JP2007280058A - Information processor, information processing method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To efficiently handle multidimensional time-series data. <P>SOLUTION: A network storage part stores a network composed of a plurality of nodes, each node retaining dynamics modeled by a dynamic approximation model. An input weight adjustment part 82 adjusts, for each dimension, an input weight factor inputted to an input layer unit of the dynamic approximation model, which is a dimension-based weight factor for input data that is time-series data observed of two or more dimensions. An output weight adjustment part 83 adjusts, for each dimension, an output weight factor outputted from an output layer unit of the dynamic approximation model, which is a dimension-based weight factor for output data of two or more dimensions. The information processor is applicable to, for example, a robot. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an information processing device, an information processing method, and a program, and more particularly to an information processing device, an information processing method, and a program that can efficiently handle multidimensional time-series data.

Robot behavior (movement) is a dynamical system determined by the law of time evolution.
s), and various behaviors are known to be realizable by certain attractor dynamics.

  For example, the walking motion of a biped robot such as a human can be described as limit cycle dynamics, which is characterized by the movement state of the system from various initial states to a specific periodic orbit. (For example, refer nonpatent literatures 1 and 2.). In addition, the reaching motion that extends the arm to a certain object is described as fixed-point dynamics, which is characterized by the fact that it settles to a specific fixed point from various initial states. Can do. Furthermore, it is said that all the movements can be realized by a combination of a discrete movement that can be realized by the fixed point dynamics and a cyclic movement that can be realized by the limit cycle dynamics.

  The problem with controlling the robot's behavior (movement) using this attractor dynamics is that how to configure the attractor dynamics according to the task, and then to the attractor dynamics based on the information obtained through sensor input. Therefore, it is necessary to generate the corresponding motor output, and in order to realize this, it is necessary to generate the robot action output in such a way that the attractor dynamics continuously interact with the environment.

  A method for learning such attractor dynamics instead of human design has been proposed. One method is to use a recurrent neural network (hereinafter referred to as RNN). It is known that RNN has a context unit connected to a network by a regression loop and can theoretically approximate any dynamic system by holding an internal state there.

  However, in a learning model that consists of one network module that is tightly coupled, when learning a large number of dynamics necessary for large-scale behavioral learning, the interference between the dynamics to be memorized is very large. There is a problem that it is difficult.

  Therefore, several learning models have been proposed that employ a modular architecture in which a plurality of network modules are combined to form one learning model. In principle, this modular architecture can easily increase the dynamics that can be stored by increasing the number of modules. However, the problem of module selection that determines which module is used to learn a given learning sample arises.

  Depending on the method of module selection, the learning method can be divided into supervised learning, in which a human determines which module a learning sample (learning data) is assigned to, and unsupervised learning, in which the learning model is autonomously determined. However, in order for robots and systems to learn autonomously, it is necessary to learn modules by unsupervised learning.

  As one of methods for learning a module by unsupervised learning, a learning model called “Mixture of RNN Expert” has been proposed (see, for example, Patent Document 1). In this learning model, the outputs of multiple RNN modules are integrated by a gating mechanism to determine the final output, and the gate is obtained by maximum likelihood estimation so that the performance of the final output is maximized. While learning, learn each RNN module.

  However, such a method based on global optimization has a problem that learning becomes difficult when the number of modules becomes large.

  On the other hand, a self-organization map (hereinafter referred to as SOM) (see, for example, Non-Patent Document 3) and neural-gas (for example, non-use) used for vector pattern category learning. In a method such as Patent Literature 4), a learning rule based on global optimization is not used, and optimality is not guaranteed, but it is possible to learn an appropriate category structure in an unsupervised manner in a self-organizing manner. Are known. In these methods, even when the number of modules becomes large, learning is practically possible.

JP-A-11-126198 G. Taga, 1998, “Self-organized control of bipedal locomotion by neural oscillators in unpredictable environment”, Biological Cybernetics, 65, 147-159 By Kentaro Taga, “Dynamic Design of Brain and Body-Nonlinear Dynamical System of Motion and Perception and Development”, Kaneko Shobo T. Kohonen, “Self-Organizing Map”, Springer Fairlark Tokyo T.M.Martinetz, S.G.Berkovich, K.J.Schulten, "" Neural-Gas "Network for Vector Quantization and its Application to Time-Series Prediction", IEEE Trans. Neural Networks, VOL.4, NO.4, p558-569, 1999

  The applicant of the present application has previously proposed a model for learning time series patterns instead of vector patterns (Japanese Patent Application No. 2004-353832).

  However, a method for efficiently handling multidimensional time series data has not been proposed.

  The present invention has been made in view of such a situation, and makes it possible to efficiently handle multi-dimensional time-series data.

  An information processing apparatus according to an aspect of the present invention includes a storage unit that stores dynamics in one node and stores a network including the plurality of nodes, and a plurality of observed units that are input to an input unit of the node. Input weight coefficient adjustment means for adjusting an input weight coefficient that is a weight coefficient for each dimension with respect to input data that is time series data of multiple dimensions, and output of a plurality of dimensions output from the output unit of the node Output weight coefficient adjustment means for adjusting an output weight coefficient, which is a weight coefficient for each dimension of data, for each dimension.

  In the time series data, reliability is added for each dimension, and the input weight coefficient adjustment unit adjusts the input weight coefficient for each dimension based on the reliability for each dimension, The output weight coefficient adjusting means can adjust the output weight coefficient for each dimension based on the reliability for each dimension.

  The information processing apparatus further includes acquisition means for acquiring the input weight coefficient and the output weight coefficient input from the outside, and the input weight coefficient adjustment means is based on the input weight coefficient acquired by the acquisition means. The input weighting coefficient is adjusted for each dimension, and the output weighting coefficient adjusting means can adjust the output weighting coefficient for each dimension based on the output weighting coefficient acquired by the acquiring means.

