be reduced equivalently by reducing the so-called quantization step.

  However, for example, even if the number of nodes constituting the SOM is increased, the number of reconstructed data considering the quantization error only increases by the number of increased nodes, that is, the number of nodes of the SOM. Is increased from N1 by N2 to (N1 + N2), the number of reconstructed data containing only N1 quantization errors can be output by N2. N1 + N2) reconstruction data including (small) quantization errors (reconstruction data considering (N1 + N2) quantization errors) can only be output, and quantization is performed. It is hard to say that errors are properly taken into account.

  The present invention has been made in view of such a situation, and makes it possible to perform processing that appropriately considers errors.

  A data processing apparatus according to a first aspect of the present invention is a data processing apparatus that performs a learning process for learning data, a main body learning module that learns data, the data, and the main body learning module that reconfigures the data An error learning module that learns a division error in which an error from the reconstructed data is divided every predetermined dimension or time.

  The data processing method according to the first aspect of the present invention is a data processing method of a data processing apparatus that performs a learning process for learning data. In a main body learning module that learns data, the data is learned, and the data, The error learning module that learns a division error in which an error from the reconstructed data obtained by reconstructing the data by the main body learning module is divided every predetermined dimension or every time includes learning the division error.

  The program according to the first aspect of the present invention is a program that causes a computer to function as a data processing device that performs learning processing for learning data. A main body learning module that learns data, the data, and the main body learning module include: The computer is caused to function as a data processing device including a learning module having an error learning module that learns a division error in which an error from the reconstructed data obtained by reconstructing the data is divided every predetermined dimension or time.

  In the first aspect of the present invention, data is learned by a main body learning module that learns data, and the data is reconstructed by the error learning module and the data is reconstructed by the main body learning module. A division error in which the error is divided every predetermined dimension or time is learned.

  The data processing device according to the second aspect of the present invention is the main body learning module that performs learning of input data and learning of output data to be output with respect to the input data in a data processing device that outputs data. Learning of divided input errors in which an error between the input data and reconstructed input data obtained by reconstructing the input data by the main body learning module is divided for each predetermined dimension or time, and the output data And an error learning module that learns a divided output error in which an error between the main body learning module and the reconstructed output data obtained by reconfiguring the output data is divided every predetermined dimension or time The main body learning module reconstructs the input data, outputs the reconstructed input data, and The output data to be output is reconstructed and the reconstructed output data is output, and the error learning module reconstructs the divided output error corresponding to the divided input error and outputs a divided reconstructed output error. The learning module adds the reconstruction output data and the partial correction data determined based on the divided reconstruction output error and outputs the result.

  A data processing method according to a second aspect of the present invention is a data processing method for a data processing apparatus that outputs data, wherein the data processing apparatus learns input data and outputs to be output for the input data. Division in which an error between the main body learning module that has performed data learning, the input data, and reconstructed input data in which the main body learning module reconstructs the input data is divided for each predetermined dimension or time Learning of input error and learning of divided output error in which an error between the output data and the reconstructed output data obtained by reconstructing the output data by the main body learning module is divided for each predetermined dimension or time. A learning module having a performed error learning module, wherein the input learning data is reconstructed in the main body learning module, Output the generated input data, reconstruct output data to be output with respect to the input data, output the reconstructed output data, and in the error learning module, the divided output error corresponding to the divided input error And the divided reconstruction output error is output, and the learning module adds and outputs the divided reconstruction output data and the partial correction data determined based on the divided reconstruction output error. Includes steps.

  A program according to a second aspect of the present invention is a program that causes a computer to function as a data processing device that outputs data, and performs learning of input data and learning of output data to be output with respect to the input data. Learning the divided input error in which the error between the main body learning module, the input data, and the reconstructed input data obtained by reconstructing the input data by the main body learning module is divided for each predetermined dimension or time; And an error learning module that learns a divided output error in which an error between the output data and reconstructed output data obtained by reconstructing the output data by the main body learning module is divided for each predetermined dimension or time. The main body learning module reconstructs the input data, and reconstructs the input data. Output data, reconstruct output data to be output with respect to the input data, output the reconstructed output data, and the error learning module reconstructs the divided output error corresponding to the divided input error. A data processing device configured to output a divided reconstruction output error, wherein the learning module adds and outputs the reconstruction output data and partial correction data determined based on the divided reconstruction output error As a computer to function.

  In the second aspect of the present invention, the body learning module reconstructs the input data and outputs the reconstructed input data, and reconstructs output data to be output with respect to the input data. Configuration output data is output. In the error learning module, the divided output error corresponding to the divided input error is reconstructed to output a divided reconstructed output error, and the divided reconstructed output data and the divided reconstructed output are output in the learning module. The partial correction data determined based on the error is added and output.

  A data processing device according to a third aspect of the present invention is a data processing device that outputs data, a main body learning module that has learned data, the data, and reconstructed data in which the main body learning module reconstructs the data And an error learning module that learns a division error in which an error is divided every predetermined dimension or time, and the main body learning module reconstructs the data to reconstruct the reconstructed data The error learning module reconstructs the division error and outputs a division reconstruction error, and the learning module is a portion determined based on the reconstruction data and the division reconstruction error. Add the correction data and output.

  A data processing method according to a third aspect of the present invention is a data processing method of a data processing apparatus that outputs data, wherein the data processing apparatus includes a main body learning module that has learned data, the data, and the main body learning module. Comprises a learning module having an error learning module that learns a division error in which an error from the reconstructed data obtained by reconstructing the data is divided for each predetermined dimension or time. In the main body learning module, the data And the reconstruction data is output in the error learning module, the division error is output in the error learning module, and the reconstruction data and the division are output in the learning module. A step of adding and outputting the partial correction data determined based on the reconstruction error.

  A program according to a third aspect of the present invention is a program that causes a computer to function as a data processing device that outputs data. An error learning module that learns a division error in which an error from the configured reconstruction data is divided for each predetermined dimension or time, and the main body learning module reconstructs the data. The reconstruction data is output, the error learning module reconstructs the division error and outputs a division reconstruction error, and the learning module is based on the reconstruction data and the division reconstruction error. The computer is caused to function as a data processing device that adds and outputs the partial correction data determined in this manner.

  In the third aspect of the present invention, the main learning module reconstructs the data and outputs the reconstructed data, and the error learning module reconstructs the division error and outputs a division reconstruction error. The learning module adds and outputs the reconstruction data and the partial correction data determined based on the division reconstruction error.

  The program can be provided by being transmitted via a transmission medium or by being recorded on a recording medium.

  Further, the data processing device may be an independent device or an internal block constituting the independent device.

  According to the first to third aspects of the present invention, it is possible to perform processing that appropriately considers an error.

  Embodiments of the present invention will be described below. Correspondences between the constituent elements of the present invention and the embodiments described in the specification or the drawings are exemplified as follows. This description is intended to confirm that the embodiments supporting the present invention are described in the specification or the drawings. Therefore, even if there is an embodiment that is described in the specification or the drawings but is not described here as an embodiment that corresponds to the constituent elements of the present invention, that is not the case. It does not mean that the form does not correspond to the constituent requirements. Conversely, even if an embodiment is described here as corresponding to a configuration requirement, that means that the embodiment does not correspond to a configuration requirement other than the configuration requirement. It's not something to do.

  The data processing apparatus according to the first aspect of the present invention includes a main body learning module (for example, FIG. 4) that learns data in a data processing apparatus (for example, the input / output system of FIG. 1) that performs learning processing for learning data. An error learning module (learning a division error in which an error between the main body learning module 101), the data, and reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time. For example, a learning module having the error learning module 102) of FIG. 4 is provided.

  A data processing method according to a first aspect of the present invention is a data processing method of a data processing apparatus (for example, the input / output system of FIG. 1) that performs a learning process for learning data. A division in which the data is learned (for example, step S131 in FIG. 16), and an error between the data and reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time. The error learning module for learning the error includes a step of learning the division error (for example, step S135 in FIG. 16).

  A program according to the first aspect of the present invention is a main body learning module that learns data in a program that causes a computer to function as a data processing device (for example, the input / output system in FIG. 1) that performs learning processing to learn data. For example, a division error in which an error between the main body learning module 101 in FIG. 4 and the data and reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time is learned. The computer is caused to function as a data processing apparatus including a learning module having an error learning module (for example, the error learning module 102 in FIG. 4).

  A data processing apparatus according to a second aspect of the present invention is a data processing apparatus that outputs data (for example, the input / output system of FIG. 1), learns input data, and outputs to be output for the input data. An error between the main body learning module that performed data learning (for example, the main body learning module 101 in FIG. 18), the input data, and reconstructed input data obtained by reconfiguring the input data by the main body learning module is a predetermined error. Learning of divided input error divided for each dimension or time, and the error between the output data and the reconstructed output data obtained by reconstructing the output data by the main body learning module for each predetermined dimension or time A learning module having an error learning module (for example, the error learning module 102 in FIG. 18) that has learned the divided output error. The main body learning module reconstructs the input data, outputs the reconstructed input data, reconstructs output data to be output with respect to the input data, and reconstructs the reconstructed output data. The error learning module reconstructs the divided output error corresponding to the divided input error and outputs a divided reconstructed output error, and the learning module outputs the reconstructed output data and the divided output data. The partial correction data determined based on the reconstruction output error is added and output.

  A data processing method according to a second aspect of the present invention is a data processing method for a data processing apparatus that outputs data, wherein the data processing apparatus learns input data and outputs to be output for the input data. An error between the main body learning module that performed data learning (for example, the main body learning module 101 in FIG. 18), the input data, and reconstructed input data obtained by reconfiguring the input data by the main body learning module is a predetermined error. Learning of divided input error divided for each dimension or time, and the error between the output data and the reconstructed output data obtained by reconstructing the output data by the main body learning module for each predetermined dimension or time And an error learning module (for example, the error learning module 102 in FIG. 18) that performs learning of the divided output error. The main body learning module reconstructs the input data, outputs the reconstructed input data (for example, step S283 in FIG. 27), and outputs output data to be output with respect to the input data. Reconstructing and outputting the reconstructed output data (for example, step S287 in FIG. 27), the error learning module reconstructs the divided output error corresponding to the divided input error, and performs divided reconstructed output. An error is output (for example, step S288 in FIG. 27), and the learning module adds and outputs the divided reconstruction output data and the partial correction data determined based on the divided reconstruction output error ( For example, step S290) of FIG. 27 is included.

  A program according to a second aspect of the present invention is a program that causes a computer to function as a data processing device that outputs data, and performs learning of input data and learning of output data to be output with respect to the input data. Further, an error between the main body learning module (for example, the main body learning module 101 of FIG. 18), the input data, and the reconstructed input data obtained by reconstructing the input data by the main body learning module is every predetermined dimension or time. Learning of divided input error divided into two, and an error between the output data and reconstructed output data obtained by reconstructing the output data by the main body learning module is divided for each predetermined dimension or time. An error learning module (for example, the error learning module 102 in FIG. 18) that performs learning of the divided output error. A learning module, wherein the main body learning module reconstructs the input data, outputs the reconstructed input data, reconstructs output data to be output with respect to the input data, and reconstructs the output Data is output, the error learning module reconstructs the divided output error corresponding to the divided input error and outputs a divided reconstructed output error, and the learning module includes the reconstructed output data, The computer is caused to function as a data processing apparatus that adds and outputs the partial correction data determined based on the divided reconstruction output error.

A data processing device according to a third aspect of the present invention is a data processing device that outputs data, a main body learning module that learns data (for example, the main body learning module 101 in FIG. 4), the data, and the main body learning. Learning having an error learning module (for example, the error learning module 102 in FIG. 4) in which a module learns a division error in which an error from reconstructed data obtained by reconstructing the data is divided every predetermined dimension or time. The main body learning module reconstructs the data and outputs the reconstructed data, the error learning module reconstructs the division error and outputs a division reconstruction error, The learning module adds and outputs the reconstruction data and the partial correction data determined based on the division reconstruction error.

  A data processing method according to a third aspect of the present invention is a data processing method for a data processing device that outputs data, wherein the data processing device is a main body learning module that has learned data (for example, the main body learning module 101 of FIG. 4). ) And the data, and an error learning module (for example, FIG. 4) that learns a division error in which an error between the data and the reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time. Error learning module 102), and the main body learning module reconstructs the data and outputs the reconstructed data (for example, step S142 in FIG. 17). Then, the division error is reconstructed and the division reconstruction error is output (for example, step S145 in FIG. 17). In the learning module, including the the reconfiguration data, the divided reconstituted by adding the partial compensation data is determined based on the error output (e.g., step S147 in FIG. 17) the step.

  The program according to the third aspect of the present invention includes a main body learning module (for example, the main body learning module 101 in FIG. 4) that has learned data in a program that causes a computer to function as a data processing device that outputs data, and the data And an error learning module (for example, the error learning module 102 in FIG. 4) that learns a division error in which an error between the main body learning module and the reconstructed data obtained by reconstructing the data is divided for each predetermined dimension or time. The main body learning module reconstructs the data and outputs the reconstructed data, and the error learning module reconstructs the division error to obtain a division reconstruction error. And the learning module is determined based on the reconstruction data and the division reconstruction error As a data processing apparatus for adding and outputting the partial correction data, it causes the computer to function.

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

  FIG. 1 is a block diagram showing a configuration example of an embodiment of an input / output system as a data processing apparatus to which the present invention is applied.

  The input / output system in FIG. 1 performs a learning process for learning data, and further, a recognition generation (cognitive behavior) process for recognizing data based on the learning result of the data and outputting the data according to the recognition result. For example, it learns robots that act autonomously in response to external stimuli (for example, talking from people, etc.), user operations, etc., and according to the learning results The present invention can be applied to home appliances represented by AV (Audio Visual) equipment that performs processing.

  The input system in FIG. 1 includes a learning module 100, and the learning module 100 includes a main body learning module 101, an error learning module 102, and a control unit 103.

  The main body learning module 101 is supplied with data D to be subjected to learning processing and data D to be subjected to recognition processing of recognition generation processing.

The main body learning module 101 learns data D to be subjected to learning processing. The main body learning module 101, depending on the learning result, with respect to the input data D to be recognition processing, reconstructs the data D, and outputs the reconstructed data D _R the resulting.

The error learning module 102, the data D to be supplied to the main learning module 101, body learning module 101, an error between the reconstructed data D _R to be output to the supplied data D error data △ D ( = DD _R ) is supplied.

The error learning module 102 learns error data ΔD. The error learning module 102, depending on the learning result, with respect to the input of the error data △ D, reconstructs the error data △ D, and outputs the reconstructed error data △ D _R the resulting.

  The control unit 103 controls the main body learning module 101 and the error learning module 102 as necessary. Further, the control unit 103 generates data to be given to the main body learning module 101 and the error learning module 102 as necessary, and supplies the data to the main body learning module 101 and the error learning module 102.

That is, for example, the control unit 103, the data D to be supplied to the main learning module 101, for the data D, by using the reconstructed data D _R of the main body learning module 101 outputs the error data △ D Generated and supplied to the error learning module 102.

  In the input / output system of FIG. 1 configured as described above, as described above, the data is recognized based on the learning process for learning the data and the learning result of the data, and the data is determined according to the recognition result. The recognition generation process to be output is performed.

  The learning process performed by the input / output system of FIG. 1 will be described with reference to the flowchart of FIG.

  This learning process is started in response to the supply of data D to be subjected to the learning process to the main body learning module 101 of the learning module 100.

  In step S101, the main body learning module 101 performs learning of the supplied data D (learning of a learning model using the data D).

In step S102, the main learning module 101 generates and outputs the reconstructed data D _R by reconstructing the data D used for learning in step S101 immediately before.

In step S103, the control unit 103, for the data D supplied to the main body learning module 101 in the previous step S101, and the reconstructed data D _R output by the main body learning module 101 in the previous step S102 for the data D. with bets, calculates an error data △ D = DD _R, supplies the error data △ D is the calculation result to the error learning module 102.

  In step S104, the error learning module 102 learns the error data ΔD supplied from the control unit 103 (learning of a learning model using the error data ΔD). This completes the learning process.

  Note that the learning process in FIG. 2 is performed, for example, every time data D to be subjected to the learning process is input to the learning module 100.

  Next, the recognition generation process performed by the input / output system of FIG. 1 will be described with reference to the flowchart of FIG.

  This recognition generation process is started in response to the input of data D to be recognized by the main body learning module 101 of the learning module 100.

  In step S111, the main body learning module 101 performs recognition processing for recognizing the supplied data D based on the learning result of the learning processing in FIG.

In step S112, the main learning module 101, based on the recognition result data D recognized in step S111 immediately before, and generates and outputs the reconstructed data D _R by reconstructing the data D.

In step S113, the control unit 103 outputs the data D supplied to the main body learning module 101 in the immediately preceding step S111 and the reconfiguration data D _R output from the main body learning module 101 in the immediately previous step S112 with respect to the data D. with bets, calculates an error data △ D = DD _R, and supplies a calculation result △ a D to error learning module 102.

In step S114, the error learning module 102 performs a recognition process for recognizing the error data ΔD from the control unit 103 based on the learning result of the learning process in FIG. In step S115, error learning module 102, based on the recognition result of the error data △ D recognized in step S114 immediately before, and generates and outputs a reconstructed error data △ D _R reconstituted the error data △ D.

In step S116, the control unit 103 adds the reconstruction data D _R output from the body learning module 101 and the reconstruction error data ΔD _R output from the error learning module 102, so that the body learning module 101 outputs the result. The reconstructed data D _R corrected is output.

That is, the control unit 103 obtains addition data D _R + ΔD _R obtained by adding the reconstruction data D _R output from the main body learning module 101 and the reconstruction error data ΔD _R output from the error learning module 102, The added data is output as final reconstructed data, that is, the reconstructed data D _R output from the main body learning module 101 is corrected. This completes the recognition generation process.

As described above, in the main learning module 101 learns the data D, the error learning module 102, apart from the training data D, the error data with the reconstructed data D _R reconstructed data D and the data D △ D is learned. Further, the main learning module 101, and reconstructs the data D, and outputs the reconstructed data D _R, the error learning module 102 reconfigures the error data △ D between the data D and the reconstructed data D _R , and it outputs the reconstructed error data △ D _R. Then, the control unit 103 adds the reconstruction data D _R and the reconstruction error data ΔD _R and outputs the addition data obtained as a result. Therefore, processing that appropriately considers the error, that is, addition data D _R + ΔD _{R obtained} by correcting the error estimated to occur in the reconstruction data D _R by the reconstruction error data ΔD _R is obtained. be able to.