  The information processing apparatus adds the output data output from the output unit immediately before and the newly observed time-series data for each dimension in a ratio corresponding to the input weighting coefficient, Input means for inputting data obtained as a result to the input unit as the input data can be further provided.

  The information processing apparatus adds the error for each dimension between the input data and the output data output from the output unit immediately before it as an output error by adding a ratio corresponding to the output weight coefficient, An internal state updating means for updating an initial value of the internal state quantity of the node based on the output error can be further provided.

  The information processing apparatus adds the error for each dimension between the input data and the output data output from the output unit immediately before, at a ratio corresponding to the output weight coefficient for each node. And determining means for determining a winner node which is a node corresponding to the dynamics most suitable for the input data based on the output error of each node, and corresponding to the distance from the winner node It is possible to provide weight updating means for updating the weight attached to the connection of each node to a certain degree.

  The information processing apparatus uses a base weight coefficient as a basis for an input weight coefficient and an output weight coefficient used when recognizing the input data or generating time-series data based on the reliability for each dimension. It is possible to further provide a base weight calculating means for calculating for each dimension.

  The information processing apparatus is obtained by updating the internal state quantity of the node for each node by inputting the input data and the previous input data to the input units of the plurality of nodes, respectively. The error for each dimension with the output data is added for each node at a ratio corresponding to the output weighting coefficient to obtain an output error for each node, and the most suitable for the input data based on the output error for each node. Determining means for determining a winner node which is a node holding the dynamics to be performed, and recognition means for outputting information representing the winner node as a recognition result of the time-series data.

  The input weight adjustment means sets the input weight coefficient for each dimension based on an error for each dimension between the input data and the output data output from the output unit immediately before the input data for each node. The output weight adjustment means can adjust the output weight coefficient for each dimension based on the error for each dimension for each node.

  The information processing apparatus further includes an acquisition unit configured to acquire a base weight coefficient that is a basis of the input weight coefficient and the output weight coefficient, calculated for each dimension when learning dynamics based on time-series data; The input weighting coefficient adjusting unit adjusts the input weighting factor for each dimension based on the base weighting factor, and the output weighting factor adjusting unit is configured to output the output for each dimension based on the base weighting factor. The weighting factor can be adjusted.

  The information processing apparatus includes: a determination unit that determines a generation node that is a node used for generation of time-series data of a plurality of dimensions among the plurality of nodes included in the network; A generation unit that generates time-series data of a plurality of dimensions while updating the internal state quantity of the generation node by inputting to the input unit can be provided.

  The input weight adjustment means adjusts the input weight coefficient for each dimension based on an error for each dimension between the input data and the output data output from the output unit immediately before the input data, and the output The weight adjusting means can adjust the output weighting factor for each dimension based on the error for each dimension.

  The information processing apparatus further includes an acquisition unit configured to acquire a base weight coefficient that is a basis of the input weight coefficient and the output weight coefficient, calculated for each dimension when learning dynamics based on time-series data; The input weighting coefficient adjusting unit adjusts the input weighting factor for each dimension based on the base weighting factor, and the output weighting factor adjusting unit is configured to output the output for each dimension based on the base weighting factor. The weighting factor can be adjusted.

  The generation means can generate the time-series data at a ratio corresponding to the output weighting coefficient for each dimension.

  An information processing method according to one aspect of the present invention is a method of maintaining time-series data of a plurality of dimensions that is input to an input unit of a node of a network that includes a plurality of nodes and that has dynamics in one node. An output weighting factor that is a weighting factor for each dimension of the output data of a plurality of dimensions, which is output from the output unit of the plurality of dimensions, by adjusting an input weighting factor that is a weighting factor for each dimension for the input data. For each dimension.

  A program according to one aspect of the present invention is time series data of a plurality of dimensions that are observed and input to an input unit of the node of a network configured by a plurality of the nodes, the dynamics being held in one node. An input weighting factor that is a weighting factor for each dimension with respect to input data is adjusted for each dimension, and an output weighting factor that is a weighting factor for each dimension with respect to the output data of a plurality of dimensions output from the output unit of the node, A computer is caused to execute a process including a step of adjusting for each dimension.

  In one aspect of the present invention, input that is time-series data of a plurality of dimensions that is observed and input to an input unit of the node of a network configured by a plurality of nodes, the dynamics being held in one node An input weighting factor that is a weighting factor for each dimension for data is adjusted for each dimension, and an output weighting factor that is a weighting factor for each dimension for output data of a plurality of dimensions output from the output unit of the node is a dimension. Adjusted every time.

  As described above, according to one aspect of the present invention, multidimensional time series data can be handled efficiently.

  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 which is described in the specification or the drawings but is not described here as an embodiment corresponding 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.

An information processing apparatus according to one aspect of the present invention includes:
Storage means (for example, the network storage unit 15 in FIG. 1) that stores the dynamics in one node and stores a network (for example, a dynamics storage network) constituted by a plurality of the nodes;
Input weight coefficient adjustment means for adjusting, for each dimension, an input weight coefficient that is a weight coefficient for each dimension with respect to input data that is time-series data of a plurality of dimensions that is input to the input unit of the node (for example, The input weight adjustment unit 82) of FIG.
Output weight coefficient adjustment means (for example, output weight adjustment unit 83 in FIG. 6) that adjusts the output weight coefficient, which is a weight coefficient for each dimension for the output data of a plurality of dimensions, output from the output unit of the node. ) And.

An information processing apparatus according to one aspect of the present invention includes:
In the time series data, reliability is added for each dimension,
The input weight coefficient adjusting means adjusts the input weight coefficient for each dimension based on the reliability for each dimension (for example, the process of step S5 in FIG. 7),
The output weight coefficient adjusting means adjusts the output weight coefficient for each dimension based on the reliability for each dimension (for example, the process of step S6 in FIG. 7).

An information processing apparatus according to one aspect of the present invention includes:
Acquisition means for acquiring the input weight coefficient and the output weight coefficient input from the outside (for example, the weight acquisition unit 101 in FIG. 8)
Further comprising
The input weight coefficient adjustment means adjusts the input weight coefficient for each dimension based on the input weight coefficient acquired by the acquisition means (for example, the process of step S25 in FIG. 9),
The output weight coefficient adjustment means adjusts the output weight coefficient for each dimension based on the output weight coefficient acquired by the acquisition means (for example, the process of step S26 in FIG. 9).

An information processing apparatus according to one aspect of the present invention includes:
The output data output from the output unit one time before and the newly observed time-series data are added for each dimension at a ratio corresponding to the input weighting factor, and the resulting data is Input means for inputting the input data to the input unit (for example, the adding unit 60 in FIG. 4)
Is further provided.

The division processing apparatus according to one aspect of the present invention includes:
The error for each dimension between the input data and the output data output from the output unit immediately before is added at a ratio corresponding to the output weighting coefficient to obtain an output error, and based on the output error. And internal state update means (for example, score calculation unit 84 in FIG. 6) for updating the initial value of the internal state quantity of the node.

An information processing apparatus according to one aspect of the present invention includes:
The error of each dimension between the input data and the output data output from the output unit immediately before is added at a ratio corresponding to the output weighting factor for each node, and an output error of each node is obtained. And determining means (for example, a winner node determining unit 85 in FIG. 6) for determining a winner node that is a node corresponding to the dynamics most suitable for the input data based on the output error of each node;
Weight update means (for example, parameter update unit 87 in FIG. 6) for updating the weight given to the connection of each node to the extent corresponding to the distance from the winner node.

An information processing apparatus according to one aspect of the present invention includes:
Based on the reliability for each dimension, a base weight coefficient used as a basis for an input weight coefficient and an output weight coefficient used when recognizing the input data or generating time-series data is calculated for each dimension. Base weight calculation means (for example, base weight determination unit 121 in FIG. 10)
Is further provided.

An information processing apparatus according to one aspect of the present invention includes:
For each dimension of the input data and output data obtained by updating the internal state quantity of each node for each node by inputting the previous input data to the input units of the plurality of nodes. Are added to each node at a ratio corresponding to the output weighting coefficient to obtain an output error of each node, and based on the output error of each node, a node that holds the dynamics most suitable for the input data A determination means (for example, the determination unit 216 in FIG. 13) for determining a winner node,
Recognizing means (for example, the output unit 217 in FIG. 13) for outputting information representing the winner node as a recognition result of the time-series data.

An information processing apparatus according to one aspect of the present invention includes:
The input weight adjustment means sets the input weight coefficient for each dimension based on an error for each dimension between the input data and the output data output from the output unit immediately before the input data for each node. (For example, the process of step S123 in FIG. 20),
The output weight adjustment means adjusts the output weight coefficient for each dimension based on the error for each dimension for each node (for example, the process of step S124 in FIG. 20).

An information processing apparatus according to one aspect of the present invention includes:
Acquisition means for acquiring a base weight coefficient that is a basis of the input weight coefficient and the output weight coefficient calculated for each dimension when learning dynamics based on time-series data (for example, the base weight acquisition unit in FIG. 22) 611)
Further comprising
The input weight coefficient adjustment means adjusts the input weight coefficient for each dimension based on the base weight coefficient (for example, the process of step S142 in FIG. 23),
The output weight coefficient adjusting means adjusts the output weight coefficient for each dimension based on the base weight coefficient (for example, the process of step S143 in FIG. 23).

An information processing apparatus according to one aspect of the present invention includes:
A determination unit (for example, a generation node determination unit 314 in FIG. 13) that determines a generation node that is a node used to generate time-series data of a plurality of dimensions among the plurality of nodes constituting the network;
Generation means for generating time series data of a plurality of dimensions while updating the internal state quantity of the generation node by inputting the input data to the input unit of the generation node (for example, the time series data of FIG. 13 Generating unit 316).

An information processing apparatus according to one aspect of the present invention includes:
The input weight adjustment means adjusts the input weight coefficient for each dimension based on an error for each dimension between the input data and the output data output from the output unit immediately before (for example, Step S134 in FIG. 21)
The output weight adjusting means adjusts the output weight coefficient for each dimension based on the error for each dimension (for example, the process of step S135 in FIG. 21).

An information processing apparatus according to one aspect of the present invention includes:
Acquisition means for acquiring a base weight coefficient that is a basis of the input weight coefficient and the output weight coefficient calculated for each dimension when learning dynamics based on time-series data (for example, the base weight acquisition unit in FIG. 22) 621)
Further comprising
The input weight coefficient adjustment means adjusts the input weight coefficient for each dimension based on the base weight coefficient (for example, the process of step S152 in FIG. 24),
The output weight coefficient adjusting means adjusts the output weight coefficient for each dimension based on the base weight coefficient (for example, the process of step S153 in FIG. 24).

An information processing apparatus according to one aspect of the present invention includes:
The generation means generates the time-series data at a ratio corresponding to the output weight coefficient for each dimension (for example, the process of step S96 in FIG. 15).