  Note that both the learning process of FIG. 2 and the recognition generation process of FIG. 3 can be performed on the input of the data D, or only one of them can be performed.

  Next, FIG. 4 shows a first configuration example of the main body learning module 101, the error learning module 102, and the control unit 103 of FIG.

  In FIG. 4, the main body learning module 101 is from the calculation unit 121 and the main body model storage unit 122, the error learning module 102 is from the calculation unit 123 and the divided error model storage unit 124, and the control unit 103 is the error calculation unit 125, data Each of the division unit 126, the correction data generation unit 127, and the addition unit 128 is configured.

  For example, data D is supplied to the calculation unit 121 from the outside. This data D is assumed to be an n-dimensional vector or time series data. The calculation unit 121 learns the data D supplied thereto. That is, the calculation unit 121 learns the main body model stored in the main body model storage unit 122 using the data D supplied thereto.

In addition, the calculation unit 121 recognizes the supplied data D using the main body model stored in the main body model storage unit 122, reconstructs the data D based on the recognition result, and obtains the result. and it outputs the reconstructed data D _R. Reconstruction data D _R by the calculation unit 121 outputs is supplied to the error calculating unit 125 and the addition unit 128 of the control unit 103.

  The main body model storage unit 122 stores a main body model as a model for learning data (learning model). As the main body model, for example, SOM, NN, or other learning models capable of unsupervised learning can be adopted.

The operation unit 123 of the error learning module 102, the divided error data △ D ₁ from the data dividing unit 126 of the control unit 103, △ D _2, · · ·, the [Delta] D _n is supplied. The arithmetic unit 123 learns the division error data ΔD ₁ (i = 1, 2,..., N) from the data division unit 126. That is, the calculation unit 123 uses the division error data ΔD _i from the data division unit 126 to learn the error model stored in the division error model storage unit 124.

In addition, the calculation unit 123 recognizes the division error data ΔD _i from the data division unit 126 using the division error model stored in the division error model storage unit 124 and, based on the recognition result, the division error data. Data ΔD _i is reconstructed, and divided reconstruction error data ΔD _R1 , ΔD _R2 ,..., ΔD _Rn obtained as a result are output. The division reconstruction error data ΔD _Ri (i = 1, 2,..., N) output from the calculation unit 123 is supplied to the correction data generation unit 127 of the control unit 103.

  The division error model storage unit 124 stores a division error model as a model for learning data. As the division error model, for example, SOM, NN, and other learning models capable of unsupervised learning can be adopted.

The error calculating section 125 of the control unit 103, as described above, except that reconstruction data D _R is supplied from the operation unit 121, the same data D and the data D to be supplied to the arithmetic unit 121 is supplied from the outside The

Error calculation unit 125 obtains the data D from the outside, the difference DD _R between the reconstructed data D _R, the error data △ D, to the data dividing unit 126. As described above, when the data D is an n-dimensional vector or time series data, the error data ΔD is also an n-dimensional vector or time series data.

The data dividing unit 126 divides the error data ΔD, which is an n-dimensional vector or time series data, every predetermined dimension or every predetermined time, and the resulting divided error data ΔD _i is obtained. Output. For example, when the error data ΔD is a two-dimensional vector, the error data ΔD is divided into ΔD ₁ and ΔD ₂ . Further, for example, error data △ D is, if a time series data of 100 steps, the error data △ D, divides △ D ₁ from one step to 50 steps from 51 Step △ D ₂ to 100 steps . The number n to be divided is arbitrary regardless of the number of dimensions and the number of steps of the error data ΔD. The division error data ΔD _i output from the data division unit 126 is supplied to the calculation unit 123 and the correction data generation unit 127.

The correction data generation unit 127 determines the dimension or time to be corrected of the reconstructed data D _R based on the division error data ΔD _i from the data division unit 126, and for each predetermined dimension divided or according to the determination result. By determining a correction weight for each time and multiplying the division reconstruction error data ΔD _Ri from the calculation unit 123 by the correction weight, partial correction data _{ΔD Rj} is generated and supplied to the adder 128.

Adding unit 128, a reconstruction data D _R to be supplied from the arithmetic unit 121, the correction by adding the partial correction data △ D _Rj supplied from the data generation unit 127, the resulting sum data D _R + △ D _Rj and the reconstruction data D _R corrected final reconstruction data (hereinafter, also referred to as the final reconstruction data) is output as.

  Here, in FIG. 4, the same learning model can be adopted as the main body model stored in the main body model storage unit 122 and the error model stored in the division error model storage unit 124, or different. A learning model can also be adopted.

  However, here, in order to simplify the explanation, the same learning model is adopted as the main body model and the error model.

Note that the main body model is a learning model that performs learning on the data that should be learned originally, whereas the division error model is obtained by reconstructing certain data D using the main body model that has learned the data. This is a learning model that calculates an error (error data ΔD) between the reconstructed data D _R and the data D and performs learning on the divided error data divided for each predetermined dimension or time. Therefore, the main body model and the division error model differ only in the learning target, and there is no essential difference as a learning model.

  Next, as the main body model and the division error model, for example, SOM can be adopted as described above.

  As the main body model and the division error model, a vector is given to the node, and a fixed-length pattern as well as a general SOM that can acquire only a fixed-length pattern (a constant-length pattern). In addition, it is possible to employ a SOM (hereinafter referred to as a pattern storage SOM as appropriate) that can acquire a variable-length pattern (a pattern whose length is not constant) such as a time-series pattern.

  First, the pattern storage SOM will be described.

  FIG. 5 schematically shows an example of the pattern storage SOM.

  The pattern storage SOM is composed of a plurality of nodes, and is common to general SOMs in that it can perform self-organized learning. Therefore, the pattern storage SOM is included in the category of the SOM, and is therefore referred to as a kind of SOM. be able to.

  However, the pattern storage SOM is an HMM that can learn a variable-length pattern such as a time-series pattern, NN, SVR, and other learning models that can acquire a pattern (hereinafter referred to as a pattern as appropriate). This is different from a general SOM in which a vector is given to a node in that a time series pattern or the like can be held by the pattern storage model.

  As described above, the pattern storage SOM is a SOM composed of a plurality of nodes having a pattern storage model capable of acquiring (representing) a variable-length pattern, and is equivalent to the number of nodes (classification is performed). Patterns can be stored.

In FIG. 5, the pattern storage SOM is composed of six nodes N _{1 to} N ₆ .

Each node N _i (in FIG. 5, i = 1, 2,..., 6) constituting the pattern storage SOM has, for example, a pattern storage model expressing a time series pattern. Further, the node N _i can have a coupling relationship with other nodes N _j (j = 1, 2,..., 6 in FIG. 5). This connection relationship is called a link. In FIG. 5, for example, the node N ₁ has a direct coupling relationship with the nodes N ₂ and N ₃ . Further, for example, the node N ₃ has a direct coupling relationship with the nodes N ₁ , N ₂ , N ₅ , and N _6, and therefore the nodes N ₅ and N ₆ are connected to the node N ₃ via the node N _3. and a N ₁ and indirect coupling relationship. Incidentally, the coupling relationship between the two nodes N _i and N _j, means the shortest binding relationship between the two nodes N _i and N _j.

  In learning of the pattern storage SOM (learning for storing (acquiring) a time-series pattern in the pattern storage SOM), time-series data is performed as learning data for learning. Further, as the learning data, time series data in which the type of category and the number of categories are unknown can be used. Therefore, unsupervised learning can be performed as the learning of the pattern storage SOM.

  Note that the learning of the pattern storage SOM is not necessarily performed so that one certain node corresponds to one certain category. That is, in the pattern storage SOM, learning may be performed so that one node corresponds to one category, or learning may be performed so that a plurality of nodes correspond to one category. Further, learning may be performed so that one node corresponds to a plurality of categories. Therefore, even if the learning data cannot be clearly categorized, learning by the pattern storage SOM can be performed.

Next, FIG. 6 shows a configuration example of a node N _i of pattern storage SOM schematically.

The node _Ni includes a pattern storage model 21 that can represent a variable-length pattern such as a time series pattern, and a learning data storage unit 22 that stores learning data used for learning the pattern storage model 21. The

Here, in FIG. 6, an HMM (continuous HMM) that is one of the state probability transition models is adopted as the pattern storage model 21. In FIG. 6, the HMM has left-to-right three states S ₁ , S ₂ , and S ₃ that have only a self-loop and a state transition to the next state (the state on the right). Note that circles in the pattern storage model 21 in FIG. 6 represent states, and arrows represent state transitions. Note that the HMM as the pattern storage model 21 is not limited to a left-to-right type or a three-state type.

  When the pattern memory model 21 is an HMM as shown in FIG. 6, the HMM as the pattern memory model 21 has a state transition probability and an output probability density function (when the HMM is a discrete HMM, it is a scalar quantity. The probability that a certain discrete symbol is output).

  The state transition probability is a probability of state transition in the HMM, and is given to each state transition indicated by an arrow in the pattern storage model 21 of FIG. The output probability density function represents the probability density of values observed from the HMM at the time of state transition. For example, a mixed normal distribution is adopted as the output probability density function. These HMM parameters (state transition probability and output probability density function) can be learned (estimated) by, for example, the Baum-Welch method.

In the node _Ni , the statistical characteristics of the learning data stored in the learning data storage unit 22, that is, the pattern of the learning data stored in the learning data storage unit 22 is learned in the pattern storage model 21, thereby The pattern storage model 21 and the learning data stored in the learning data storage unit 22 have a correspondence relationship.

The learning pattern storage SOM, thus, learning of the node N _i of pattern storage model 21, the pattern storage SOM, data to be learning is performed by on-line learning for learning for each given. Thus, the pattern parameter memory SOM, i.e., the parameter of the node N _i of pattern storage model 21 (when the pattern storage model 21 is HMM, as described above, the output probability density function state transition probability), the pattern Every time data to be learned is given to the stored SOM, it is updated little by little.

  That is, as will be described later, as the learning of the pattern storage SOM progresses, the learning data stored in the learning data storage unit 22 is updated with the data given to the pattern storage SOM, thereby changing little by little. Then, by learning the pattern storage model 21 with the learning data that changes little by little, the parameters of the pattern storage model 21 also change little by little.

  FIG. 7 schematically shows another example of the pattern storage SOM.

In FIG. 7, the pattern storage SOM is composed of nine nodes N _{1 to} N ₉ , and these nine nodes N _{1 to} N ₉ are two-dimensionally arranged. That is, in FIG. 7, nine nodes N _{1 to} N ₉ are arranged on the two-dimensional plane so that the horizontal × vertical is 3 × 3.

Further, in FIG. 7, links (coupling relationships) between the nodes adjacent in the horizontal direction and the nodes adjacent in the vertical direction among the nine nodes N _{1 to} N ₉ arranged two-dimensionally. Is given. By providing such links, the nodes constituting the pattern storage SOM are given an arrangement structure that is spatially arranged two-dimensionally.

  In the pattern memory SOM, the distance between two nodes on the space can be defined based on the spatial arrangement structure of the nodes given by the link. It can be used as an inter-pattern distance (similarity between patterns) of a pattern expressed by the pattern storage model 21 of each node.

  As the distance between the two nodes, for example, the number of links constituting the shortest path connecting the two nodes can be employed. In this case, when paying attention to a certain node, a node having a direct link with the target node (a node adjacent in the horizontal direction or vertical direction of the target node in FIG. 7) is the closest to the target node, A node that can be reached by following a previous link from a node having a direct link with the node becomes farther from the target node as the number of links to be reached increases.

  The links given to the nodes are not limited to those shown in FIGS. The links shown in FIG. 5 and FIG. 7 give a two-dimensional arrangement structure to the node, but the links give other one-dimensional arrangement structures, three-dimensional arrangement structures and the like. It may be.

  Next, FIG. 8 shows a configuration example of a learning apparatus that performs learning of the pattern storage SOM.

  In FIG. 8, the learning device includes a score calculation unit 41, a winner node determination unit 42, a weight determination unit 43, a learning data update unit 44, and a model learning unit 45.

  Time series data to be learned is input from the outside to the learning device, and the time series data from the outside is supplied to the score calculation unit 41, the weight determination unit 43, and the learning data update unit 44.

  The score calculation unit 41 obtains, for each node constituting the pattern storage SOM, a degree of adaptability of the node to the time-series data (observed value) from the outside, and supplies the score to the winner node determination unit 42. That is, when the pattern storage model 21 included in the node is an HMM as illustrated in FIG. 6, for example, the score calculation unit 41 performs time series from the outside from the HMM as the pattern storage model 21 included in the node. The likelihood that the data is observed is obtained and supplied to the winner node determination unit 42 as the score of the node.

  The winner node determination unit 42 obtains a node that best matches the external time-series data in the pattern storage SOM, and determines that node as a winner node.

  That is, the winner node determination unit 42 determines the node having the highest score from the score calculation unit 41 among the nodes constituting the pattern storage SOM as the winner node. Then, the winner node determination unit 42 supplies information representing the winner node to the weight determination unit 43.

  Here, a node label, which is a label for identifying each node, can be attached to the nodes constituting the pattern storage SOM. And a node label is employable as information showing a winner node and information showing other nodes. The node label is a label for identifying the node itself, and has nothing to do with the correct answer label indicating what the correct answer is.

  The weight determination unit 43 determines an update weight, which will be described later, for each node constituting the pattern storage SOM based on the winner node represented by the node label supplied from the winner node determination unit 42 and supplies the update weight to the learning data update unit 44. .

  That is, the weight determination unit 43 determines the update weight of each node (including the winner node) constituting the pattern storage SOM based on the inter-pattern distance between the node and the winner node, and the learning data update unit 44 Supply.

  Here, the pattern storage model 21 (FIG. 6) possessed by a node is updated using (new) time-series data supplied from the outside. The update weight of a node is the pattern storage model possessed by the node. By updating 21, the degree of the influence of the external time-series data received by the pattern storage model 21 is expressed. Therefore, if the update weight of a node is 0, the pattern storage model 21 that the node has is not affected (not updated) by the time-series data from the outside.

  The learning data update unit 44 updates the learning data stored in the learning data storage unit 22 (FIG. 6) included in each node of the pattern storage SOM.

  That is, the learning data update unit 44 uses the learning data already stored in the learning data storage unit 22 of the node and the time series data from the outside as the update weight of the corresponding node from the weight determination unit 43. Therefore, mixing is performed, and the mixing result is stored in the learning data storage unit 22 as new learning data, thereby updating the storage content of the learning data storage unit 22.

  As described above, when the learning data stored in the learning data storage unit 22 (FIG. 6) is updated according to the update weight, the learning data update unit 44 notifies the model learning unit that the update is completed. 45.

  When the model learning unit 45 receives an end notification from the learning data update unit 44, each node uses the learning data stored in the learning data storage unit 22 (FIG. 6) updated by the learning data update unit 44. The pattern storage model 21 is updated by learning the pattern storage model 21 that is included.

  Accordingly, the model learning unit 45 updates the pattern storage model 21 included in the node with the learning data stored in the learning data storage unit 22 (FIG. 6) included in the node and a new one from the outside. Based on time series data.

  Next, FIG. 9 shows a determination method for determining the update weight in the weight determination unit 43 of FIG.

  The weight determination unit 43 follows, for example, a curve (hereinafter referred to as a distance / weight curve) representing a relationship in which the update weight α decreases with respect to an increase in the inter-pattern distance d with the winner node as shown in FIG. The update weight (node update weight) α for the node is determined. According to the distance / weight curve, a node with a shorter inter-pattern distance d with the winner node determines a larger update weight α, and a node with a longer inter-pattern distance d determines a smaller update weight α.

  In the distance / weight curve of FIG. 9, the horizontal axis (from left to right) represents the update weight α, and the vertical axis (from top to bottom) represents the inter-pattern distance d.

In FIG. 9, along the vertical axis, the six nodes N _{1 to} N ₆ constituting the pattern storage SOM are located at positions (vertical axis positions) corresponding to the inter-pattern distance d between each node _Ni and the winner node. Are listed.

In FIG. 9, the inter-pattern distance d between the _six nodes N _{1 to} N ₆ constituting the pattern storage SOM and the winner node is closer in that order. Among the _six nodes N _{1 to} N ₆ constituting the pattern storage SOM, the node having the shortest pattern distance d to the winner node, that is, the node N ₁ having the zero pattern distance d to the winner node is , The winner node.

Here, if the pattern storage SOM has a two-dimensional arrangement structure as shown in FIG. 7 and the winner node is, for example, the node N ₆ , the winner node N ₆ and the node N ₆ Is the closest (1st) 0, and the inter-pattern distance d between the node N ₆ and the winner node N ₆ is also 0. The pattern of the winning node N _6, node N _3, N ₅ or N ₉ distances respectively, is 1 second closest, a node N _3, N ₅ or N ₉ respectively winning node N _6, The distance d is also 1.

Moreover, the pattern of the winning node N _6, node N _2, N ₄ or N ₈ distance respectively, is 2 close to the third, and the node N _2, N ₄ or N ₈ respectively winning node N _6, The distance d is also 2. Further, the distance between the winner node N ₆ and each of the nodes N ₁ or N ₇ is the farthest (fourth closest) 3, and the inter-pattern distance d between each of the nodes N ₁ or N ₇ and the winner node N ₆ is also 3

  Next, for example, a distance / weight curve as shown in FIG. 9 representing the relationship between the update weight α and the inter-pattern distance d is given by, for example, Expression (1).

^{α = G × γ (d /} △)
... (1)

  Here, in Expression (1), the coefficient G is a constant representing the update weight of the winner node, and the coefficient γ is a constant in the range of 0 <γ <1. In addition, the coefficient Δ is a coefficient (hereinafter referred to as an attenuation coefficient as appropriate) for adjusting a node range that affects new data (here, new time-series data from outside) in so-called neighborhood learning.

  As described above, 0 is given as the inter-pattern distance d of the node that is the winner node, and hereinafter, as the inter-pattern distance d of other nodes, according to the distance to the winner node, 1, 2,. In the equation (1), for example, if G = 8, γ = 0.5, Δ = 1, 8 (= G) is obtained as the update weight α of the node that is the winner node. . In the following, as the distance from the winner node increases, the update weight α of the node is determined to be 4, 2, 1,.

  Here, when the attenuation coefficient Δ in Equation (1) is a large value, the change in the update weight α with respect to the change in the inter-pattern distance d becomes moderate, and conversely, the attenuation coefficient Δ is a value close to 0. In this case, the change in the update weight α with respect to the change in the inter-pattern distance d becomes steep.

  Therefore, when the attenuation coefficient Δ is adjusted to gradually approach 0 from 1 as described above, for example, the change in the update weight α with respect to the change in the inter-pattern distance d becomes steep, and the update weight α is As the distance d increases, the value decreases. When the attenuation coefficient Δ is close to 0, the update weight α of the nodes other than the winner node (being a node) is almost 0.

  For example, the attenuation coefficient Δ can be set to a large value at the start of updating (learning) of the pattern storage SOM, and can be set to a small value as time passes, that is, as the number of updates increases. In this case, at the start of the update of the pattern storage SOM, the update weight α of each node of the pattern storage SOM is determined according to a distance / weight curve in which the change of the update weight α with respect to the change of the inter-pattern distance d is gentle. As (learning) progresses (advances), the update weight α of each node in the pattern storage SOM is determined according to the distance / weight curve in which the change of the update weight α with respect to the change in the inter-pattern distance d becomes steep. .

  That is, in this case, the update of the winner node (the pattern storage model 21 that the winner node has) is performed so as to be strongly influenced by the new time-series data regardless of the progress of learning (update). On the other hand, the update of the nodes other than the winner node (the pattern storage model 21 possessed) is somewhat large at the start of learning over a relatively wide range of nodes (from a node having a small inter-pattern distance d to the winner node) Up to the node), so that it is affected by the new time-series data. Then, as learning progresses, the update of nodes other than the winner node is performed so as to be influenced by new time-series data only for nodes in a narrow range gradually.

  The weight determination unit 43 in FIG. 8 determines the update weight α of each node of the pattern storage SOM as described above, and the learning data update unit 44 stores the learning data stored in the learning data storage unit 22 of each node. Data is updated based on the update weight α of the node.

  Next, an update method for updating the learning data stored in the learning data storage unit 22 included in the node will be described with reference to FIG.

Now, the learning data storage unit 22 included in a node N _i, use learning data has already been stored, the pattern storage model 21 of the node N _i is already learning data stored in the learning data storage unit 22 Suppose you have already learned.

Learning data update unit 44, for example, the node N _i is the learning data that has already been stored in the learning data storage unit 22 having (hereinafter, appropriately referred to as old learning data) and the new time series data from the outside , they were mixed in accordance with the updating weight α of the node N _i from the weight determiner 43, the mixing result as a new learning data, by storing in the learning data storage unit 22, stored contents of the learning data storage unit 22 Is updated to new learning data.

In other words, the learning data update unit 44 adds new time series data to the old learning data to obtain new learning data in which the old learning data and the new time series data are mixed. additional time series data new to the data (mixed with the old training data and the new time series data) is performed according to the ratio corresponding to the updating weight α of the node N _i.

Here, the node updates the N _i of the pattern storage model 21 (FIG. 6) is to be done by a learning using a new learning data, by changing the ratio of mixing the new time series data and the old learning data The degree (intensity) of the influence of new time series data received by the pattern storage model 21 by the update can be changed.

In the node N _i, as a percentage of mixing the new time series data and the old learning data, a node value corresponding to the updating weight α of N _i is employed, for example, as the updating weight α is large, the new The value is such that the ratio of the time series data is large (the ratio of the old learning data is small).

Specifically, in the learning data storage unit 22 of the node N _i is a possible time-series data of a certain number (learning data) is stored, the number of its constant and H. In this case, the learning pattern storage model 21 of the node N _i is always carried out using the H-number of learning data (time-series data).

When a constant number H of learning data is always stored in the learning data storage unit 22, the number of new learning data obtained by mixing new time-series data and old learning data needs to be H. There is such a mixture of the new time series data and the old learning data, as a method of performing in a ratio corresponding to the updating weight α of the node N _i, and a new time series data and the old learning data, There is a method of mixing at a ratio α: H-α.

  As a specific method of mixing new time-series data and old learning data at a ratio α: H-α, as shown in FIG. 10, H-α old out of H old learning data are used. There is a method of obtaining H new learning data by adding α new time-series data to the learning data.

In this case, the number H of the time series data stored in the learning data storage unit 22, for example, a 100, node N _i updating weight α of, for example, if 8, the stored contents of the learning data storage unit 22 Is updated to 100 new learning data obtained by adding 8 new time-series data to 92 old learning data out of 100 old learning data.

  There is a method of adding α new time series data to H-α old learning data of H old learning data after α new time series data is obtained, In this method, the contents stored in the learning data storage unit 22 cannot be updated every time one new time series data is obtained.

  Therefore, the update of the stored contents of the learning data storage unit 22 adds only α new time series data to the H-α old learning data every time one new time series data is obtained. Can be done. That is, one new time-series data is copied to form α new time-series data, and the α new time-series data is excluded from the H old learning data in the order of oldness. By adding to the remaining H-α old learning data, the storage content of the learning data storage unit 22 is updated. Thereby, whenever one new time series data is obtained, the memory content of the learning data memory | storage part 22 can be updated.

  As described above, by updating the storage contents of the learning data storage unit 22, only the H time-series data in the new order are always held as learning data in the learning data storage unit 22. Thus, the ratio (ratio) of new time-series data in the learning data is adjusted by the update weight α.

  Next, learning processing for learning the pattern storage SOM performed by the learning device in FIG. 8 will be described with reference to the flowchart in FIG. 11.

  In step S1, the learning device of the learning unit 4 (FIG. 10) initializes the parameters of the pattern storage SOM, that is, the parameters of the HMM, for example, as the pattern storage model 21 (FIG. 6) possessed by each node of the pattern storage SOM. Perform initialization processing. By this initialization process, appropriate initial values are given as parameters (state transition probability and output probability density function) of the HMM. In the initialization process, how the initial value is given as the HMM parameter is not particularly limited.

  After step S1, the process proceeds to step S2, and new time-series data such as voice data of one voice section (feature parameters extracted from) is input from the outside to the learning apparatus in FIG. Then, the processing proceeds in sequence to steps S3 to S7, and the learning device performs self-organized learning that uses the new time-series data to update the pattern storage SOM in a self-organized manner.

  In other words, in step S3, the score calculation unit 41 obtains a score representing the degree to which each node constituting the pattern storage SOM matches the new time-series data from the outside.

  Specifically, when the pattern storage model 21 (FIG. 6) possessed by the node is, for example, an HMM, a logarithmic likelihood at which new time series data is observed is obtained from the HMM as a score. Here, as a log likelihood calculation method, for example, a Viterbi algorithm can be employed.

  When the score calculation unit 41 calculates the score for the new time-series data for all the nodes included in the pattern storage SOM, the score calculation unit 41 supplies the score for each node to the winner node determination unit 42, and the processing is performed in step S3. To step S4.

  In step S4, the winner node determination unit 42 obtains a node having the highest score from the score calculation unit 41 among the nodes constituting the pattern storage SOM, and determines that node as a winner node. Then, the winner node determination unit 42 supplies a node label as information representing the winner node to the weight determination unit 43, and the process proceeds from step S4 to step S5.

  In step S5, the weight determination unit 43 determines the update weight of each node constituting the pattern storage SOM using the winner node represented by the node label from the winner node determination unit 42 as a reference.

  That is, as described with reference to FIG. 9, the weight determination unit 43 determines that the change in the update weight α with respect to the change in the inter-pattern distance d becomes steeper as the update (learning) of the pattern storage SOM progresses. The update weight α of each node in the pattern storage SOM is determined according to the distance / weight curve represented by () and supplied to the learning data update unit 44, and the process proceeds from step S5 to step S6.

  In step S6, the learning data updating unit 44 uses the learning data stored in the learning data storage unit 22 (FIG. 6) included in each node of the pattern storage SOM in step S6 as a corresponding node from the weight determination unit 43. Update according to the update weight. That is, as described with reference to FIG. 10, the learning data updating unit 44 uses the new time-series data from the outside and the old learning data stored in the learning data storage unit 22 of the node as the update weight of the node. By mixing at a ratio α: H−α corresponding to α, H new learning data are obtained, and the stored contents of the learning data storage unit 22 are updated with the H new learning data.

  When the learning data update unit 44 updates the stored contents of the learning data storage unit 22 (FIG. 6) of all the nodes of the pattern storage SOM, the learning data update unit 44 supplies the model learning unit 45 with an end notification indicating that the update has been completed. Advances from step S6 to step S7.

  In step S <b> 7, the model learning unit 45 updates the parameters of the pattern storage SOM in response to the end notification from the learning data update unit 44.

  That is, the model learning unit 45 learns the pattern storage model 21 for each node of the pattern storage SOM using new learning data stored in the learning data storage unit 22 updated by the learning data update unit 44. As a result, the pattern storage model 21 is updated.

  Specifically, when the pattern storage model 21 included in a node is, for example, an HMM, HMM learning is performed using new learning data stored in the learning data storage unit 22 included in the node. In this learning, for example, the current state transition probability (obtained by learning using old learning data) and the output probability density function of the HMM are set as initial values, and new learning data is used, and Baum-Welch method is used. New state transition probabilities and output probability density functions are obtained. Then, the state transition probability and output probability density function of the HMM are updated by the new state transition probability and output probability density function, respectively.

  Thereafter, the processing returns from step S7 to step S2, waits for the next new time-series data to be input from the outside to the learning device, and thereafter the same processing is repeated.

  Next, FIG. 12 shows a configuration example of a recognition generation apparatus that performs recognition generation processing using a pattern storage SOM.

  In FIG. 12, the recognition generation device includes a score calculation unit 51, a winner node determination unit 52, and a reconstruction unit 53. The score calculation unit 51 and the winner node determination unit 52 perform recognition processing for recognizing data. The reconstruction unit 53 performs a generation process of acquiring (generating) data (reconstruction data) obtained by reconstructing the data recognized by the recognition process.

  That is, for example, time series data as data to be recognized is input from the outside to the recognition generation apparatus, and the time series data is supplied to the score calculation unit 51.

  The score calculation unit 51, like the score calculation unit 41 of the learning device (FIG. 8), for each node constituting the pattern storage SOM, the degree to which the node matches the external time-series data (observed values thereof) Is obtained and supplied to the winner node determination unit 52.

  Similarly to the winner node determination unit 42 of the learning device (FIG. 8), the winner node determination unit 52 obtains a node that best matches the external time-series data in the pattern storage SOM, and determines that node as the winner node. To do.

  That is, the winner node determination unit 52 determines the node having the highest score from the score calculation unit 51 among the nodes constituting the pattern storage SOM as the winner node. Then, the winner node determination unit 52 supplies a node label as information representing the winner node to the reconstruction unit 53 as a recognition result of time-series data from the outside.

  Based on the node label from the winner node determination unit 52, the reconfiguration unit 53 reconfigures and outputs external time-series data.

  That is, the reconfiguration unit 53 uses the node (winner node) represented by the node label from the winner node determination unit 52 as a generation node used to generate reconfiguration data obtained by reconstructing time-series data from the outside. Based on the pattern storage model 21 (FIG. 6) of the generation node, the reconstruction data is acquired (generated) and output.

  Specifically, when the pattern storage model 21 is, for example, an HMM, the reconfiguration unit 53 indicates, for example, the likelihood that the time series data is observed in the HMM as the pattern storage model 21 included in the generation node. Generate time-series data that maximizes the output probability to be represented.

  For example, K. Tokuda, T. Yoshimura, T. Masuko, T. Kobayashi, T. Kitamura, "SPEECH PARAMETER GENERATION ALGORITHMS FOR HMM-BASED SPEECH SYNTHESIS" , Proc. Of ICASSP 2000, vol.3, pp.1315-1318, June 2000, Tetsuya Inagi, Hiroaki Tanie, Hitoshi Nakamura, “Keyframe Extraction of Time Series Data Using Continuously Distributed Hidden Markov Model and Its Restoration ", JSME Robotics and Mechatronics Lecture 2003 Proceedings, 2P1-3F-C6, 2003, etc.

  Here, for example, the time series data is voice data (feature parameters extracted from), and the reconstruction unit 53 observes the time series data using the HMM as the pattern storage model 21 of the generation node. When generating voice data as time-series data that maximizes the output probability representing the likelihood, distortion in which the waveform becomes discontinuous (hereinafter, referred to as generation round as appropriate) may occur in the voice data.

  Therefore, the reconfiguration unit 53 can acquire time series data as learning data stored in the learning data storage unit 22 of the generation node by randomly selecting the time series data and outputting the time series data as reconfiguration data. it can.

  The time series data selected at random from the time series data as the learning data stored in the learning data storage unit 22 included in the generation node follows the probability distribution of the learning data stored in the learning data storage unit 22 included in the generation node. Therefore, from the viewpoint of the expected value, it can be regarded as time-series data generated using the HMM as the pattern storage model 21 possessed by the generation node, and does not have the above-described generation round.

  Next, recognition processing (hereinafter referred to as recognition generation processing) for recognizing time-series data by the recognition generation apparatus in FIG. 12 will be described with reference to the flowchart in FIG.

  In step S21, for example, when time series data such as voice data of one voice section (feature parameters extracted from) is input from the outside to the recognition generation device in FIG. 12, the processing is performed in steps S22 and S23. Then, a recognition process for recognizing time-series data from the outside is performed.

  That is, in step S22, the score calculation unit 51 obtains a score representing the degree to which each node constituting the pattern storage SOM matches the external time-series data, and supplies the score to the winner node determination unit 52. The process proceeds to step S23.

  In step S23, the winner node determination unit 52 obtains the node having the highest score from the score calculation unit 51 among the nodes constituting the pattern storage SOM, and determines that node as the winner node. Then, the winner node determination unit 52 supplies a node label as information representing the winner node to the reconstruction unit 53 as a recognition result of time-series data from the outside.

  After step S23, the process proceeds to step S24, and the reconstruction unit 53 reconstructs time-series data from the outside based on the recognition result obtained in step S23, and the reconstruction obtained as a result. Output data.

  That is, in step S24, the reconfiguration unit 53 uses the generation node as a generation node for the node represented by the node label as the recognition result from the winner node determination unit 52, and the generation node acquires the pattern. Is acquired as described above, and is output as reconstructed data obtained by reconstructing the time-series data from the outside. This completes the recognition generation process.

  In addition, according to the reconstructed data acquired in the above recognition generation process, for example, the recognition generation apparatus itself can confirm how time-series data from the outside is recognized. it can.

  Next, FIG. 14 schematically shows the storage contents of the main body model storage unit 122 and the divided error model storage unit 124 of FIG. 4 when the pattern storage SOM is adopted as the main body model and the division error model. ing.

  In FIG. 14, the main body model storage unit 122 stores a pattern storage SOM as a main body model, and the division error model storage unit 124 stores a pattern storage SOM as a division error model.

  Here, hereinafter, the pattern storage SOM as the main body model is also referred to as a main body SOM, and the pattern storage SOM as the division error model is also referred to as a division error SOM.

  FIG. 15 is a diagram for explaining data that can be input and output with respect to the main body SOM and the division error SOM in FIG. 14.

Data D to be learned and data D to be recognized are inputted (supplied) to the main body SOM. In addition, from the main body SOM, to the data D to be input thereto, it is possible to output the reconstructed data D _R reconstructed the data D.

For division error SOM, and the data D to be input to the body SOM, error data △ D is predetermined dimension or a error between reconstructed data D _R to be outputted from the main SOM for the data D Division error data ΔD _i (i = 1,..., N) divided every time is input. Further, from the division error SOM, the division reconstruction error data ΔD _Ri (i = 1,..., Reconstructed from the division error data ΔDi for the division error data ΔD _i input thereto. n) can be output.

  Next, as shown in FIGS. 14 and 15, the main body learning module 101, the error learning module 102, and the control unit in FIG. 4 when the pattern storage SOM is adopted as the main body model and the division error model. The learning process by 103 will be described with reference to the flowchart of FIG.

  In this learning process, data D to be subjected to the learning process is supplied from the outside to the calculation unit 121 and the error calculation unit 125.

  In the learning process, in step S131, the calculation unit 121 learns the data D supplied thereto, that is, uses the data D to update the main body SOM stored in the main body model storage unit 122 in a self-organizing manner. The self-organized learning of the main body SOM is performed as described in steps S3 to S7 in FIG.

  In step S <b> 132, the arithmetic unit 121 determines a winner node for the data D from the nodes that constitute the main body SOM stored in the main body model storage unit 122. Here, the node having the highest score for the data D among the nodes constituting the main body SOM is determined as the winner node for the data D.

  Since the winner node for data D is determined by self-organized learning in step S131, the winner node can be employed as it is in step S132. However, in step S132, separately from step S131, for each node constituting the main body SOM, a score for data D can be obtained, and the node with the highest score can be determined as the winner node for data D.

In step S133, the arithmetic unit 121 uses the winning node determined in the step S132, as the generation process of generating the reconstruction data D _R by reconstructing the data D, described in step S24 in FIG. 13 To do. Then, the arithmetic unit 121, the reconstructed data D _R generated, and supplies the error calculating unit 125 and the addition unit 128.

In step S134, the error calculation unit 125 calculates the data D from the outside, the error data △ D is the difference between the reconstructed data D _R from the arithmetic unit 121, to the data dividing unit 126. The data dividing unit 126 divides the error data ΔD, which is an n-dimensional vector or time series data, for each predetermined dimension or every predetermined time, and obtains the divided error data ΔD _i (i = 1,..., N) are supplied to the calculation unit 123 and the correction data generation unit 127.

In step S135, the arithmetic unit 123 learns the division error data ΔD _i supplied thereto, that is, uses the division error data ΔD _i, and calculates the division error SOM stored in the division error model storage unit 124 by itself. Self-organized learning of the division error SOM that is systematically updated is performed as described in steps S3 to S7 in FIG. This completes the learning process.

As described above, in the learning process of FIG. 16, the main body SOM learns the data D to be originally learned, and the division error SOM obtains the data D and the main body SOM reconstructing the data D. The division error data ΔD _i obtained by dividing the error data ΔD, which is an error with the reconstructed data D _R to be generated, is learned.

  Note that the learning process in FIG. 16 is performed, for example, every time data D to be subjected to the learning process is input.

  Next, as shown in FIGS. 14 and 15, the main body learning module 101, the error learning module 102, and the control unit 103 in FIG. 4 when the pattern storage SOM is adopted as the main body model and the division error model. The recognition generation process by means of will be described with reference to the flowchart of FIG.

  In this recognition generation processing, data D to be recognized is supplied from the outside to the calculation unit 121 and the error calculation unit 125.

  In step S141, the calculation unit 121 performs recognition processing for recognizing the data D supplied thereto based on the learning result of the learning processing in FIG.

  That is, the operation unit 121 determines a winner node for the data D from the nodes constituting the main body SOM stored in the main body model storage unit 122, and recognizes the winner node (a node label representing the data D). As a result.