A learning method or program according to one aspect of the present invention includes:
A weighting factor for each dimension with respect to input data that is time-series data of a plurality of dimensions that is input to the input unit of the node of the network that holds the dynamics in one node and is configured by a plurality of the nodes. A certain input weight coefficient is adjusted for each dimension (for example, step S5 in FIG. 7),
An output weighting factor, which is a weighting factor for each dimension for output data of a plurality of dimensions, output from the output unit of the node is adjusted for each dimension (for example, step S6 in FIG. 7).
Includes steps.

  Hereinafter, specific embodiments to which the present invention is applied will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing a configuration example of an embodiment of an information processing apparatus 1 to which the present invention is applied.

  The information processing apparatus 1 in FIG. 1 stores a network storage unit 15 that stores a dynamics storage network having, as one node, a dynamic system approximation model having an internal state quantity among dynamic system approximation models that are models approximating a dynamic system. And a learning unit 14 that updates the parameters of the dynamics storage network in a self-organizing manner.

  Each node of the dynamics storage network holds the dynamic characteristics of time-series data as dynamics. The dynamics defined by the parameters updated by the learning unit 14 and held in the nodes of the dynamics storage network are used for recognition and generation of time series data.

  The information processing apparatus 1 in FIG. 1 is used for recognition and generation of a control signal such as a robot. The information processing apparatus 1 is used for learning, recognition, or generation of input / output signals of sensors and motors, particularly in autonomous systems and autonomous robots.

  In the information processing apparatus 1 in FIG. 1, both a signal input to the information processing apparatus 1 and a signal output from the information processing apparatus 1 are input to the signal input unit 12 as an observation signal 11 that is an observed signal. The The observation signal 11 is, for example, a sound or image signal, a signal representing the brightness of an LED (Light Emitting Diode), a signal representing a rotation angle or a rotation angular velocity of a motor, or the like.

  The signal input unit 12 outputs an electrical signal corresponding to the input observation signal 11 to the feature extraction unit 13. Specifically, the signal input unit 12 corresponds to, for example, a microphone when the observation signal 11 is an audio signal, and corresponds to, for example, a camera when the observation signal 11 is an image signal. When the observation signal 11 is a signal representing the rotation angle or rotation speed of the motor, the signal input unit 12 corresponds to, for example, a measuring device that measures the rotation angle or rotation speed of the motor.

  Hereinafter, not only a signal input to the signal input unit 12 but also a signal output from the signal input unit 12 is referred to as an observation signal 11. The observation signal 11 may be a stationary signal or a non-stationary signal that changes with time.

  Further, hereinafter, a sensor motor signal mainly in the robot system will be described as an example of the observation signal 11. The sensor motor signal is a vector having, for example, a signal output from the sensor or a control signal (hereinafter referred to as a motor signal) for controlling the motor input to the motor as a component. Of course, the observation signal 11 is not limited to the sensor motor signal.

  In addition, the signal input unit 12 includes a section detection device and the like, and outputs the sensor motor signal divided by a predetermined section. The signal input unit 12 only needs to output a sensor motor signal divided into an appropriate length, and the way of dividing is not particularly limited. Therefore, the sensor motor signal divided into an appropriate length by the best method may be output from the signal input unit 12 as the observation signal 11 according to the input sensor motor signal.

  The feature extraction unit 13 extracts feature amounts in time series from the observation signal 11 output from the signal input unit 12. For example, the feature extraction unit 13 performs processing such as frequency analysis on a sound signal that is one of the sensor signals at regular time intervals, and extracts feature quantities such as a mel cepstrum in time series. Here, the mel cepstrum is a feature amount widely used for voice recognition and the like.

  The feature extraction unit 13 generates time series data of feature quantities obtained by extracting feature quantities from the observation signal 11 in time series (hereinafter simply referred to as time series data), a learning unit 14, a recognition unit 16, and a generation unit. Supplied to the unit 19.

  The learning unit 14 uses the time series data supplied from the feature extraction unit 13 to learn the temporal change feature of the time series data as dynamics at a predetermined degree. Specifically, the learning unit 14 updates the parameters of the dynamics storage network that holds the dynamics to a predetermined degree.

  Details of learning by the learning unit 14 will be described later. Basically, when time-series data without a label is repeatedly given, the learning unit 14 self-organizes characteristic dynamics in the time-series data. Execute unsupervised learning that can be acquired in a chemical manner. As a result, typical dynamics are efficiently held in the dynamics storage network stored in the network storage unit 15. The recognized dynamics can be used at any time by the recognition unit 16 and the generation unit 19 as necessary.

  Here, the dynamics represents a dynamic system that changes with time, and can be expressed by a specific function, for example. In the dynamics storage network, the temporal change characteristics of time-series data are retained as dynamics.

  The recognizing unit 16 compares the time series data supplied from the feature extracting unit 13 with the dynamics held in the dynamics storage network obtained as a result of learning so far, and determines the most similar dynamics.

  Specifically, the recognizing unit 16 inputs the time series data supplied from the feature extracting unit 13 to the dynamics storage network, and obtains output data output in response to the input. Based on the output data and the time series data supplied from the feature extraction unit 13, the recognition unit 16 determines the most similar dynamics to the time series data. The recognition unit 16 outputs the determination result as a recognition result 17.

  On the other hand, from the dynamics stored in the dynamics storage network, time series data can be generated as needed. The generation unit 19 performs a generation process for generating time-series data.

  Specifically, the generation unit 19 acquires a control signal 18 that specifies from which dynamics the time-series data is generated. Based on the control signal 18 and the time series data supplied from the feature extraction unit 13, the generation unit 19 inputs the time series data to a node that holds the designated dynamics. And the production | generation part 19 produces | generates time series data based on the output data output with respect to the input. Then, the generation unit 19 outputs the time series data as a generation result 20.

  The internal state storage unit 21 holds an internal state amount of each node of the dynamics storage network. The internal state quantity stored in the internal state storage unit 21 is updated by, for example, the recognition unit 16 and is used by the generation unit 19 for generation processing.

[About learning]
Next, learning performed by the information processing apparatus 1 in FIG. 1 will be described with reference to FIGS.

2 and 3, it is assumed that an RNN having a regression loop from the output layer to the input layer of the three-layer neural network (NN) is used as the dynamic system approximation model. Using this RNN, the state vector X _t at time t in the time series data is input, and learning to predict and output the state vector X _{t + 1} at time t + 1 is input to that input, that is, prediction learning ( prediction learning), it is possible to learn the time evolution law of time series data.

  In general, a BPTT (Back-Propagation Through Time) method is used as a parameter estimation method for a dynamical approximate model having an internal state quantity such as RNN. The BPTT method is a learning method based on the steepest descent method. In the BPTT method, learning is performed by a gradient method based on iterative calculation.

  For the BPTT method, see, for example, DE Rumelhart, GE Hinton & RE Williams, 1986 “Learning internal representations by error propagation”, In DE Rumelhart & J. McClelland, “Parallel distributed processing”, pp. 318-364, Cambridge, MA: MIT Press, RJ Williams and D. Zipser, “A learning algorithm for continuously running fully recurrent neural networks”, Neural Computation, 1: 270-280, 1989, etc.

  FIG. 2 shows a dynamic system approximation model 31 when learning is performed using two-dimensional time series data, and FIG. 3 shows a dynamic system approximation model 41 when learning is performed using six-dimensional time series data. Represents. Here, the dimension represents the type of time-series data distinguished in learning, recognition, or generation, and the number of dimensions corresponds to the number of RNN units.

  2 and 3, a sensor signal representing sound (hereinafter referred to as an audio signal), a sensor signal representing an image (hereinafter referred to as an image signal), a sensor signal representing a tactile sense (hereinafter referred to as a tactile signal), A motor signal for controlling the motor to be operated (hereinafter referred to as a “hand signal”), a motor signal for controlling the motor for operating both feet (hereinafter referred to as a “both foot signal”), and a motor signal for controlling the motor to operate the neck (hereinafter referred to as the neck signal) It is assumed that 6-dimensional time series data respectively generated from the above are obtained as time series data.

In FIG. 2, among the obtained 6-dimensional time series data, the two-dimensional time series data of the audio signal and the both foot signals are input to two units of the input layer of the dynamic system approximation model 31, respectively. The dynamical approximate model 31 receives the state vector X _t at time t in the two-dimensional time series data, and predicts the state vector X _{t + 1} at time t + 1 in the two-dimensional time series data with respect to the input. And learn to output. Therefore, the number of units of the input layer and the output layer of the dynamic system approximation model 31 in FIG. 2 is two each.

In FIG. 3, all of the obtained 6-dimensional time-series data is input to each of the six units in the input layer of the dynamic system approximation model 32. The dynamical approximate model 32 receives the state vector X _t at time t in the 6-dimensional time series data as input, and predicts the state vector X _{t + 1} at time t + 1 in the 6-dimensional time series data with respect to the input. And learn to output. Therefore, the number of units of the input layer and the output layer of the dynamic system approximation model 32 in FIG.

  As described above, in the dynamic system approximation model 32 of FIG. 3 that is learned using 6-dimensional time series data, the dynamic system approximation model 31 of FIG. 3 that is learned using two-dimensional time series data is input. The number of units in the layer and output layer is greatly increased and the scale is increased.

  Here, it is generally known that it is difficult to converge parameters when the scale of a neural network increases. In addition, learning is performed using the average value of the prediction error calculated from the predicted value of the time series data of each dimension output from each unit of the output layer, but when the number of dimensions of the time series data increases, The proportion of time-series data of one dimension that affects the average value of the prediction error becomes very small. Therefore, it is not easy to perform learning using the average value of the prediction errors.

  As described above, when the number of dimensions of time series data increases, learning based on the time series data becomes difficult.

  For example, when learning the behavior of walking in a direction in which sound can be heard, the time-series data indicating the direction and size of the sound obtained from the sound signal and the time-series data corresponding to the both foot signals for performing the walking motion It is necessary to learn the dynamics formed based on the action. At this time, as shown in FIG. 2, when learning is performed using two-dimensional time-series data corresponding to the audio signal and the both foot signals among the six-dimensional time-series data, the dynamics can be easily learned. However, as shown in FIG. 3, when learning is performed using all of the six-dimensional time-series data, the learning becomes difficult.

  Also, when learning the action of bringing the hand closer when the red ball is seen in front of the eyes, time series data representing the position coordinates of the red ball obtained from the image signal and the time series corresponding to the two-hand signal for moving the hand It is necessary to learn the dynamics formed based on the interaction of data. At this time, as shown in FIG. 3, when learning is performed using all of the 6-dimensional time-series data, 2-dimensional time-series data corresponding to the image signal and the two-handed signal is selected from the 6-dimensional time-series data. The learning becomes difficult as compared to the case where learning is performed using the learning method.

  Therefore, it is desirable to learn dynamics using not only time-series data of all dimensions obtained but only time-series data of a dimension to be noted in learning.

  However, when supervised learning is performed, time series data to be focused on can be determined in advance for each node, but when unsupervised learning is performed, it cannot be determined in advance. .

  Therefore, the learning unit 14 based on the time series data supplied from the feature extraction unit 13, weight factors for each dimension (hereinafter referred to as input weights) for the time series data input to the dynamic system approximate model, and the dynamic system Weights for each dimension (hereinafter referred to as output weights) for time-series data output from the approximate model are determined so that input weights and output weights of dimensions to be noted in learning are increased.