In step S142, the arithmetic unit 121, based on the recognition result of the data D in step S141 immediately before, to generate the reconstructed data D _R by reconstructing the data D. That is, the arithmetic unit 121 as the recognition result of the data D, with the winning node for the data D, to generate the reconstructed data D _R reconstructed data D, supplied to the error calculating unit 125 and adding unit 128 To do.

Here, reconstructed data D _R may be with reference to FIG. 12 obtained as described above. In other words, reconstruction data D _R can be generated using a pattern storage model 21 of the winning node (FIG. 6) has for the data D. Also, reconstruction data D _R can be obtained by selecting one of the learning data stored in the learning data storage unit 22 included in the winning node (FIG. 6) for the data D.

In step S143, the error calculation unit 125 calculates the data D supplied from the outside, the error data △ D to the reconstructed data D _R by the calculation unit 121 is reconstructed for the data D data dividing unit 126. The data dividing unit 126 divides the error data ΔD, which is an n-dimensional vector or time series data, for each predetermined dimension or every predetermined time, and obtains the divided error data ΔD _i (i = 1,..., N) are supplied to the calculation unit 123 and the correction data generation unit 127.

In step S144, the calculation unit 123 performs a recognition process for recognizing the supplied division error data ΔD _i based on the learning result of the learning process of FIG. That is, the calculation unit 123 determines a winner node for the division error data ΔD _i from the nodes constituting the error SOM stored in the division error model storage unit 124, and determines the winner node (a node label representing the winner node). And the recognition result of the division error data ΔD _i .

In step S145, the arithmetic unit 123, based on the recognition result of the division error data △ D _i in step S144 immediately before, the division error data △ D _i divided reconstructed reconstruction error data △ D _Ri (i = 1 ,..., N) are generated and supplied to the correction data generation unit 127. That is, the arithmetic unit 123, as a recognition result of the division error data △ D _i, with the winning node for the division error data △ D _i, division error data △ D _i reconstituted divided reconstruction error data △ D _Ri is generated and supplied to the correction data generation unit 127.

Here, reconstruction error data △ D _R, like the reconstructed data D _R, generated using a pattern storage model 21 of the winning node (FIG. 6) to the error data △ D, or error data △ D Can be obtained by selecting one of the learning data stored in the learning data storage unit 22 of the winner node.

In step S146, the correction data generation unit 127 determines the dimension or time to be corrected of the reconstructed data D _R based on the division error data ΔD _i from the data division unit 126 (for example, a predetermined number from the larger error). In addition, a correction weight for each dimension or time is determined based on the determination result, and the division reconstruction error data ΔD _Ri from the calculation unit 123 is multiplied by the correction weight. As a result, partial correction data ΔD _j is generated and supplied to the adder 128.

In step S147, the adding unit 128 adds the reconstructed data D _R from the calculation unit 121 and the partial correction data ΔD _j from the correction data generating unit 127, and the addition data D _R + Δ obtained as a result. the D _Rj, and outputs the reconstructed data D _R as corrected final reconstruction data. This completes the recognition generation process.

As described above, in the learning process, the data D, is calculated as well as learning body SOM, the error data △ D to the reconstructed data D _R reconstructed data D by using the data D and the body SOM, further divided division error data △ D _i learned by division error SOM, the recognition generating processing, the reconstructed data D _R reconstructed data D by using the body SOM, was reconstituted with division error SOM division Correction is performed by adding the partial correction data _{ΔD Rj} generated based on the reconstruction error data ΔD _Ri , and the resulting addition data D _R + _{ΔD Rj} is converted into the final reconstruction of the data D. Since it is output as configuration data (final reconstructed data), a process that appropriately considers an error every predetermined dimension or every predetermined time is performed, and as a result, a quantization error can be reduced.

  That is, in the recognition generation process, the recognition process for determining the winner node for the data D from the nodes constituting the main body SOM is the pattern of the data D among the finite number of nodes constituting the main body SOM. Is the process of converting to the node that has acquired the most similar pattern, and is equivalent to quantizing the data D.

Furthermore, the recognition generating processing, among the nodes of the body SOM, with the winning node for the data D, generation processing for generating the reconstruction data D _R, which reconstructs the data D is quantized result data D Is converted to data of a pattern acquired by the winner node, which is equivalent to inverse quantization of the winner node.

  Therefore, according to the recognition generation process, a quantization error occurs as in the case of performing quantization and inverse quantization.

Therefore, in the learning process, error data ΔD, which is a quantization error caused by quantization and dequantization, is divided for each predetermined dimension or time, and learning is performed with the division error SOM. The sum data D obtained by adding the reconstructed data D _R reconstructed using the data D and the partial correction data _{ΔD Rj} generated based on the divided reconstructed error data ΔD _Ri reconstructed using the segmented error SOM. _{By using R} + _{ΔD Rj} as final reconstructed data obtained by reconstructing data D, the quantization error can be efficiently reduced.

  In other words, when the number of nodes constituting the SOM is N1, when the number of nodes in the SOM is increased by N2 to (N1 + N2), N1 types of quantization errors are included. Only configuration data could be output, but reconstruction data including (N1 + N2) quantization errors increased by N2 (reconstruction data considering N1 + N2 quantization errors) It becomes possible to output. Therefore, when the number of nodes in the SOM is increased by N2, the number of reconstructed data that can be output only increases by N2, which is the increase in the number of nodes.

In contrast, with comprises the body SOM at N1 nodes, the division error SOM constituted by N2 nodes, and reconstructed data D _R reconstructed data D by using the body SOM, division error SOM the final reconstruction of the added data D _R + △ D _Rj obtained by adding the partial correction data △ D _Rj generated based on the divided reconstruction error data △ D _Ri reconstituted, reconstructed data D using In the case of data, N1 × N2 final reconstruction data can be output.

  Specifically, if the number of nodes N1 and N2 are both 10 for example, when the number of nodes constituting the SOM is increased, the number of reconstructed data that can be output is 20 (= 10 + In contrast, when the main body SOM and the division error SOM are used, the number of reconstructed data that can be output is 100 (= 10 × 10). Therefore, when the main body SOM and the division error SOM are used, the quantization error can be further reduced as compared with the case where the number of nodes constituting the SOM is increased.

  That is, when the number of nodes constituting the SOM is increased, each node holds a pattern of reconstructed data in consideration of the quantization error, so that one quantization error is generated for one node. Only reconstructed data can be obtained in consideration of

In contrast, in the case of using a body SOM and division error SOM is the division error SOM, split reconstruction can be commonly used to correct the reconstructed data D _R obtained from each node of the main body SOM Since the pattern of error data ΔD _Ri (division error data ΔD _Ri ) is stored, one node of the main body SOM is several times the number of nodes constituting the division error SOM (in the correction data generation division 127). It is possible to obtain the addition data D _R + _{ΔD Rj} as final reconstructed data in consideration of the quantization error of the number of types of correction weights to be determined), and as a result, to further reduce the quantization error Can do.

  Here, in order to output the same number of reconstructed data as in the case of using the main body SOM and the division error SOM by increasing the number of nodes constituting the SOM, in the above example, the nodes constituting the SOM Needs to be 100. And since the storage capacity such as the memory required to store the SOM is proportional to the number of nodes that make up the SOM, the number of nodes that make up the SOM is increased when using the main SOM and the division error SOM. Compared with this case, the amount of memory such as a memory for storing the SOM (main body SOM and error SOM) can be reduced in order to reduce the quantization error to the same extent. That is, it is possible to reduce the quantization error while efficiently using the memory.

  Note that both of the learning process of FIG. 16 and the recognition generation process of FIG. 17 can be performed on the input of the data D, or only one of them can be performed.

  Next, FIG. 18 illustrates a second configuration example of the main body learning module 101, the error learning module 102, and the control unit 103 in FIG.

  In FIG. 18, the main body learning module 101 is from the calculation unit 201 and the main body model storage unit 202, the error learning module 102 is from the calculation unit 203 and the divided error model storage unit 204, and the control unit 103 is the error calculation unit 205, data Each of the division unit 206, the correction data generation unit 207, and the addition unit 208 is configured.

  For example, input data S and output data A are supplied from the outside as data D in FIG. 1 to the arithmetic unit 201 of the main body learning module 101.

  Here, as the input data S and the output data A, for example, as described above, sensor data and action data can be employed. The input data S and output data A are assumed to be n-dimensional vectors or time series data.

  That is, for example, voice data or image data (feature parameters extracted from) is adopted as the input data S, and the robot arm is used as the output data A according to the voice data and the image data as the input data S. Parameters used for generating synthesized speech (speech synthesis) in accordance with control data for controlling the actuator to be driven, audio data as the input data S, and image data can be employed.

  Further, for example, sensor data of frame #t can be adopted as input data S, and sensor data of frame # t + 1 can be adopted as output data A. Further, for example, as input data S, a combination of sensor data of frame #t and action data of frame #t is adopted, and as output data A, sensor data of frame # t + 1 and frame #t A combination with +1 action data can be adopted.

  At the time of learning, teaching data in which the input data S and the output data A to be output with respect to the input data S are set is supplied to the calculation unit 201 (and the error calculation unit 205). Only the input data S is supplied to the calculation unit 201.

  Here, for example, when the input / output system of FIG. 1 is applied to a robot, when the robot learns the task of rolling the ball in front of the user to the left or right, the user places the ball in front of the robot. Hold the robot arm and move the arm to roll the ball left and right.

  In this case, assuming that the image data obtained by photographing (detecting) the state of the ball rolling to the left and right is adopted as the input data S, as the output data A to be output with respect to the input data S, the arm of the robot, Control data for controlling the actuator for driving as if the user is moving can be employed.

  After giving the teaching data which is a set of the input data S and the output data A as described above to the input / output system and performing the learning process, in the recognition generation process, the image data obtained by photographing the state of the ball rolling left and right, When the input data S is given to the input / output system, the input / output system outputs control data for controlling the actuator so as to drive the arm of the robot in accordance with the state of the ball.

  The arithmetic unit 201 learns the input data S and output data A supplied thereto. That is, the arithmetic unit 201 learns the main body model stored in the main body model storage unit 202 using the input data S and output data A supplied thereto.

Further, the arithmetic unit 201 reconstructs the input data S using the main body model stored in the main body model storage unit 202 for the input data S supplied thereto, and the reconstructed input obtained as a result Data S _R is output. Further, the arithmetic unit 201 uses the main body model stored in the main body model storage unit 202 to reconstruct the output data A to be output with respect to the input data S supplied thereto, and obtains the reproduction data obtained as a result. and it outputs the configuration output data a _R. Further, the calculation unit 201 reconstructs the output data A using the main body model stored in the main body model storage unit 202 for the output data A supplied thereto, and the reconstructed output obtained as a result Data A _R is output.

The reconstructed input data S _R output from the calculation unit 201 is supplied to the error calculation unit 205 of the control unit 103, and the reconstructed output data A _R output from the calculation unit 201 is added to the error calculation unit 205 of the control unit 103 and added. Supplied to the unit 208.

  The main body model storage unit 202 stores a main body model as a model for learning data. As the main body model, for example, SOM, NN, or other learning models capable of unsupervised learning can be adopted.

The calculation unit 203 of the error learning module 102 receives the divided input error data ΔS _i (i = 1, 2,..., N) and the divided output error data ΔA _i (i) from the data dividing unit 206 of the control unit 103. = 1, 2,..., N) is supplied as error data ΔD in FIG.

Calculation unit 203 performs learning of the divided input error data △ S _i and the divided output error data △ A _i from the data dividing unit 206. That is, the calculation unit 203 uses the divided input error data △ S _i and the divided output error data △ A _i from the data dividing unit 206 performs learning of the error model stored in the division error model storage unit 204.

Further, the calculation unit 203 reconstructs the divided output error data ΔA _i corresponding to the divided input error data ΔS _i supplied thereto using the divided error model stored in the divided error model storage unit 204. Then, the divided reconstruction output error data ΔA _{R i} (i = 1, 2,..., N) obtained as a result is output.

Here, the divided output error data △ A _i corresponding to the divided input error data △ S _i is divided outputs divided input error data △ S _i is obtained from the output data A to be output to the input data S obtained This means error data ΔA _i .

Further, the arithmetic unit 203 re-divides the divided output error data ΔA _i using the divided error model stored in the divided error model storage unit 204 for the divided output error data ΔA _i supplied thereto. The divided reconstruction output error data ΔA _Ri obtained as a result is output.

The divided reconstruction output error data ΔA _Ri output from the calculation unit 203 is supplied to the correction data generation unit 207 of the control unit 103.

  The division error model storage unit 204 stores an error model as a model for learning data. As the division error model, for example, SOM, NN, and other learning models capable of unsupervised learning can be adopted.

As described above, the error calculation unit 205 of the control unit 103 is supplied with the reconstruction input data S _R and the reconstruction output data A _R from the computation unit 201, and the input data S and The same input data S and output data A as the output data A are supplied from the outside.

Error calculation unit 205 obtains the input data S from the outside, the difference SS _R between the reconstructed input data S _R from the arithmetic unit 201 as input error data △ S, to the data dividing unit 206.

Further, the error calculation unit 205 obtains a difference AAR between the output data A from the outside and the reconstructed output data A _R from the calculation unit 201 and supplies the difference AA _R to the data division unit 206 as output error data ΔA.

The data dividing unit 206 divides the input error data ΔS and the output error data ΔA, which are n-dimensional vectors or time series data, for each predetermined dimension or every predetermined time, and obtains the result. The divided input error data ΔS _i and the divided output error data ΔA _i are output.

For example, when the input error data ΔS and the output error data ΔA are two-dimensional vectors, the input error data ΔS is divided into ΔS ₁ and ΔS ₂ and the output error data ΔA is divided into ΔA ₁ and Δ Divide into A ₂ . Further, for example, input error data △ S and the output error data △ A is 100 when a time series data of step, input error data △ S and output error data △ A, 1 step from up to 50 steps △ S ₁ and △ a ₁ and divides the 51 step △ S ₂ and △ a ₂ to 100 steps. The number n to be divided is arbitrary irrespective of the number of dimensions and the number of steps of the input error data ΔS and the output error data ΔA.

The divided input error data ΔS _i is supplied to the calculation unit 203 and the correction data generation unit 207, and the divided output error data ΔA _i is supplied to the calculation unit 203.

The correction data generation unit 207 determines the dimension or time to be corrected of the reconstructed output data S _R based on the divided input error data ΔS _i from the data dividing unit 206, or randomly determines, according to the determination result, Partial correction data ΔA _Rj is generated by determining a correction weight for each predetermined dimension or time and dividing the division reconstruction output error data ΔA _Ri from the calculation unit 203 by the correction weight. Then, it is supplied to the adding unit 208.

The adding unit 208 adds the reconstructed output data A _R supplied from the calculation unit 201 and the partial correction data ΔA _Rj supplied from the correction data generating unit 207, and the resulting addition data A _R ′ = A _R + ΔA _Rj is output as final reconstructed output data (hereinafter also referred to as final reconstructed output data as appropriate) obtained by correcting the reconstructed output data A _R.

  Here, in FIG. 18, as the main body model stored in the main body model storage unit 202 and the division error model stored in the division error model storage unit 204, the same learning model is used as in FIG. However, in order to simplify the explanation, the same learning model is adopted as the main body model and the error model.

In FIG. 18, the main body model is a learning model that performs learning on the input data S and the output data S that are data to be originally learned, whereas the division error model has learned data. Using the body model, the error (input error data ΔS) between certain input data S and reconstructed input data S _R obtained by reconfiguring the input data S is divided for each predetermined dimension or every predetermined time. the divided input error data △ S _i, and is the output data a and the error between the reconstructed output data a _R obtained by reconstructing the output data a (output error data △ a) a predetermined dimension or every predetermined This is a learning model in which learning is performed on the divided output error data ΔA _i divided every time. Accordingly, in FIG. 18, as in FIG. 4, the main body model and the division error model differ only in the learning target, and there is no essential difference as a learning model.

As the main body model for learning the input data S and the output data A, and the divided error model for learning the divided input error data ΔS _i and the divided output error data ΔA _i , for example, a SOM including a pattern storage SOM is used. An input / output relationship model can be adopted.

  Here, the input / output relationship model will be described with reference to FIGS.

  FIG. 19 schematically shows a configuration example of an input / output relationship model using a pattern storage SOM.

In FIG. 19, the input / output relationship model has two pattern storages SOM # 1 and # 2. Further, the input / output relationship model includes each node N _i (i = 1, 2,..., Total number of nodes) of the pattern storage SOM # 1 and each node N ′ _j (j = 1) of the pattern storage SOM # 2. , 2,..., The total number of nodes).

Here, in FIG. 19, the node N _i of pattern storage SOM # 1, 'the arrow between _j, the node N _i and N' node N of pattern memory SOM # 2 represents the binding of _j .

Note that the pattern storage SOMs # 1 and # 2 may have the same number of nodes and the same links, or may have different numbers of nodes or different links. Further, the pattern storage model 21 of the node N _i of pattern storage SOM # 1 (FIG. 6), the pattern storage model 21 of the node of pattern storage SOM # 2 also may be the same pattern storage model, Different pattern storage models may be used.

  Next, FIG. 20 illustrates a configuration example of a learning device that performs learning processing and a recognition generation device that performs recognition generation processing using an input / output relationship model.

  20, the learning device includes a learning unit 221 and a connection weight update unit 222, and the recognition generation device includes a score calculation unit 231, a winner node determination unit 232, a generation node determination unit 233, and a reconfiguration unit 234. Is done.

  Teaching data, which is a set of input data (observation values) to be learned and output data (observation values) to be output with respect to the input data, is supplied to the learning device. The learning device uses the teaching data supplied thereto to update the pattern storage SOMs # 1 and # 2 constituting the input / output relationship model (FIG. 19), and the pattern storages SOM # 1 and # 2 Update the connection relationship between the two nodes.

  That is, in the learning device of FIG. 20, the learning unit 221 is configured in the same manner as the learning device shown in FIG. 8, and uses the input data of the teaching data supplied to the learning device to input / output relation model (FIG. The pattern storage SOM # 1 in 19) is updated in a self-organizing manner as in the learning apparatus of FIG. In addition, the learning unit 221 uses output data (output data set with input data in the teaching data) of the teaching data supplied to the learning unit 221 in the input / output relation model (FIG. 19). The pattern storage SOM # 2 is updated in a self-organizing manner as in the learning apparatus of FIG.

Furthermore, the learning unit 221, has become the winning node to update the pattern storage SOM # 1, the pattern storage SOM # 1 of the node N _i of the node label (hereinafter, input label referred to) and the pattern memory was the winning node to update the SOM # 2, the pattern storage SOM # 2 of the node N _'j of node label (hereinafter, output label referred to) a label set to the set, the weight updating To the unit 222.