  As a result, the input / output time-series data is weighted, and the dynamics can be learned based on the time-series data to be noted among all the time-series data. As a result, it is possible to efficiently handle input multidimensional time-series data in learning.

  FIG. 4 is a diagram showing details of one node 41 of the dynamics storage network stored in the network storage unit 15.

  The node 41 includes a dynamic system approximation model 51 having an internal state quantity, an input weight storage unit 52, an output weight storage unit 53, and a learning degree storage unit 54.

The dynamic system approximation model 51 is an RNN, and includes an adding unit 60, an input layer 61, a hidden layer 62, and an output layer 63. The addition unit 60 is a new observation supplied from the learning unit 14 at a ratio corresponding to the input weight of each dimension (hereinafter referred to as a learning input weight) stored in the input weight storage unit 52 and used for learning. State vector X _{t at} time t in the time series data corresponding to the observed signal and the output data output immediately before from the output layer 63, that is, the predicted value of the state vector at time t-1 in the time series data. Certain output data is added for each dimension, and the resulting data for each dimension is input to each unit of the input layer 61 for each dimension as learning data at time t.

  The learning data at time t input to each unit of the input layer 61 is output from the output layer 63 via the hidden layer 62. That is, based on learning data at time t input to each unit of the input layer 61 and weights (hereinafter referred to as “combining weights”) given to the connections between the units of the input layer 61 and the units of the hidden layer 62. , A predetermined calculation is performed, and based on the data obtained as a result and the coupling weight of each unit of the hidden layer 62 and each unit of the output layer 63, the predetermined calculation is performed and the resulting data is The output data is output from the output layer 63.

  The learning unit 14 uses the output data output from the output layer 63 to update the coupling weight of the dynamic system approximation model 51 as a parameter with the degree represented by the degree information stored in the learning degree storage unit 54. That is, the dynamical system approximation model 51 learns the time series pattern of the learning data input to each unit of the input layer 61 as dynamics with the degree represented by the degree information stored in the learning degree storage unit 54.

  The learning performed by the learning unit 14 is online learning. In other words, every time the observation signal 11 is input, the learning unit 14 updates the parameters of the dynamic system approximation model 51 little by little based on the learning data corresponding to the observation signal 11.

The input weight storage unit 52 is a learning input weight supplied from the learning unit 14, an input weight at the time of recognition supplied from the recognition unit 16 (hereinafter referred to as a recognition input weight), and a generation time supplied from the generation unit 19. An input weight (hereinafter referred to as a generated input weight) is stored. The adder 60 calculates the state vector X _{t at the} time t in the time series data supplied from the learning unit 14 and the output data output immediately before from the output layer 63 at a ratio corresponding to the learning input weight. Add together. That is, the adder 60 weights the time series data supplied from the learning unit 14 based on the learning input weight. The weighting by the recognition input weight and the generation input weight will be described later with reference to FIG.

  The output weight storage unit 53 is an output weight (hereinafter referred to as learning output weight) used during learning supplied from the learning unit 14, an output weight during recognition (hereinafter referred to as recognition output weight) supplied from the recognition unit 16, And the output weight at the time of generation (hereinafter referred to as a generation output weight) supplied from the generation unit 17 is stored. Based on the learning output weights, the learning unit 14 weights the time-series data of each dimension output from each unit of the output layer of the dynamical approximate model 51. Details of the weighting based on the learning input weight and the learning output weight will be described later with reference to FIG. The weighting by the recognition output weight and the generation output weight will be described later with reference to FIG.

  The learning degree storage unit 54 stores information indicating the degree of learning of the parameters of the dynamic system approximation model 51 (hereinafter referred to as degree information). Here, as the degree information, for example, the number of repetitions of repeated calculation in the BPTT method is used.

  The learning unit 14 adjusts the degree of learning in accordance with this degree information, and thereby adjusts the degree to which the parameters of the dynamic system approximation model 51 are affected by the learning data.

  Next, weighting based on the learning input weight and the learning output weight will be described with reference to FIG.

In the example of FIG. 5, the state vector x1 _t and x2 _{t at} time t in the two-dimensional time-series data corresponding to the sensor signal and the two-dimensional time corresponding to the motor signal are added to the adding unit 60 of the dynamical approximate model 51. State vectors x3 _t and x4 _{t at} time t in the series data are input from the learning unit 14. That is, state vectors x1 _{t to} x4 _{t at} time t in the four-dimensional time series data are input from the learning unit 14 to the adding unit 60.

The adding unit 60 reads the learning input weight stored in the input weight storage unit 52, and based on the learning input weight, from the state vectors x1 _{t to} x4 _t from the learning unit 14 and each unit of the output layer 63. output to the previous (one time before), the output data XO1 _t to x O4 _t is the predicted value of the state vector at time t, the sum according to the following equation (1), the data obtained as a result of It inputs into the input layer 61 as learning data.

In Expression (1), xk _t , Xik _t , Xok _t , and αs _k (0 ≦ αs _k ≦ 1) are k dimensions (k = 1, 2, 3, and 4 in the example of FIG. 5), respectively. It represents the eye state vector, input data input to the input layer, output data output from the output layer, and learning input weight.

According to equation (1), when the input weights .alpha.s _k is 1, only the state vector xk _t from the learning unit 14 is input to the input layer 61, when the input weights .alpha.s _k is 0, the output data Xok _t Only is input to the input layer 61. Therefore, by adjusting the value of the input weight αs _k , it is possible to adjust the dimension of the time series data that should be noted during learning.