The weight updating unit 222, based on the label set supplied from the learning unit 221, and a node N _i of pattern storage SOM # 1 in the input-output relationship model (Fig. 19), and a node N _'j pattern storage SOM # 2 Update the join relationship.

Here, the label set supplied from the learning unit 221 to the connection weight updating unit 222 is a set of an input label and an output label, and the input label is a pattern storage SOM # 1 using the input data of the teaching data. was the winning node to update the, because the node label of the pattern memory SOM # 1 of the node N _i, in the pattern storage SOM # 1, the node label of the most compatible node N _i to the input data.

Similarly, output label has become the winning node to update the pattern storage SOM # 2 using the output data of the teaching data, since the node label of the pattern storage SOM # 2 node N _'j In the pattern storage SOM # 2, this is the node label of the node N ′ _j that best matches the output data.

In the weight updating unit 222, among the nodes of pattern storage SOM # 1, and the winning node N _i which is the most compatible node to the input data in the teaching data, the binding relationship between each node of the pattern memory SOM # 2 is updated In addition, the connection relationship between the winner node N ′ _j that is the most suitable node for the output data in the teaching data among the nodes of the pattern storage SOM # 2 and the nodes of the pattern storage SOM # 1 is updated.

  Here, the connection relationship between the node of the pattern storage SOM # 1 and the node of the pattern storage SOM # 2 in the input / output relationship model is represented by a connection weight that increases as the degree of the connection increases. The update of the connection relationship between the nodes means the update of the connection weight. Details of the method of updating the connection relationship between nodes by the connection weight update unit 222 will be described later.

Input data is supplied to the recognition generator. Recognition generating apparatus, the pattern storage SOM # 1 in the input-output relationship model, to determine the most compatible winning node N _i to the input data, the strongest link weight between the winning node N _i, the pattern storage SOM # 2 The node is determined as a generation node N ′ _j that generates reconstructed output data (estimated value of output data) obtained by reconstructing output data to be output with respect to the input data. Further, the recognition generation apparatus generates and outputs reconstructed output data obtained by reconstructing output data to be output with respect to input data, using the pattern storage model 21 (FIG. 6) of the generation node N ′ _j . .

That is, in the recognition generating apparatus of FIG. 20, the score calculation unit 231, the input data supplied to the recognition generator, the degree nodes N _i of pattern storage SOM # 1 constituting the input-output relationship model fits Is calculated in the same manner as in the score calculation unit 51 of the recognition generation apparatus of FIG. 12 and is supplied to the winner node determination unit 232.

Winning node determining unit 232, as in the case of recognition generating device winning node determining unit 52 in FIG. 12, the node N _i of highest score pattern storage SOM # 1 supplied from the score computing unit 231 to the winning node The node label (input label) representing the winner node is supplied to the generation node determination unit 233.

Generation node determination unit 233, among the patterns stored SOM # 2 that constitute the input-output relationship model nodes, the strongest link weight between the nodes N _i represented by the input label from the winning node determining unit 232 (the strongest) The node N ′ _j is determined as a generation node, and an output label representing the generation node N ′ _j is supplied to the reconfiguration unit 234.

The reconfiguration unit 234 stores the pattern storage model 21 (FIG. 6) included in the node N ′ _j indicated by the output label from the generation node determination unit 233 among the nodes of the pattern storage SOM # 2 constituting the input / output relationship model. Based on the input data supplied to the recognition generation device, the reconstruction output data obtained by reconstructing the output data to be output is acquired in the same manner as in the reconstruction unit 53 of the recognition generation device in FIG. (Generate) and output.

  Next, the learning process of the input / output relation model performed by the learning device of FIG. 20 will be described with reference to the flowchart of FIG.

  In step S201, when teaching data that is a set of input data and output data is input to the learning apparatus in FIG. 20, the teaching data is supplied to the learning unit 221 and the process proceeds to step S202.

  In step S202, the learning unit 221 uses the input data of the teaching data to update the pattern storage SOM # 1 in the input / output relation model (FIG. 19) in a self-organizing manner. In step S203, the learning unit 221 uses the output data of the teaching data to update the pattern storage SOM # 2 in the input / output relation model in a self-organizing manner.

Then, the learning unit 221, has become the winning node to update the pattern storage SOM # 1, the input label of the pattern memory SOM # 1 of the node N _i, winning node to update a pattern storage SOM # 2 The label set that is a set with the output label of the node N ′ _j of the pattern storage SOM # 2 is supplied to the connection weight update unit 222.

  In step S204, the connection weight updating unit 222, based on the label set supplied from the learning unit 221, each node of the pattern storage SOM # 1 and each node of the pattern storage SOM # 2 in the input / output relationship model (FIG. 19). And the process returns to step S201, waits for the next teaching data to be input, and thereafter the same processing is repeated.

  Note that the processes of steps S202 and S203 can be performed in parallel or in the reverse order of the case of FIG.

  Next, update of the connection weight (connection relationship between nodes) performed by the connection weight update unit 222 (FIG. 20) in step S204 of FIG. 21 will be described with reference to FIG.

  In addition, hereinafter, in the input / output relation model, a pattern storage SOM to which input data is given, that is, a pattern storage SOM that performs learning using the input data and acquires a pattern of the input data (pattern storage SOM # 1) Is also called an input SOM, a pattern storage SOM to which output data is given, that is, a pattern storage SOM (pattern storage SOM # 2) that is trained using the output data and acquires the pattern of the output data is also called an output SOM. Say.

  Further, hereinafter, the node of the input SOM is also referred to as an input node, and the node of the output SOM is also referred to as an output node.

  FIG. 22 illustrates an input / output relationship model in which the input node of the input SOM # 1 and the output node of the output SOM # 2 are coupled as illustrated in FIG.

  That is, on the left side of FIG. 22, the input / output relationship model has an input SOM # 1 and an output SOM # 2, and each input node of the input SOM # 1 and each output node of the output SOM # 2 are coupled. ing. In FIG. 22, both the input SOM # 1 and the output SOM # 2 are composed of eight nodes.

  On the left side of FIG. 22, each input node of input SOM # 1 is coupled to all output nodes of output SOM # 2 (so each output node of output SOM # 2 is also connected to all inputs of input SOM # 1). There is a connection weight w for all combinations of the input node of the input SOM # 1 and the output node of the output SOM # 2.

  Now, for the two pattern storage SOMs in which the nodes are connected, one of the pattern storage SOMs is made to correspond to each row, and the other pattern storage SOM is made to correspond to each column. A matrix in which the connection weights w between the i-th node of one pattern storage SOM and the j-th node of the other pattern storage SOM are arranged in j column elements is referred to as a connection weight matrix MTX. The weight update unit 222 (FIG. 20) updates the connection weight w that is each element of the connection weight matrix MTX.

  The right side of FIG. 22 shows the connection weight matrix MTX for the input / output relationship model on the left side of FIG.

  In the connection weight matrix MTX on the right side of FIG. 22, the input node corresponds to each row, the output node corresponds to each column, the i-th input node, and the j-th output node as elements of the i-th row and j-th column. And the connection weight w is arranged. The connection weight updating unit 222 (FIG. 20) updates the connection weight w that is each element of the connection weight matrix MTX.

  That is, the connection weight updating unit 222 initializes all connection weights w in the connection weight matrix MTX to initial values (for example, 0), for example, when the power is first turned on. Then, the connection weight updating unit 222 receives teaching data, that is, a set of input data and output data, from the learning device in FIG. 20, and thereby an input label representing a winner node in the input SOM # 1 Each time a label set with an output label representing a winner node in the output SOM # 2 is given from the learning unit 221, the connection weight is updated with the connection between the winner nodes as the center.

  Specifically, the connection weight updating unit 222 updates the connection weight w of the connection weight matrix MTX, for example, according to the equation (2).

w = (1-β) w + β △ w
... (2)

  Here, β is a learning rate representing the degree of updating the connection weight w, and is given in the range of 0 to 1. As the learning rate β decreases, the connection weight w does not change much. When the learning rate β is 0, the connection weight w does not change. On the other hand, as the learning rate β is increased, the connection weight w is also greatly changed. When the learning rate β is 1, the connection weight w is updated to the update reference value Δw.

  Further, the reference value Δw is given by, for example, Expression (3).

△ w = 1 / (1 + d)
... (3)

  Here, d represents the inter-pattern distance with the winner node, similarly to the case of Expression (1), and d = 0 for the node that is the winner node. Therefore, the reference value Δw is 1 for the winner node (the node that becomes), and the reference value Δw approaches 0 as the inter-pattern distance d from the winner node increases.

Now, the input node input label supplied from the learning unit 221 (FIG. 20) to the weight updating section 222 is represented, i.e., a winning node of the input SOM # 1, together with representative of the input node N _i, the learning unit 221 (FIG. 20), when the output node represented by the output label supplied to the connection weight update unit 222, that is, the winner node in the output SOM # 2, is expressed as the output node N ′ _j , the connection weight update unit 222 (FIG. 20) In accordance with Expression (2) (and Expression (3)), the connection weight w of the connection weight matrix MTX is updated as follows.

That is, the weight updating unit 222, for each output node of the output SOM # 2, using the inter-pattern distance d between the output node N _'j is a winning node in the output SOM # 2, according to equation (3), the reference It obtains the value △ w, further the use of a reference value △ w, according to equation (2), and updates the weight w of the i-th input node N _i is the winning node of the input SOM # 1.

Thus, corresponding to the input node N _i is the winning node of the input SOM # 1, the i-th row of bond weight matrix MTX (each column) weight w is updated.

Further, the weight updating unit 222, for each input node of the input SOM # 1, using the inter-pattern distance d between the input node N _i is a winning node of the input SOM # 1, in accordance with the formula (3), the reference value Δw is obtained, and further, using the reference value Δw, the connection weight w with the _jth output node N ′ _j that is the winner node of the output SOM # 2 is updated according to the equation (2).

As a result, the connection weight w (for each row) in the j-th column of the connection weight matrix MTX corresponding to the output node N ′ _j that is the winner node of the output SOM # 2 is updated.

Therefore, the winning node N _i inputs SOM # 1, link weight between the winning node N _'j of the output SOM # 2 is updated to enhance the most the degree of binding.

Note that the input node N _i is the winning node of the input SOM # 1, update of the weight w and the output is the winning node of SOM # 2 output node N _'j is the output SOM # 2 for each output node , when updating of weight w and the input node N _i is a winning node, or, for each input node of the input SOM # 1, the weight w of the output node N _'j is a winning node update time of This is done only in either one.

  The update of the connection weight w (connection weight matrix MTX) as described above is performed every time a set of input data and output data as teaching data is input to the learning apparatus in FIG.

  Further, in the learning using the input data of the input SOM # 1 and the learning using the output data of the output SOM # 2 included in the input / output relation model, a set of input data and output data as teaching data is shown in FIG. It is performed every time it is input to the learning device.

  As learning of the input SOM # 1 and the output SOM # 2 progresses, the pattern storage model 21 possessed by the input node expresses (acquires) a specific pattern, and the pattern storage model 21 possessed by the output node also , To express other specific patterns.

  As a result, when there is some relationship between input data of a specific pattern and output data of another specific pattern, a set (teaching data) of such input data and output data is given. The input node having the pattern storage model 21 expressing a specific pattern in the input SOM # 1 becomes the winner node, and the output node having the pattern storage model 21 expressing another specific pattern in the output SOM # 2 Becomes the winner node.

  Furthermore, as described above, with the winner node of input SOM # 1 as the center, the connection weight of each input node of input SOM # 1 and the winner node of output SOM # 2 is updated, and output SOM # The connection weight between each output node of the output SOM # 2 and the winner node of the input SOM # 1 is updated around the 2 winner nodes.

  That is, the connection weight between each input node of the input SOM # 1 and the winner node of the output SOM # 2 is made stronger (enhanced) as the input distance between the patterns d closer to the winner node of the input SOM # 1 is shorter. Updated to Also, the connection weight between each output node of the output SOM # 2 and the winner node of the input SOM # 1 is updated so that the output node having a shorter inter-pattern distance d with the winner node of the output SOM # 2 is updated. .

  In other words, the connection weight between each input node of input SOM # 1 and the winner node of output SOM # 2 becomes weaker as the inter-pattern distance d from the winner node of input SOM # 1 is farther (weakened) Updated). Also, the connection weight between each output node of the output SOM # 2 and the winner node of the input SOM # 1 is also updated so that the output node whose pattern distance d from the winner node of the output SOM # 2 is farther becomes weaker. .

  Given a lot of teaching data, learning of input SOM # 1 and output SOM # 2 progresses, and when the connection weight is updated, input data (pattern) and output data (pattern) are determined by the connection weight. ) Can be obtained.

  Then, according to the input / output relationship model that relates input data and output data, given input data, it is possible to determine the winner node that best matches the input data in input SOM # 1. The output node of the output SOM # 2 having the strongest connection weight with the winner node can be determined. Furthermore, reconstructed output data obtained by reconstructing output data to be output with respect to input data can be obtained using the output node (the pattern storage model 21 included in the output node).

  Next, in the recognition generation apparatus in FIG. 29, recognition processing for generating reconstructed output data by recognizing input data and reconstructing output data to be output with respect to the input data based on the recognition result. This will be described with reference to the flowchart of FIG.

  In this recognition generation process, in step S211, input data is input to the recognition generation apparatus in FIG. 20, and the recognition generation apparatus recognizes the input data.

That is, the input data is supplied to the score calculation unit 231, the score calculation unit 231, at step S212, the relative input data, the scores for each node N _i of pattern storage SOM # 1 constituting the input-output relationship model Calculate and supply to the winner node determination unit 232.

In step S213, winning node determining unit 232 in the pattern storage SOM # 1 of nodes constituting the input-output relationship model, the highest node score from the score computing unit 231 determines the winning node N _i, the input label representing the winning node N _i, as a recognition result of the input data, and supplies the generated node determiner 233.

In step S214, the binding of the generation node determiner 233, in the pattern storing SOM # 2 that constitute the input-output relationship model node, a node N _i represented by the input label as a recognition result from the winning node determining unit 232 The node N ′ _j having the strongest weight is determined as a generation node, and an output label representing the generation node N ′ _j is supplied to the reconstruction unit 234.

In step S215, the reconfiguration unit 234 uses the node N ′ _j represented by the output label from the generation node determination unit 233 among the nodes of the pattern storage SOM # 2 constituting the input / output relationship model, that is, the generation node. The reconstructed output data obtained by reconstructing the output data as the estimated value of the output data to be output with respect to the input data supplied to the recognition generator is acquired in the same manner as the reconstructing unit 53 in FIG. Generated) and output. This completes the recognition generation process.

  Next, FIG. 24 schematically shows the storage contents of the main body model storage unit 202 and the division error model storage unit 204 in FIG. 18 when the input / output relation model is adopted as the main body model and the division error model. Show.

  24, the main body model storage unit 202 stores an input / output relation model as a main body model, and the division error model storage section 204 stores an input / output relation model as a division error model.

  Hereafter, the input / output relationship model as the main body model is also referred to as a main body input / output relationship model, and the input / output relationship model as the division error model is also referred to as an error input / output relationship model.

  The main body input / output relation model and the error input / output relation model are input / output relation models in which two pattern storage SOMs are connected and nodes of the two SOMs are combined.

  That is, the main body input / output relation model has a pattern storage SOM for learning the input data S and a pattern storage SOM for learning the output data A. Then, the pattern memory SOM for learning the input data S and the pattern memory SOM for learning the output data A are referred to as the input body SOM and the output body SOM, respectively. The nodes of the SOM and the output body SOM are coupled to each other, and the strength of coupling between the input body SOM and the output body SOM is represented by a coupling weight.

  Here, hereinafter, the connection weight of the main body input / output relation model is also referred to as a main body connection weight.

The error input / output relation model has a pattern storage SOM for learning the divided input error data ΔS _i and a pattern storage SOM for learning the divided output error data ΔA _i . The pattern storage SOM for learning the divided input error data ΔS _i and the pattern memory SOM for learning the divided output error data ΔA _i are referred to as a divided input error SOM and a divided output error SOM, respectively. In the error input / output relationship model, the nodes of the divided input error SOM and the divided output error SOM are combined, and the strength of the connection between the divided input error SOM and the divided output error SOM depends on the combination weight. expressed.

  Hereinafter, the connection weight of the error input / output relationship model is also referred to as an error connection weight as appropriate.

  FIG. 25 is a diagram for explaining data that can be input / output with respect to the main body input / output relation model and the error input / output relation model of FIG.

Input data S to be learned (input data S constituting teaching data) and input data S to be recognized are input to the input body SOM of the body input / output relation model. Further, the input body SOM can output reconstructed input data S _R obtained by reconfiguring the input data S with respect to the input data S input thereto.

To the output body SOM of the body input / output relation model, output data A (output data A constituting the teaching data together with the input data S inputted to the input body SOM) is inputted. In addition, the output body SOM outputs the reconstructed output data A _R obtained by reconstructing the output data A input thereto and the output data A to be output with respect to the input data S input to the input body SOM. be able to.

For the divided input error SOM of the error input / output relationship model, the error between the input data S input to the input body SOM and the reconstructed input data S _R output from the input body SOM with respect to the input data S The divided input error data ΔS _i is obtained by dividing the divided input error data ΔS, which is the above, for each predetermined dimension or every predetermined time.

For the divided output error SOM of the error input / output relationship model, the error between the output data A input to the output body SOM and the reconstructed output data A _R output from the output body SOM with respect to the output data A The divided output error data ΔA _i obtained by dividing the divided output error data ΔA is divided for each predetermined dimension or every predetermined time. Also, from the divided output error SOM, division and outputs error data △ A _i input thereto, the divided output error data △ A _i corresponding to the divided input error data △ S _i which is input to the divided input error SOM re The configured divided reconstruction output error data ΔA _Ri can be output.

Here, the divided output error data ΔA _i corresponding to the divided input error data ΔS _i is, as described above, the output data A to be output with respect to the input data S from which the divided input error data ΔS _i is obtained. The divided output error data ΔA _i obtained from

  Next, as shown in FIGS. 24 and 25, when the input / output relation model is adopted as the main body model and the division error model, the main body learning module 101, the error learning module 102, and the control unit in FIG. The learning process according to 103 will be described with reference to the flowchart of FIG.

  In this learning process, teaching data that is a set of input data S and output data A to be subjected to the learning process is supplied from the outside to the calculation unit 201 and the error calculation unit 205.

  In step S251, the arithmetic unit 201 learns the input data S constituting the supplied teaching data, that is, the input body SOM of the body input / output relation model stored in the body model storage unit 202 using the input data S. The self-organized learning of the input main body SOM is performed as described in steps S3 to S7 in FIG.

  In step S252, the calculation unit 201 learns the output data A (output data A to be output with respect to the input data S learned in step S251) constituting the supplied teaching data, that is, the output data A. The self-organized learning of the output main body SOM, which uses the self-organized update of the output main body SOM of the main body input / output relation model stored in the main body model storage unit 202, has been described in steps S3 to S7 in FIG. Do as follows.

  In step S <b> 253, the calculation unit 201 determines a winner node for the input data S from among the nodes constituting the input body SOM of the body input / output relation model stored in the body model storage unit 202. Here, the node having the highest score for the input data S among the nodes constituting the input body SOM is determined as the winner node for the input data S.

  Since the winner node for the input data S is determined by self-organized learning in step S251, the winner node can be employed as it is in step S253. However, in step S253, separately from step S251, for each node constituting the input body SOM, the score for the input data S can be obtained, and the node with the highest score can be determined as the winner node for the input data S. .

  In step S254, the arithmetic unit 201 determines a winner node for the output data A from the nodes constituting the output main body SOM of the main body input / output relation model stored in the main body model storage unit 202. Here, as in step S253, the winner node determined by the self-organizing learning in step S252 can be adopted as it is, and, apart from step S252, the output data for each node constituting the output main body SOM is used. A score for A can be obtained, and the node with the highest score can be determined as the winner node for the output data A.

  In step S255, as in the connection weight update unit 222 (FIG. 20), the calculation unit 201 obtains each node of the input body SOM from the winner node of the input body SOM obtained in step S253, and obtains in step S254. Update the body connection weight with the winner node of the output body SOM, and update the body connection weight between each node of the output body SOM and the winner node of the input body SOM, centering on the winner node of the output body SOM .

In step S256, the arithmetic unit 201 uses the winning node of the input body SOM determined in step S253, the reconstructed input data S _R which reconstructs the input data S, in the same manner as reconstruction unit 53 in FIG. 12 Generated (obtained) and supplied to the error calculation unit 205.

In step S257, the error calculation unit 205 uses the input data S from the outside and the reconstructed input data S _R output from the calculation unit 201 with respect to the input data S, and the input error data ΔS = SS _R. Is calculated and supplied to the data dividing unit 206. The data dividing unit 206 divides the input error data ΔS for each predetermined dimension or every predetermined time, and supplies the divided input error data ΔS _i obtained as a result to the arithmetic unit 203. .

In step S258, the calculation unit 203 learns the divided input error data ΔS _i from the data dividing unit 206, that is, uses the divided input error data ΔS _i to input the error input stored in the divided error model storage unit 204. Self-organized learning of the divided input error SOM for updating the divided input error SOM of the output relation model in a self-organized manner is performed as described in steps S3 to S7 in FIG.

In step S259, the calculation unit 201 uses the winner node of the output main body SOM determined in step S254 to reconstruct the reconstructed output data A _R in the same manner as the reconstructing unit 53 in FIG. Are generated (obtained) and supplied to the error calculator 205 and the adder 208.

In step S260, the error calculation unit 205 uses the output data A from the outside and the reconstructed output data A _R output from the calculation unit 201 with respect to the output data A, that is, output error data ΔA, that is, Output error data ΔA = AA _R corresponding to the divided input error data ΔS calculated in step S 257 is calculated and supplied to the data dividing unit 206. The data dividing unit 206 divides the output error data ΔA for each predetermined dimension or every predetermined time, and supplies the obtained divided output error data ΔA _i to the arithmetic unit 203. .

In step S261, the arithmetic unit 203 learns the divided output error data ΔA _i from the data dividing unit 206, that is, uses the divided output error data ΔA _i to input the error input stored in the divided error model storage unit 204. Self-organized learning of the divided output error SOM for updating the divided output error SOM of the output relation model in a self-organized manner is performed as described in steps S3 to S7 in FIG.

In step S262, the arithmetic unit 203 determines a winner node for the divided input error data ΔS _i from the nodes constituting the divided input error SOM of the error input / output relation model stored in the divided error model storage unit 204. . Here, as in step S253, the winner node determined by the self-organizing learning in step S258 can be adopted as it is, and, apart from step S258, for each node constituting the divided input error SOM, obtains a score for the input error data △ S _i, the highest node of the scores may be determined winning node for splitting the input error data △ S _i.

In step S263, the arithmetic unit 203 determines a winner node for the divided output error data ΔA _i from the nodes constituting the divided output error SOM of the error input / output relation model stored in the divided error model storage unit 204. . Here, as in step S253, the winner node determined by the self-organizing learning in step S261 can be used as it is, or, separately from step S261, for each node constituting the divided output error SOM, It obtains a score for output error data △ a _i, the highest node of the scores may be determined winning node for dividing the output error data △ a _i.

  In step S264, as in the connection weight updating unit 222 (FIG. 20), the calculation unit 203 performs the process of step S263 with each node of the divided input error SOM around the winner node of the divided input error SOM obtained in step S262. In addition to updating the error coupling weight with the winner node of the divided output error SOM obtained in step 1, with the winner node of the divided output error SOM as the center, each node of the divided output error SOM and the winner node of the divided input error SOM Update error coupling weights. This completes the learning process.

As described above, in the learning process of FIG. 26, in the main body input / output relation model, the input data S and the output data A to be originally learned are learned, and the input main body SOM and the output main body SOM of the main body input / output relation model The body connection weights for connecting the nodes are updated. Further, in the learning process of FIG. 26, in the error input / output relation model, the input data S constituting the teaching data and the reconstructed input data S _R obtained by reconstructing the input data S by the main body input / output relation model. The error between the input error data △ S that is the error between the output data A and the output data A that constitutes the teaching data, and the reconstructed output data A _{R that} is obtained by reconfiguring the output data S by the main body input / output relation model Output error data ΔA (output error data ΔA corresponding to input error data ΔS) divided into predetermined input data or divided input error data ΔS _i and divided output error data ΔA. Learning with _i is performed, and at the same time, error coupling weights for coupling nodes of the divided input error SOM and the divided output error SOM of the error input / output relation model are updated.

That is, in the learning process of FIG. 26, in addition to learning of the input data S and output data A, and updating of the body coupling weight relating the input data S and the output data A, the input data S is reconstructed. The divided input error data ΔS _i obtained by dividing the input error data ΔS, which is the error of the reconstructed input data S _R , and the reconstructed output data A obtained by reconstructing the output data A to be output for the input data S The learning of the divided output error data ΔA _i obtained by dividing the output error data ΔA, which is the error of _R , and the error coupling weight for associating the divided input error data ΔS _i with the divided output error data ΔA _i Updates are made.

The divided input error data ΔS _i and the divided output error data ΔA _i correspond to quantization errors caused by quantization and inverse quantization equivalently performed in the recognition generation process. In the learning process, the quantization error By learning with the error input / output relationship model, the recognition generation process can perform processing that appropriately considers this quantization error, and as a result, the quantization error generated in the recognition generation process can be efficiently processed. Can be reduced.

  Note that the learning process in FIG. 26 is performed each time teaching data that is a set of input data S and output data A to be subjected to the learning process is input, for example.

Further, when the learning process of FIG. 26 is performed by preparing a certain number of teaching data in advance, learning of the input data S and output data A, and updating of the body connection weight (hereinafter referred to as body learning as appropriate). ) and, learning of the divided input error data △ S _i and the divided output error data △ a _i, and the updating of the error connection weights (hereinafter, as appropriate, and is referred to as error learning), Good be carried out simultaneously, performed separately May be.

  That is, in the learning process, each time one piece of pre-prepared teaching data is given, both main body learning and error learning may be performed, or main body learning is performed using all the pre-prepared teaching data, Thereafter, error learning may be performed using all the prepared teaching data.

  Next, as shown in FIGS. 24 and 25, when the input / output relation model is adopted as the main body model and the division error model, the main body learning module 101, the error learning module 102, and the control unit in FIG. The recognition generation process by 103 will be described with reference to the flowchart of FIG.

  In this recognition generation process, input data S to be recognized is supplied from the outside to the calculation unit 201 and the error calculation unit 205.

  In step S281, the calculation unit 201 performs a recognition process for recognizing the supplied input data S based on the learning result of the learning process in FIG. That is, the computing unit 201 determines a winner node for the input data S from the nodes constituting the input body SOM of the body input / output relation model stored in the body model storage unit 202, and represents the winner node Label) is the recognition result of the input data S.

In step S282, the calculation unit 201 is used to generate (acquire) reconstructed output data A _R obtained by reconstructing the output data A to be output for the input data S based on the recognition result of the input data S. The generation node is determined from among the nodes constituting the output body SOM. That is, the arithmetic unit 201 detects a node of the output body SOM having the strongest body coupling weight with the winner node of the input body SOM represented by the node label as the recognition result of the input data S, and the node is input to the input data S. On the other hand, it is determined as a generation node used to generate reconstructed output data A _R obtained by reconstructing output data A to be output.

In step S283, the arithmetic unit 201, based on the recognition result of the input data S in step S281, and generates a reconstructed input data S _R which reconstructs the input data S, and supplies the error calculation unit 205. That is, the arithmetic unit 201 uses the winner node for the input data S as the recognition result of the input data S to reconstruct the reconfigured input data S _R with the reconstructing unit 53 of FIG. Similarly, it is generated (acquired) and supplied to the error calculation unit 205.

In step S284, the error calculation unit 205, calculates a supplied input data S, an input error data △ S and reconstructed input data S _R by the calculation unit 201 is reconstructed in step S283 with respect to the input data S Then, the data is supplied to the data dividing unit 206. The data dividing unit 206 divides the input error data ΔS for each predetermined dimension or every predetermined time, and generates the divided input error data ΔS _i obtained as a result of the calculation unit 203 and the correction data generation. Supplied to the unit 207.

In step S285, the calculation unit 203 performs a recognition process based on the learning result of the learning process of FIG. 26 based on the divided input error data ΔS _i from the data dividing unit 206. That is, the calculation unit 203 determines a winner node for the divided input error data ΔS _i from among the nodes constituting the divided input error SOM of the error input / output relation model stored in the divided error model storage unit 204, The winner node (representing the node label) is set as the recognition result of the divided input error data ΔS _i .

In step S286, the arithmetic unit 203 divides the input error data △ S _i of based on the recognition result, the divided input error data △ S divided output error data △ divided reconstructed output error data to reconstruct the A _i corresponding to the _i A generation node used to generate (acquire) ΔA _Ri is determined from the nodes constituting the divided output error SOM. That is, the arithmetic unit 203 detects a node of the divided output error SOM having the strongest error coupling weight with the winner node of the divided input error SOM represented by the node label as the recognition result of the divided input error data ΔS _i , and the node _Is determined as a generation node used to generate divided reconstructed output error data ΔA _Ri obtained by reconstructing the divided output error data ΔA _i corresponding to the divided input error data ΔS _i .

In step S287, the arithmetic unit 201, in step S282, the output body SOM determined to generate a node used to produce a reconstructed output data A _R reconstructed output data A to be output to the input data S using node, to reconfigure the output data a to be output to the input data S, and outputs the reconstructed output data a _R the resulting.

That is, the arithmetic unit 201 uses the node of the output body SOM having the strongest body connection weight with the winner node of the input body SOM for the input data S, and reconstructs the output data A to be output for the input data S. The configuration output data A _R is acquired (generated) and output in the same manner as the reconstruction unit 53 in FIG. 12 (reconstruction unit 234 in FIG. 20). The reconstructed output data A _R output from the calculation unit 201 is supplied to the error calculation unit 205 and the addition unit 208.

In step S288, the calculation unit 203 generates the divided reconstructed output error data ΔA _Ri obtained by reconstructing the divided output error data ΔA _i corresponding to the divided input error data ΔS _i in step S286. Using the node of the divided output error SOM determined as a node, the divided output error data ΔA _i corresponding to the input error data ΔS obtained from the input data S is reconstructed, and the resulting divided reconstructed output error Data △ A _Ri is output.

That is, the calculation unit 203 uses the node of the divided output error SOM having the strongest error coupling weight with the winner node of the divided input error SOM with respect to the divided input error data ΔS _i obtained from the input data S. the resulting divided input error data △ S _i divided reconstructed output error data △ a _Ri reconstituted divided output error data △ a _i corresponding to the reconstruction unit 234 of the reconstruction unit 53 (FIG. 20 in FIG. 12 ) To obtain (generate) and output. The divided reconstruction output data A _Ri output from the calculation unit 203 is supplied to the correction data generation unit 207.

In step S289, the correction data generation unit 207, based on the divided input error data △ S _i input from the data dividing unit 206 in step S284, it determines the reconstructed output data S dimension or time to be corrected in the _R, determined In accordance with the result, a correction weight for each predetermined dimension or time is determined, and the partial reconstructed output error data ΔA _Ri from the calculation unit 203 is multiplied by the correction weight to thereby obtain partial correction data ΔA. _Rj is generated and supplied to the adding unit 208.

In step S290, the adding unit 208 adds the reconstructed output data A _R from the arithmetic unit 201 and the partial correction data ΔA _Rj from the correction data generating unit 207 to add data A _R ′ = A _R + ΔA. _Rj is calculated and output as final reconstructed output data obtained by correcting the reconstructed output data A _R. This completes the recognition generation process.

As described above, reconstructed output data obtained by reconstructing the output data A to be output for the input data S using the main body input / output relation model in the recognition generation process, by performing main body learning and error learning in the learning process. A _R is a partial correction generated based on the divided reconstructed output error data ΔA _Ri obtained by reconstructing the divided output error data ΔA _i corresponding to the divided input error data ΔS _i using the error input / output relation model. The final reconstructed output data obtained by reconstructing the output data A to be output with respect to the input data S is corrected by adding the data △ A _Rj to the corrected data A _R + ΔA _Rj. (Final reconstructed output data) is output, so that a process that appropriately considers the error is performed. As a result, the quantization error can be reduced as described with reference to FIGS.

  By the way, as described above, the main body input / output relation model for learning the input data and output data to be originally learned (main body learning), and the division that divides the error between the input data and output data and the data reconstructed from them. An error input / output relationship model that performs learning (error learning) of input error data and divided output error data can be used instead of the input / output relationship model for a technique that can use the input / output relationship model. it can.

  As a technique that can use the input / output relationship model, for example, there is the above-described listening imitation learning. Therefore, in the following, instead of the (one) input / output relation model, listening imitation learning using the main body input / output relation model and the error input / output relation model will be described. A brief explanation of listening imitation learning using.

  In the listening imitation learning using one input / output relation model having two input SOMs and output SOMs, for example, the parameter of the node of the output SOM in the input / output relation model is changed, and the parameters after the change are Generates articulation parameters for generating audio data as output data, and recognizes acoustic parameters as feature parameters extracted from the audio data as input data by the input SOM. By repeating until the voice data input from SOM matches the recognition result in SOM, when there is input of unknown voice data from the outside, the articulation parameters for generating voice data imitating the voice data are It can be output as output data.

  Specifically, for example, in the input SOM of the input / output relation model, the acoustic parameters of the phonetic data “A”, “I”, “U” are learned, and the acoustic parameter pattern is acquired. At the same time, in the output SOM, learning of articulation parameters for generating speech data of phonemes `` a '', `` i '', `` u '' is performed, and when the articulation parameter pattern is acquired, from the user, When speech data of unknown phoneme “e” is given, the pattern acquired by a node with an output SOM is slightly changed in listening imitation learning, and speech data is generated from the articulation parameters of the changed pattern. The

  Furthermore, if an acoustic parameter is extracted from the speech data and the acoustic parameter is close to the acoustic parameter of the speech data of the unknown phoneme “e” given by the user, the unknown parameter given by the user is input in the input SOM. The acoustic parameter of the speech data of the phoneme “e” is learned, and the articulation parameter of the changed pattern is learned in the output SOM.

  When the voice data of unknown phoneme “e” is given by the user, the above processing is performed, so that the input SOM obtains the acoustic parameter pattern of the voice data of unknown phoneme “e”. At the same time, in the output SOM, articulation parameter patterns for generating speech data of the unknown phoneme “e” are acquired.

  FIG. 28 shows an outline of listening imitation learning using an input / output relationship model.

  Audio data from the outside is supplied to the feature extraction device 301. The feature extraction device 301 extracts acoustic parameters by analyzing external audio data, and outputs the time series of the acoustic parameters as input data S.

  The input data S output by the feature extraction device 301 is input to the pattern storage SOM # 1 as the input SOM of the input / output relation model. Then, for each node constituting the pattern storage SOM # 1, the score for the input data S is calculated, and the node N having the highest score is determined as the winner node for the input data S.

  That is, in the pattern storage SOM # 1, the input data S is abstracted (quantized) to the node (winner node) N having the highest score for the input data S and is called the winner node N, so to speak, input / output relationship Represented by the internal representation of the model.

Furthermore, in the listening imitation learning, the node N ′ having the strongest connection weight with the winner node N of the pattern storage SOM # 1 is detected from the nodes of the pattern storage SOM # 2 as the output SOM of the input / output relation model, The node N ′ is determined as a generation node used for generating (acquiring) reconstructed output data A _R obtained by reconfiguring output data A to be output with respect to input data S.

In the listening imitation learning, the generation of voice data (sound data) as reconstructed output data A _{R obtained} by reconstructing the output data A to be output with respect to the input data S using the generation node of the pattern storage SOM # 2 A time series of articulation parameters used for synthesis is generated.

The articulation parameters (the time series) as the reconstructed output data A _R are supplied to the speech synthesizer 302. Speech synthesizer 302 performs speech synthesis by using the articulatory parameters as reconstructed output data A _R, and outputs the synthesized speech data is audio data obtained as a result.

  The synthesized speech data output from the speech synthesizer 302 is supplied to the feature extraction device 301.

  The feature extraction device 301 analyzes the synthesized speech data from the speech synthesizer 302 to extract acoustic parameters, and outputs the time series of the acoustic parameters as new input data S.

  The new input data S output from the feature extraction device 301 is input to the pattern storage SOM # 1 as the input SOM of the input / output relation model. Then, the score for the new input data S is calculated for each node constituting the pattern storage SOM # 1, and the node with the highest score is determined as the winner node for the new input data S.

  If the winner node of pattern storage SOM # 1 for new input data S obtained from synthesized speech data does not match the winner node N for input data S obtained from external speech data (the recognition results match) If the synthesized speech data does not resemble external speech data, the parameters of the pattern storage model 21 (FIG. 6) of the generation node N ′ are changed.