For example, when the input weights αs ₁ and αs ₃ are 1 and the input weights αs ₂ and αs ₄ are 0, out of the state vectors x1 _{t to} x4 _{t in} the four-dimensional time series data input to the adder 60 Only the state vector x1 _t in the one-dimensional time series data corresponding to the sensor signal and the state vector x3 _t in the one-dimensional time series data corresponding to the motor signal are input to the input layer 61, and as a result, the state vector x1 Learning is performed focusing only on _t and x3 _t .

In this case, the unit of the input layer 61 corresponding to the state vectors x2 _t and x4 _t performs the same function as the context unit combined in the regression loop of the input layer 61.

Further, when the input weights α _{1 to} α ₄ are all 0.5, attention is paid to the state vectors x1 _{t to} x4 _t in all the four-dimensional time series data input to the adding unit 60 with the same weight. Learning is performed. In this case, the adding unit 60 adds the state vectors x1 _{t to} x4 _t and the previous output data Xo1 _{t to} Xo4 _t for each dimension at the same ratio according to the equation (1), respectively. Is input to each unit of the input layer 61.

As described above, since weighting is performed on the time series data of each dimension input from the learning unit 14 based on the learning input weight αs _k , among the time series data of each dimension input, at the time of learning It is possible to efficiently learn dynamics by focusing only on time-series data that should be noted. As a result, the learning process load can be reduced.

Note that the output data output immediately before from the context unit of the output layer 63 is input to the context unit of the input layer 61 as the context C _{t at} time t.

On the other hand, the learning unit 14 learns to predict and output the state vector at time t + 1 with respect to the state vectors x1 _{t to} x4 _t at time t. Therefore, the following evaluation values are used as learning values. The predicted value of the state vector at time t + 1 and the weighted mean square error E of the true value obtained according to the equation (2) are used.

In Equation (2), βs _k (0 ≦ βs _k ≦ 1) represents an output weight of the kth dimension. Further, n represents the number of dimensions, and is 4 in the example of FIG.

According to the equation (2), the state vector xk _{t + 1} at time t + 1 in the time series data supplied from the learning unit 14 is set as the true value at time t + 1, and the state vector xk _{t + 1} and the state at time t + 1. mean square error between the output data Xok t + ₁ obtained by inputting the state vector xk _t of the predicted value is output data Xok t + _1, i.e. one time before the time t of the vector is determined for each dimension, The mean square error is weighted by the output weight βs _k and added. Therefore, by adjusting the value of the output weight βs _k , the dimension of output data to be noted as the evaluation value at the time of learning can be adjusted.

For example, output weights .beta.s ₁ and .beta.s ₃ is 1, and output weights .beta.s ₂ and .beta.s ₄ is 0, among the output data XO1 t + ₁ to x O4 t + ₁ output from each unit of the output layer 63 The mean square error E obtained by paying attention only to the state vector Xo1 _t in the one-dimensional time-series data corresponding to the sensor signal and the state vector Xo3 _t in the one-dimensional time-series data corresponding to the motor signal is used as the evaluation value. Used.

  The learning unit 14 estimates a parameter that minimizes the mean square error E obtained in this way, and updates the parameter.

As described above, the learning unit 14 performs weighting on the mean square error of each dimension based on the learning output weight βs _k , and therefore, by using the obtained mean square error E as an evaluation value, output is performed. Learning can be performed using an evaluation value focused on only time-series data to be focused on as an evaluation value at the time of learning among the time-series data of each dimension. As a result, dynamics can be efficiently learned based on multidimensional time-series data.

  FIG. 6 is a block diagram illustrating a detailed configuration example of the learning unit 14 of FIG.

  The learning unit 14 includes a reliability extraction unit 81, an input weight adjustment unit 82, an output weight adjustment unit 83, a score calculation unit 84, a winner node determination unit 85, a learning weight determination unit 86, and a parameter update unit 87. The learning unit 14 receives time-series data from the feature extraction unit 13 in FIG. 1, and the time-series data is supplied to the reliability extraction unit 81, the score calculation unit 84, and the parameter update unit 87.

It is assumed that reliability ρ _k (0 ≦ ρ _k ≦ 1) of each dimension is added to the time series data input from the feature extraction unit 13 as additional information. Here, the reliability ρ _k is the observation probability of the observation signal 11 corresponding to the k-dimensional time-series data. That is, when the observation signal 11 corresponding to the k-dimensional time-series data is observed, the reliability ρ _k is high, and when the observation signal 11 is not observed, the reliability ρ _k is low. Here, the reliability [rho _k is high, the value of the reliability [rho _k is that large, and the reliability [rho _k is low, is that the value of the confidence [rho _k is small.

For example, when the power of the sound signal that is the observation signal 11 is large, that is, when the sound is heard, the feature extraction unit 13 sets the reliability ρ _k to be added to the time-series data of the dimension corresponding to the sound signal to 1, When the power of the signal is small, that is, when the voice cannot be heard, the reliability ρ _{k is set} to zero.

Further, when the voltage of the motor signal as the observation signal 11 is high, that is, when the motor is driven, the reliability ρ _k added to the time-series data of the dimension corresponding to the motor signal is set to 1, and the voltage of the motor signal Is low, that is, when the motor is not driven, the reliability ρ _{k is set} to zero.

Note that the method for obtaining the reliability ρ _k is not limited to the above-described method. For example, the magnitude of the noise signal superimposed on the observation signal 11 corresponding to the time-series data of each dimension is measured or predicted, and the noise signal is calculated. A method may be used in which the reliability ρ added to the corresponding time-series data is increased when the noise signal is large, and the reliability ρ is decreased when the noise signal is small.

The method for obtaining the reliability ρ _k can be determined according to what observation signal 11 is input, what feature quantity the feature extraction unit 13 extracts.