  Specifically, for example, when the pattern storage model 21 is an HMM, the state transition probability and the output probability density function as parameters of the HMM are changed by a certain amount of change.

Then, using the changed parameters, the time series of the articulation parameters as reconstructed output data A _{R obtained} by reconstructing the output data A to be output with respect to the input data S is generated again, and is transmitted to the speech synthesizer 302. Supplied. Speech synthesizer 302 performs speech synthesis by using the articulatory parameters as reconstructed output data A _R, and outputs the new synthesized speech data.

  The new synthesized speech data output from the speech synthesizer 302 is supplied to the feature extraction device 301. The feature extraction device 301 analyzes the new synthesized speech data from the speech synthesizer 302 and extracts acoustic parameters. It is output as new input data S.

  Newer input data S output from the feature extraction device 301 is input to the pattern storage SOM # 1 as an input SOM of the input / output relation model, and a winner node for the input data S is determined.

  Then, the winner node of the pattern storage SOM # 1 for the input data S obtained from the synthesized voice data (and the new input data S) does not match the winner node N for the input data S obtained from the external voice data. In this case, the parameters of the pattern storage model 21 (FIG. 6) of the generation node N ′ are changed again, and the same processing is repeated thereafter.

Further, when the winner node of the pattern storage SOM # 1 for the input data S obtained from the synthesized speech data matches the winner node N for the input data S obtained from the speech data from the outside, that is, the synthesized speech data is When the synthesized voice data that is similar to the voice data from the outside and imitated the voice data from the outside can be output, the pattern storage SOM # 1 can receive the voice data from the outside (or from the outside. (Synthetic voice data similar to the voice data of), and is updated in a self-organized manner using the input data S as acoustic parameters obtained from the sound parameters, and the pattern storage SOM # 2 is similar to the voice data from the outside synthesized speech data is updated a self-organizing manner by using the reconstructed output data a _R as articulatory parameters obtained.

Furthermore, the winner node of the pattern storage SOM # 1 with respect to the input data S as the acoustic parameter obtained from the external audio data, and the articulation parameter from which the synthesized audio data similar to the external audio data was obtained around the winning node of pattern memory SOM # 2 to the reconstruction output data a _R, the coupling weight of the input-output relationship model is updated.

That is, it is similar to each node of the pattern storage SOM # 1 and the audio data from the outside, centering on the winner node of the pattern storage SOM # 1 with respect to the input data S as the acoustic parameters obtained from the audio data from the outside connection weight between the winning node of pattern memory SOM # 2 to the reconstruction output data a _R as articulatory parameters synthesized speech data is obtained is updated.

Further, around the winning node of pattern memory SOM # 2 to the reconstruction output data A _R as articulatory parameters synthesized speech data is obtained which is similar to the audio data from the outside, and each node of the pattern memory SOM # 2 Then, the connection weight with the winner node of the pattern storage SOM # 1 for the input data S as the acoustic parameter obtained from the sound data from the outside is updated.

  This completes the explanation of the outline of the listening imitation learning using the input / output relationship model.

  By the way, in the listening imitation learning using the input / output relation model as described above, the articulation parameter is changed in order to obtain synthesized voice data similar to the voice data from the outside. This is nothing other than searching for articulation parameters (hereinafter, referred to as target articulation parameters, as appropriate) for generating synthesized speech data that can obtain acoustic parameters similar to those obtained from the speech data.

  As a search method for searching for the target articulation parameter, for example, there is a method of performing a random search in which the articulation parameter is changed at random.

  However, since the articulation parameter used for speech synthesis is time-series data, it consists of a plurality of frames that are not necessarily a fixed number of frames. Furthermore, since the articulation parameters are generally multidimensional, and the number of values that can be taken by the multidimensional articulation parameters is extremely large, the articulation parameters having a large number of such possible values can be obtained by random search for each frame. Even if it is changed, it is unlikely that the changed articulation parameter approaches the target articulation parameter, and it may take a long time to search for the target articulation parameter.

  Therefore, a search method for searching for the target articulation parameter by changing the articulation parameter in a direction approaching the target articulation parameter from the past search results using the gradient method is conceivable.

  When using the gradient method, it is necessary to change the articulation parameters in a direction that reduces the error between the acoustic parameters of the synthesized speech data generated from the modified articulation parameters and the acoustic parameters of the external speech data. is there.

However, the acoustic parameters extracted by analyzing the speech data, that is, the input data S input to the pattern storage SOM # 1 in FIG. 28 and the articulation parameters used for speech synthesis, that is, the pattern storage SOM in FIG. If the modality differs from the reconstructed output data A _R (output data A) output from # 2 (for example, the acoustic parameter is a cepstrum coefficient and the articulation parameter is a linear prediction coefficient) The change in the error between the acoustic parameters of the synthesized speech data generated from the articulation parameters of the sound and the acoustic parameters of the external speech data is used as a gradient that determines the direction in which the articulation parameters having different modalities from the acoustic parameters are changed. It is difficult.

  In the articulation parameter space (articulation parameter space), when the articulation parameter is changed in any direction, the acoustic parameter of the synthesized speech data generated from the changed articulation parameter becomes the target acoustic parameter, that is, the external parameter. If it is known whether or not the sound parameter of the sound data from is approaching, the target articulation parameter can be searched quickly.

  That is, from the acoustic parameter gradient in the acoustic parameter space (acoustic parameter space) between the acoustic parameter of the synthesized speech data generated from the current articulation parameter and the acoustic parameter of the external speech data, the current articulation is performed. If the gradient of the articulation parameter in the articulation parameter space between the parameter and the target articulation parameter can be obtained, the target articulation parameter can be searched quickly.

  In listening imitation learning using the main body input / output relation model and the error input / output relation model (hereinafter also referred to as newspaper imitation learning as appropriate), the gradient in the articulation parameter space changes from the gradient in the acoustic parameter space to the error input / output relation model. Thus, the target articulation parameter is quickly searched for.

  FIG. 29 shows a configuration example (third configuration example) of the main body learning module 101, the error learning module 102, and the control unit 103 of FIG. 1 that performs newspaper imitation learning.

  In the figure, portions corresponding to those in FIG. 18 are denoted by the same reference numerals, and description thereof will be omitted as appropriate.

  The main body learning module 101 is common to the case of FIG. 18 in that it has a main body model storage unit 202, but differs from the case of FIG. 18 in that a calculation unit 321 is provided instead of the calculation unit 201.

  The error learning module 102 is common to the case of FIG. 18 in that it has a division error model storage unit 204, but is different from the case of FIG. 18 in that a calculation unit 323 is provided instead of the calculation unit 203. .

  The control unit 103 is common to the case of FIG. 18 in that it includes an error calculation unit 205, a data division unit 206, a correction data generation unit 207, and an addition unit 208, but further includes a listening imitation control unit 327. This is different from the case of FIG.

  For example, input data S and output data A are supplied to the calculation unit 321 of the main body learning module 101 as data D in FIG.

The computing unit 321 performs the same processing as the computing unit 201 in FIG. That is, the calculation unit 321 learns the main body model stored in the main body model storage unit 202 using the supplied input data S and output data A. Further, the operating section 321, the input data S supplied, by using a body model stored in the body model storage unit 202, and generates and reconstructed input data S _R which reconstructs the input data S Output. The arithmetic unit 321 uses the body model stored in the body model storage unit 202, generates the reconstructed output data A _R reconstructed output data A to be output to the input data S supplied And output. Further, the arithmetic unit 321, the output data A supplied by using a body model stored in the body model storage unit 202, and generates the reconstructed output data A _R reconstructed output data A Output.

  In addition, the calculation unit 321 performs processing according to control information supplied from the listening imitation control unit 327.

The calculation unit 323 of the error learning module 102 performs the same processing as the calculation unit 203 of FIG. In other words, the divided input error data ΔS _i and the divided output error data ΔA _i are supplied from the data dividing unit 206 as the error data ΔD in FIG.

Calculation unit 323, from the data dividing unit 206 using the divided input error data △ S _i and the divided output error data △ A _i, performs learning of the division error model stored in the division error model storage unit 204. Further, the calculation unit 323 uses the division error model stored in the division error model storage unit 204 to use the division output error data ΔA _i corresponding to the division input error data ΔS _i supplied from the data division unit 206. _Is generated and output as divided reconstructed output error data ΔA _Ri . Further, the calculation unit 323 uses the division error model stored in the division error model storage unit 204 for the division output error data ΔA _i supplied from the data division unit 206, and uses the division output error data ΔA. Divided reconstructed output error data ΔA _Ri reconstructed from _i is generated and output.

  In addition, the calculation unit 323 performs processing according to the control information supplied from the listening imitation control unit 327.

The listening imitation control unit 327 performs various processes for performing newspaper imitation learning. In other words, the listening imitation control unit 327 is supplied with reconstructed input data S _R output from the calculation unit 321, addition data A _R + ΔA _Rj output from the addition unit 208, target input data T, and the like. .

  Note that the target input data T corresponds to an acoustic parameter (target acoustic parameter) of audio data from the outside in the listening imitation learning described with reference to FIG.

  Here, in order to simplify the description, an expression S = g (A) using a certain function g () between a certain input data S and output data A to be output with respect to the input data S is expressed by Suppose you have a relationship that you have followed. Here, for example, as shown in FIG. 28, if the input data S is an acoustic parameter and the output data A is an articulation parameter, the function g () performs speech synthesis using the articulation parameter A, This corresponds to extracting the acoustic parameter S from the resultant synthesized speech data.

When the function value S = g (A) is referred to as input data corresponding to the output data A, in newspaper imitation learning, the addition data A _R ′ = A _R + ΔA that is the final reconstructed output data is used. _The addition data A _R ′ is searched so that the input data g (A _R ′) corresponding to _Rj is close to the target input data T (ideally matches).

For this search, the listening imitation control unit 327 is supplied with the addition data A _R ′ output from the addition unit 208, target input data T, and reconstructed input data S _R. The listening imitation control unit 327 determines whether or not the input data g (A _R ′) corresponding to the addition data A _R ′ has approached the target input data T.

Specifically, the difference between the input data g (A _R ′) corresponding to the addition data A _R ′ and the target input data T is such that the target input data T is supplied to the calculation unit 321 as the input data S. If the difference between the reconstructed input data S _R output from the calculation unit 321 and the target input data T correspondingly decreases, the input data g (A _R corresponding to the addition data A _R ′ It is determined that ') has approached the target input data T.

When the hearing imitation control unit 327 determines that the input data g (A _R ′) corresponding to the addition data A _R ′ has approached the target input data T, the listening imitation control unit 327 learns the main body model and the division error model. Therefore, the input data g (A _R ′) close to the input data T is set as the new input data S, and the addition data A _R ′ is supplied as the new output data A corresponding thereto to the arithmetic unit 321 and the like.

  Furthermore, the listening imitation control unit 327 includes a calculation unit 321, a calculation unit 323, and a correction data unit 207 in order to cause the calculation unit 321, the calculation unit 323, and the correction data unit 207 to perform processing necessary for newspaper imitation learning. Control information for control is supplied to the calculation unit 321, the calculation unit 323, and the correction data unit 207.

  Next, newspaper imitation learning processing by the main body learning module 101, the error learning module 102, and the control unit 103 in FIG. 29 will be described with reference to the flowchart in FIG.

  Here, as shown in FIGS. 24 and 25, as the main body model stored in the main body model storage unit 202, a main body input / output relation model that is an input / output relation model is adopted, and a division error is also generated. Assume that an error input / output relationship model, which is an input / output relationship model, is adopted as the division error model stored in the model storage unit 204.

  For the main body input / output relation model and the error input / output relation model, the learning process of FIG. 26 using a certain number of teaching data is performed, and the main body input / output relation model and the error input / output relation model are It is assumed that at least a part of the patterns that can be acquired by the main body input / output relation model and the error input / output relation model has already been acquired.

  In this newspaper imitation learning, in step S301, for example, target input data T (hereinafter also referred to as target data T as appropriate) is supplied from the outside. The listening imitation control unit 327 supplies the supplied target data T as input data S to the calculation unit 321 and the error calculation unit 205 and controls the calculation unit 321 and the calculation unit 323 so as to perform recognition generation processing. .

  According to the control from the listening imitation control unit 327, the calculation unit 321 performs the recognition generation process described with reference to FIG.

  That is, in step S302, as described with reference to FIG. 27, the calculation unit 321 includes the calculation unit 321 from among the nodes constituting the input body SOM of the body input / output relation model stored in the body model storage unit 202. A winner node for the input data S (here, target data T) is determined, and the winner node (a node label representing the winner node) is set as a recognition result of the input data S.

In step S303, the calculation unit 321 detects the node of the output body SOM having the strongest body connection weight with the winner node of the input body SOM represented by the node label as the recognition result of the input data S, and the node is detected as the input data. The generation node used to generate the reconstructed output data A _R obtained by reconfiguring the output data A to be output to S is determined.

In step S304, the calculation unit 321 uses the winner node for the input data S as the recognition result of the input data S to reconstruct the input data S _R from the reconstructed input data S _R in FIG. 53, and is supplied to the error calculation unit 205 and the listening imitation control unit 327.

In step S305, the error calculation unit 205 inputs the input data S (here, target data T) from the listening imitation control unit 327 and the reconstructed input data S _R reconstructed by the calculation unit 321 with respect to the input data S. The input error data ΔS is calculated and supplied to the data dividing unit 206. The data dividing unit 206 divides the input error data ΔS for each predetermined dimension or every predetermined time, and the divided input error data ΔS _i obtained as a result is generated by the calculation unit 323 and the correction data generation. Supplied to the unit 207.

In step S 306, the calculation unit 323 divides the divided input error data ΔS _i from the data dividing unit 206 from the nodes constituting the divided input error SOM of the error input / output relation model stored in the divided error model storage unit 204. The winner node is determined, and the winner node (representing the node label) is set as the recognition result of the divided input error data ΔS _i .

In step S307, the calculation unit 323 detects the node of the divided output error SOM having the strongest error coupling weight with the winner node of the divided input error SOM represented by the node label as the recognition result of the divided input error data ΔS _i . The node is determined as a generation node used to generate divided reconstructed output error data ΔA _Ri obtained by reconstructing the divided output error data ΔA _i corresponding to the divided input error data ΔS _i .

In step S308, the calculation unit 321 outputs the output body having the strongest body connection weight with the winner node of the input body SOM for the input data S (here, target data T) from the listening imitation control unit 327 determined in step S303. Using the SOM generation node, the reconstructed output data A _R obtained by reconstructing the output data A to be output with respect to the input data S from the listening imitation control unit 327 is converted into the reconstructing unit 53 (reconstructed in FIG. 20). This is generated in the same manner as the configuration unit 234) and supplied to the error calculation unit 205 and the addition unit 208.

In step S309, the calculation unit 323 uses the generation node of the divided output error SOM having the strongest error coupling weight with the winner node of the divided input error SOM for the divided input error data ΔS _i determined in step S307. The divided reconstructed output error data ΔA _Ri obtained by reconstructing the divided output error data ΔA _i corresponding to the error data ΔS _i is the same as the reconstructing unit 53 in FIG. 12 (the reconstructing unit 234 in FIG. 20). Generated and supplied to the correction data generation unit 207.

In step S310, the correction data generation unit 207, based on the divided input error data △ S _i input from the data dividing unit 206 in step S305, it determines the reconstructed output data S dimension or time to be corrected in the _R, determined According to the result, a correction weight for each predetermined dimension or time is determined.

In step S <b> 311, the correction data generation unit 207 multiplies each division reconstruction output error data ΔA _Ri from the calculation unit 203 by the correction weight for each predetermined dimension or time determined previously. , Partial correction data ΔA _Rj is generated and supplied to the adding unit 208.

In step S312, the adding unit 208 adds the reconstructed output data A _R supplied from the calculation unit 321 in step S308 and the partial correction data ΔA _Rj supplied from the correction data generating unit 207 in step S311. The addition data A ′ _R = A _R + ΔA _Rj obtained as a result is calculated as final reconstructed output data obtained by correcting the reconstructed output data A _R and supplied to the listening imitation control unit 327.

As described above, the final reconstruction output data A _R ′ is supplied from the adder 208 and the reconstruction input data S _R is supplied from the operation unit 321 to the listening imitation control unit 327, and the process proceeds to step S313.
In step S313, the listening imitation control unit 327 determines whether or not the acoustic parameter series of speech synthesized by the final reconstructed output data A _R ′ has approached the target input data T.

Specifically, sound imitation control unit 327, the input data g of 'a, the final reconstructed output data A _R' final reconstructed output data A _R supplied from the adder 208 corresponding to (A _R ') ( (Hereinafter also referred to as “corresponding data g (A _R ′)”), the difference between the input data g (A _R ′) and the target input data T is calculated as the target input data T as the input data S. If the difference is smaller than the difference between the reconstructed input data S _R supplied to 321 and output from the operation unit 321 corresponding thereto and the target input data T, the speech synthesized by the final reconstructed output data A _R ′ Is determined to have approached the target input data T.

If it is determined in step S313 that the acoustic parameter series of the speech synthesized by the final reconstructed output data A _R ′ is not approaching the target input data T, the process proceeds to step S314. In step S314, the correction data generation unit 207 randomly determines correction weights used when generating the partial correction data ΔA _Rj according to the control information from the listening imitation control unit 327. Thereafter, the processing returns to step S311 and the subsequent processing is repeated. If it is determined in step S313 that the acoustic parameter series of the speech synthesized by the final reconstructed output data A _R ′ has approached the target input data T, the process proceeds to step S315.

In step S315, the listening imitation control unit 327 supplies the corresponding data g (A _R ') corresponding to the final reconstructed output data A _R ' as input data S to the calculation unit 321 and the error calculation unit 205, The final reconstructed output data A _R ′ is supplied as output data A to the calculation unit 321 and the error calculation unit 205, and the calculation unit is configured to perform learning processing using the set of input data S and output data A as teaching data. 321 and the calculation unit 323 are controlled.

As a result, the calculation unit 321 and the calculation unit 323 perform the learning process described with reference to FIG. 26 using the corresponding data g (A _R ′) as the input data S and the addition data A _R ′ as the output data A. Is called.

  Thereafter, the process returns to step S302, and the subsequent processes are repeated. Therefore, the learning of the main body input / output relation model and the learning of the miscalculation input / output relation model are advanced. More accurate (similar) listening imitation can be performed.

  As described above, the error between the data learned by the main body model and the reconstructed data reconstructed using the main body model is learned by the split error model. It can be performed.

  That is, as described above, reconstructing highly accurate data with reduced quantization error, searching for data similar to target data more quickly (efficiently) than random search, and memory It can be used efficiently.

  In particular, since the error from the reconstructed data obtained by reconstructing data using the main body model is divided and learned every predetermined dimension or every time, more various corrections can be performed.

  Furthermore, when SOM (input / output relation model using SOM) is adopted as the main body model and the division error model, it is handled by the same system regardless of scale (data, granularity, etc.). Can be handled regardless of the dynamic range).

  Next, the series of processes described above can be performed by hardware or software. When a series of processing is performed by software, a program constituting the software is installed in a general-purpose computer or the like.

  FIG. 31 is a block diagram illustrating a hardware configuration example of a computer that executes the above-described series of processing by a program.

  In the computer 500, a CPU (Central Processing Unit) 501, a ROM (Read Only Memory) 502, and a RAM (Random Access Memory) 503 are connected to each other by a bus 504.

  An input / output interface 505 is further connected to the bus 504. The input / output interface 505 includes an input unit 506 including a keyboard, a mouse, and a microphone, an output unit 507 including a display and a speaker, a storage unit 508 including a hard disk and a non-volatile memory, and a communication unit 509 including a network interface. A drive 510 for driving a removable medium 511 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory is connected.

  In the computer 500 configured as described above, for example, the CPU 501 loads the program stored in the storage unit 508 to the RAM 503 via the input / output interface 505 and the bus 504 and executes the program. A series of processing is performed.

  The program executed by the computer may be a program that is processed in time series in the order described in this specification, or in parallel or at a necessary timing such as when a call is made. It may be a program for processing.

  The program may be processed by a single computer, or may be distributedly processed by a plurality of computers. Furthermore, the program may be transferred to a remote computer and executed.

  Further, in this specification, the system represents the entire apparatus constituted by a plurality of apparatuses.

  The embodiment of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention.

  Further, the present invention can be applied to a user interface (human interface) in, for example, an AV (AudioVisual) device such as a television receiver and an HD recorder, a computer, and other devices in addition to a robot.

It is a block diagram which shows the structural example of one Embodiment of the input / output system to which this invention is applied. It is a flowchart explaining a learning process. It is a flowchart explaining a recognition production | generation process. FIG. 2 is a block diagram illustrating a first configuration example of the input / output system of FIG. 1. It is a figure which shows the 1st structural example of pattern memory | storage SOM. It is a figure which shows the structural example of a node. It is a figure which shows the 2nd structural example of pattern memory | storage SOM. It is a block diagram which shows the structural example of a learning apparatus. It is a figure for demonstrating the determination method which determines an update weight. It is a figure for demonstrating the update method which updates the learning data memorize | stored in the learning data storage part. It is a flowchart explaining a learning process. It is a block diagram which shows the structural example of a recognition production | generation apparatus. It is a flowchart explaining a recognition production | generation process. It is a figure which shows typically the main body model and division | segmentation error model which employ | adopted pattern memory | storage SOM. It is a figure explaining the data which can be input / output with respect to the main body SOM and the division | segmentation error SOM. It is a flowchart explaining a learning process. It is a flowchart explaining a recognition production | generation process. It is a block diagram which shows the 2nd structural example of the input / output system of FIG. It is a figure which shows the structural example of an input / output relationship model. It is a block diagram which shows the structural example of the learning apparatus which performs the process using an input-output relationship model, and a recognition production | generation apparatus. It is a flowchart explaining the learning process of an input / output relationship model. It is a figure which shows a connection weight matrix. It is a flowchart explaining the recognition production | generation process using an input / output relationship model. It is a figure which shows typically the main body model and error model in which the input / output relation model was employ | adopted. It is a figure explaining the data which can be input / output with respect to a main body input / output relationship model and an error input / output relationship model. It is a flowchart explaining a learning process. It is a flowchart explaining a recognition production | generation process. It is a figure for demonstrating listening imitation learning. It is a block diagram which shows the 3rd structural example of the input / output system of FIG. It is a flowchart explaining the process of new hearing imitation learning. It is a block diagram which shows the structural example of one Embodiment of the computer to which this invention is applied.

Explanation of symbols

  21 pattern storage model, 22 learning data storage unit, 41 score calculation unit, 42 winner node determination unit, 43 weight determination unit, 44 learning data update unit, 45 model learning unit, 51 score calculation unit, 52 winner node determination unit, 53 Reconfiguration unit, 100 learning module, 101 main body learning module, 102 error learning module, 103 control unit, 121 arithmetic unit, 122 main body model storage unit, 123 arithmetic unit, 124 division error model storage unit, 125 error calculation unit, 126 data Dividing unit, 127 correction data generating unit, 128 adding unit, 201 calculating unit, 202 main body model storage unit, 203 calculating unit, 204 dividing error model storing unit, 205 error calculating unit, 206 data dividing unit, 207 correction data generating unit, 208 Adder unit, 221 learning unit, 222 connection weight update unit, 231 score calculation unit, 232 winner node determination unit, 233 generation node determination unit, 234 reconstruction unit, 301 feature extraction device, 302 speech synthesizer, 321, 323 calculation unit , 327 Hearing imitation control unit

Claims (21)

In a data processing apparatus that performs a learning process for learning data,
A body learning module to learn data,
An error learning module comprising: an error learning module that learns a division error in which an error between the data and the reconstructed data obtained by reconstructing the data by the main body learning module is divided every predetermined dimension or time Processing equipment. The main body learning module performs learning of input data, and learning of output data to be output for the input data,
The error learning module is an input division error in which an input error that is an error between the input data and reconstructed input data obtained by reconstructing the input data by the main body learning module is divided for each predetermined dimension or time Learning, and an output division error in which an output error that is an error between the output data and the reconstructed output data obtained by reconstructing the output data by the main body learning module is divided for each predetermined dimension or time. The data processing apparatus according to claim 1, wherein learning is performed. The body learning module
It has a main body input / output relationship model in which nodes of an input main body SOM that is a SOM (Self-Organization Map) for learning the input data and an output main body SOM that is a SOM for learning the output data are combined. ,
Self-organized learning of the input body SOM is performed using the input data,
Self-organized learning of the output body SOM is performed using the output data,
The degree of coupling between the winner node for the input data of the plurality of nodes constituting the input body SOM and the winner node for the output data of the plurality of nodes constituting the output body SOM is increased. And update the connection weight that represents the degree of connection between the nodes,
The error learning module
An error input / output relationship model in which nodes of a divided input error SOM that is a SOM for learning the divided input error and a divided output error SOM that is a SOM for learning the divided output error are combined;
Self-organized learning of the divided input error SOM is performed using the divided input error,
Self-organized learning of the divided output error SOM is performed using the divided output error,
Of a plurality of nodes constituting the divided input error SOM, a winner node for the divided input error and a combination of a winner node for the divided output error among the plurality of nodes constituting the divided output error SOM. The data processing apparatus according to claim 2, wherein a connection weight representing a degree of connection between nodes is updated so as to increase the degree. The body learning module further includes:
A winner node for the input data is determined from a plurality of nodes constituting the input body SOM,
Using the winner node of the input body SOM, reconstructing the input data, outputting the reconstructed input data,
Using the node of the output body SOM that has the strongest connection with the winner node of the input body SOM, reconstructing the output data to be output for the input data, and outputting the reconstructed output data,
The error learning module further includes:
A winner node for the divided input error is determined from among a plurality of nodes constituting the divided input error SOM,
Using the node of the divided output error SOM that has the strongest coupling with the winner node of the divided input error SOM, reconstructing the divided output error, and outputting a divided reconstructed output error,
The data processing apparatus according to claim 3, wherein the learning module adds and outputs the reconstruction output data and partial correction data determined based on the divided reconstruction output error. The body learning module
It has a main body input / output relationship model in which nodes of an input main body SOM that is a SOM (Self-Organization Map) for learning the input data and an output main body SOM that is a SOM for learning the output data are combined. ,
A winner node for the input data is determined from a plurality of nodes constituting the input body SOM,
Using the winner node of the input body SOM, reconstructing the input data, outputting the reconstructed input data,
Using the node of the output body SOM that has the strongest connection with the winner node of the input body SOM, reconstructing the output data to be output for the input data, and outputting the reconstructed output data,
The division error learning module includes:
An error input / output relationship model in which nodes of a divided input error SOM that is a SOM for learning the divided input error and a divided output error SOM that is a SOM for learning the divided output error are combined;
A winner node for the divided input error is determined from among a plurality of nodes constituting the divided input error SOM,
Using the node of the divided output error SOM that has the strongest coupling with the winner node of the divided input error SOM, reconstructing the divided output error, and outputting a divided reconstructed output error,
The body learning module
Self-organizing learning of the input body SOM is performed using correspondence data that is input data corresponding to addition data obtained by adding the reconstruction output data and the partial correction data determined based on the divided reconstruction output error Done,
Self-organized learning of the output body SOM is performed using the addition data,
The degree of coupling between the winner node for the corresponding data among the plurality of nodes constituting the input body SOM and the winner node for the added data among the plurality of nodes constituting the output body SOM is increased. And updating the connection weight of the body input / output relation model,
The error learning module
In the self-organized learning of the divided input error SOM, the error between the correspondence data and the reconstruction correspondence data obtained by reconstructing the correspondence data by the main body learning module is divided for each predetermined dimension or time. Is used as the split input error,
In the self-organized learning of the divided output error SOM, an error between the addition data and reconstructed addition data obtained by reconstructing the addition data by the main body learning module is divided for each predetermined dimension or time. Is used as the divided output error,
Of a plurality of nodes constituting the divided input error SOM, a winner node for the divided input error and a combination of a winner node for the divided output error among the plurality of nodes constituting the divided output error SOM. The data processing apparatus according to claim 2, wherein the connection weight of the error input / output relation model is updated so as to increase the degree. When the correspondence data is more similar to the target input data than the input data corresponding to the reconstructed output data,
The main body learning module performs self-organized learning of the input main body SOM, self-organized learning of the output main body SOM, and update of the connection weight of the main body input / output relation model,
The error learning module performs self-organized learning of the divided input error SOM, self-organized learning of the divided output error SOM, and update of the connection weight of the error input / output relation model. The data processing apparatus described. The body learning module further includes:
Reconstructing the input data and outputting the reconstructed input data;
Reconstructing output data to be output with respect to the input data, and outputting the reconstructed output data;
The error learning module further includes:
Reconstructing the divided output error corresponding to a divided input error in which an error between the input data and the reconstructed input data obtained by reconstructing the input data by the main body learning module is divided for each predetermined dimension or time To output the split reconstruction output error,
The data processing apparatus according to claim 2, wherein the learning module adds and outputs the reconstruction output data and partial correction data determined based on the divided reconstruction output error. The body learning module
Reconstructing the input data and outputting the reconstructed input data;
Reconstructing the output data to be output with respect to the input data, and outputting the reconstructed output data;
The error learning module includes:
Reconstructing the divided output error corresponding to the divided input error in which an error between the input data and the reconstructed input data output by the main body learning module is divided for each predetermined dimension or time, Output split reconstruction output error,
The body learning module
As learning of the input data, learning of corresponding data which is input data corresponding to addition data obtained by adding the reconstructed output data and partial correction data determined based on the divided reconstructed output error,
As the learning of the input data, the addition data is learned,
The error learning module
As learning of the divided input error, learning is performed in which the error between the correspondence data and the reconstructed correspondence data in which the main body learning module reconstructs the correspondence data is divided for each predetermined dimension or time,
The learning of the output error is performed when the error between the addition data and the reconstructed addition data obtained by reconstructing the addition data by the main body learning module is divided for each predetermined dimension or time. 2. A data processing apparatus according to 2. The body learning module
A main body SOM that is a SOM (Self-Organization Map) for learning the data,
Perform self-organized learning of the main body SOM using the data,
The error learning module
A division error SOM that is a SOM for learning the division error;
The data processing apparatus according to claim 1, wherein self-organized learning of the division error SOM is performed using the division error. The body learning module further includes:
From among a plurality of nodes constituting the main body SOM, a winner node for the data is determined,
Using the winner node of the main body SOM, reconstructing the data, outputting the reconstructed data,
The error learning module further includes:
A winner node for the division error is determined from a plurality of nodes constituting the division error SOM,
Using the winner node of the split error SOM, reconstruct the split error, and output a split reconstruction error,
The data processing apparatus according to claim 9, wherein the learning module adds and outputs the reconstruction data and partial correction data determined based on the division reconstruction error. The main body learning module further reconstructs the data and outputs the reconstructed data,
The error learning module further reconstructs the division error in which the error between the data and the reconstructed data output from the main body learning module is divided for each predetermined dimension or time, Output reconstruction error,
The data processing device according to claim 1, wherein the learning module adds and outputs the reconstruction data and partial correction data determined based on the division reconstruction error. In a data processing method of a data processing apparatus that performs learning processing for learning data,
In the main body learning module for learning data, learning the data,
An error learning module that learns a division error in which an error between the data and reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time, and learning the division error A data processing method including steps. In a program that causes a computer to function as a data processing device that performs learning processing for learning data,
A body learning module to learn data,
An error learning module comprising: an error learning module that learns a division error in which an error between the data and the reconstructed data obtained by reconstructing the data by the main body learning module is divided every predetermined dimension or time A program that causes a computer to function as a processing device. In a data processing device that outputs data,
The main body learning module that has learned input data and learned output data to be output with respect to the input data;
Learning of divided input error in which an error between the input data and reconstructed input data obtained by reconstructing the input data by the main body learning module is divided for each predetermined dimension or time, and the output data, An error learning module that learns a divided output error in which an error from the reconstructed output data in which the main body learning module reconstructs the output data is divided for each predetermined dimension or time. ,
The body learning module
Reconstructing the input data and outputting the reconstructed input data;
Reconstructing output data to be output with respect to the input data, and outputting the reconstructed output data;
The error learning module includes:
Reconstructing the divided output error corresponding to the divided input error, and outputting a divided reconstructed output error;
The learning module adds and outputs the reconstruction output data and partial correction data determined based on the divided reconstruction output error. The body learning module
It has a main body input / output relationship model in which nodes of an input main body SOM that is a SOM (Self-Organization Map) for learning the input data and an output main body SOM that is a SOM for learning the output data are combined. ,
A winner node for the input data is determined from a plurality of nodes constituting the input body SOM,
Using the winner node of the input body SOM, reconstructing the input data, outputting the reconstructed input data,
Using the node of the output body SOM that has the strongest connection with the winner node of the input body SOM, reconstructing the output data to be output for the input data, and outputting the reconstructed output data,
The error learning module
An error input / output relationship model in which nodes of a divided input error SOM that is a SOM for learning the divided input error and a divided output error SOM that is a SOM for learning the divided output error are combined;
From among a plurality of nodes constituting the divided input error SOM, determine a winner node for the divided input error,
Using the node of the divided output error SOM that has the strongest coupling with the winner node of the divided input error SOM, reconstructing the divided output error, and outputting a divided reconstructed output error,
The data processing apparatus according to claim 14, wherein the learning module adds and outputs the reconstruction output data and partial correction data determined based on the divided reconstruction output error. In a data processing method of a data processing device that outputs data,
The data processing device includes:
The main body learning module that has learned input data and learned output data to be output with respect to the input data;
Learning of divided input error in which an error between the input data and reconstructed input data obtained by reconstructing the input data by the main body learning module is divided for each predetermined dimension or time, and the output data, An error learning module that learns a divided output error in which an error from the reconstructed output data in which the main body learning module reconstructs the output data is divided for each predetermined dimension or time. ,
In the main body learning module,
Reconstructing the input data and outputting the reconstructed input data;
Reconstructing output data to be output with respect to the input data, and outputting the reconstructed output data;
In the error learning module,
Reconstructing the divided output error corresponding to the divided input error, and outputting a divided reconstructed output error;
A data processing method comprising a step of adding and outputting the division reconstruction output data and partial correction data determined based on the division reconstruction output error in the learning module. In a program that causes a computer to function as a data processing device that outputs data,
The main body learning module that has learned input data and learned output data to be output with respect to the input data;
Learning of divided input error in which an error between the input data and reconstructed input data obtained by reconstructing the input data by the main body learning module is divided for each predetermined dimension or time, and the output data, An error learning module that learns a divided output error in which an error from the reconstructed output data in which the main body learning module reconstructs the output data is divided for each predetermined dimension or time. ,
The body learning module
Reconstructing the input data and outputting the reconstructed input data;
Reconstructing output data to be output with respect to the input data, and outputting the reconstructed output data;
The error learning module
Reconstructing the divided output error corresponding to the divided input error, and outputting a divided reconstructed output error;
The learning module is a program that causes a computer to function as a data processing device that adds and outputs the reconstruction output data and partial correction data determined based on the divided reconstruction output error. In a data processing device that outputs data,
The main body learning module that learned the data,
An error learning module comprising: an error learning module that learns a division error in which an error between the data and the reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time;
The main body learning module reconstructs the data and outputs the reconstructed data,
The error learning module reconstructs the division error and outputs a division reconstruction error;
The learning module adds and outputs the reconstruction data and partial correction data determined based on the division reconstruction error. The body learning module
A main body SOM that is a SOM (Self-Organization Map) for learning the data,
From among a plurality of nodes constituting the main body SOM, a winner node for the data is determined,
Using the winner node of the main body SOM, reconstructing the data, outputting the reconstructed data,
The error learning module includes:
A division error SOM that is a SOM for learning the division error;
A winner node for the division error is determined from a plurality of nodes constituting the division error SOM,
Using the winner node of the split error SOM, reconstruct the split error, and output a split reconstruction error,
The data processing device according to claim 18, wherein the learning module adds and outputs the reconstruction data and partial correction data determined based on the division reconstruction error. In a data processing method of a data processing device that outputs data,
The data processing device includes:
The main body learning module that learned the data,
An error learning module comprising: an error learning module that learns a division error in which an error between the data and the reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time;
In the main body learning module, the data is reconstructed and the reconstructed data is output,
In the error learning module, the division error is reconstructed and a division reconstruction error is output,
A data processing method comprising the step of adding and outputting the reconstruction data and partial correction data determined based on the division reconstruction error in the learning module. In a program that causes a computer to function as a data processing device that outputs data,
The main body learning module that learned the data,
An error learning module comprising: an error learning module that learns a division error in which an error between the data and the reconstructed data obtained by reconstructing the data by the main body learning module is divided for each predetermined dimension or time;
The main body learning module reconstructs the data and outputs the reconstructed data,
The error learning module reconstructs the division error and outputs a division reconstruction error;
The learning module is a program that causes a computer to function as a data processing device that adds and outputs the reconstruction data and partial correction data determined based on the division reconstruction error.

Images