Confidence extractor 81, thus to the respective dimensions of the time-series data reliability [rho _k is added which sought to extract the reliability [rho _k for each dimension, the output-weight adjuster to the input-weight adjuster 82 To the unit 83.

Based on the reliability ρ _k of each dimension from the reliability extraction unit 81, the input weight adjustment unit 82 adjusts the learning input weight αs _k stored in the input weight storage unit 52 of FIG. 4 for each dimension. Specifically, the input-weight adjuster 82 is based on the reliability [rho _k, the reliability [rho _k higher learning input weights .alpha.s _k dimension increases, the reliability [rho _k low dimensional learning input weights .alpha.s _k as smaller, determines the learning input weights .alpha.s _k, stores and supplies the learning input weights .alpha.s _k to the input-weight storage unit 52.

Similar to the input weight adjustment unit 82, the output weight adjustment unit 83 determines the learning output weight βs _k stored in the output weight storage unit 53 based on the reliability ρ _k of each dimension from the reliability extraction unit 81. Adjust for each dimension.

  The score calculation unit 84 calculates the score of the dynamic system approximation model 51 of each node 41 (FIG. 4) included in the dynamics storage network stored in the network storage unit 15 with respect to the time series data from the feature extraction unit 13. This is done while updating the internal state quantity.

  Specifically, the score calculation unit 84 inputs time series data to the addition unit 60 of the dynamic system approximation model 51 of each node 41, and outputs output data corresponding to the input to the output of the dynamic system approximation model 51. Obtained from layer 63. Then, the score calculation unit 84 calculates the mean square error E as an evaluation value for each node 41 according to Expression (2). The score calculation unit 84 assigns the mean square error E to each node 41 as a score.

The score calculation unit 84 sets the score among the values obtained by updating the predetermined value according to the following formula (3) with respect to the predetermined value for each node 41 as the reference. The value to be decreased is determined as the initial value C ₀ of the context of the dynamical approximate model 51 as the internal state quantity, and score calculation is performed while updating the context from the initial value.

In Equation (3), C _{0, m} (s) is the initial value of the context input to the m-th context unit, which is updated for the s (s = 0, 1,...) Times. ΔC _{0, m} (s + 1) is the update amount of the s + 1-th update of the initial value of the context input to the m-th context unit, and is obtained by the following equation (4).

In Expression (4), η represents a learning coefficient, and γ represents an inertia coefficient. Further, δ _{C0, m} represents an error amount of the initial value C _{0, m} (S) of the context obtained by the BPTT method using the mean square error E.

In the BPTT method, the error amount δ _c (t + 1) of the context C _{t + 1} at the time t + 1 input to the context unit of the input layer 61 is used as the context C _{t of the} time t output from the context unit of the output layer 61. When backpropagating to the error amount δ _c (t), the time constant of the context is adjusted by dividing by an arbitrary positive coefficient m as shown in the following equation (5).

  By adopting equation (5) in the BPTT method, the influence degree of one time step ahead of the context can be adjusted.

The predetermined value used for determining the initial value of the context, that is, the initial value C _{0, m} (0) is, for example, a random value or the final value of the context obtained when learning the previous dynamical system approximation model 51. An update value (hereinafter referred to as a previous update value) or the like can be adopted.

For example, when it is known that there is no relationship between the learning data used in the current learning and the learning data used in the previous learning, the initial value C _{0, m} (0) is Random values can be adopted.

In addition, for example, when it is known that the learning data used in the current learning and the learning data used in the previous learning have some relationship such as continuous time-series data, As the value C _{0, m} (0), the previous updated value can be adopted. When the previous update value is adopted as the initial value C _{0, m} (0), the previous update value can be determined as the initial value of the context without updating.

  The score calculation unit 84 supplies the score given to each node 41 to the winner node determination unit 85 as a result of the score calculation. The winner node determination unit 85 compares the scores of the respective nodes 41 from the score calculation unit 84, and determines the node 41 having the smallest score value as the winner node that is the most suitable node for the learning data. The winner node determination unit 85 supplies information specifying the winner node to the learning weight determination unit 86.

  The learning weight determination unit 86 refers to the dynamics storage network stored in the network storage unit 15 and calculates the distance d from the winner node specified by the information from the winner node determination unit 85 for each node 41. Then, the learning weight determination unit 86 determines the learning weight for each node 41 so that the value for the winner node becomes the largest according to the distance d, and the value becomes smaller as the distance d from the winner node becomes larger. To decide.

  Further, the learning weight determination unit 86 generates the degree information of each node 41 based on the learning weight. Specifically, the learning weight determination unit 86 increases the degree of learning of the node 41 having a large learning weight and decreases the degree of learning of the node 41 having a small learning weight according to the learning weight. The degree information of each node 41 is generated. Then, the learning weight determination unit 86 supplies the generated degree information to the learning degree storage unit 54 of each node 41 and stores it.

  As a result, the learning level is adjusted such that the strongest learning is performed at the winner node and the learning level decreases as the distance d from the winner node increases.

  For each node 41, the parameter update unit 87 reads the degree information stored in the learning degree storage unit 54, and dynamics the time series pattern of the learning data supplied from the feature extraction unit 13 with the degree represented by the degree information. To learn as.

  Specifically, the parameter updating unit 87 inputs time series data for each node 41 to the adding unit 60 of each node 41, and outputs output data corresponding to the input to the output layer 63 of the dynamic system approximation model 51. Get from. Then, the score calculation unit 84 calculates the mean square error E as an evaluation value for each node 41 according to Expression (2).

  Based on the mean square error E, the time series data, and the degree information, the parameter update unit 87 performs repetitive calculation in the BPTT method and updates the parameters of the dynamic system approximation model 51 according to the following equation (6). .

In Equation (6), w _{i, j} (s) is the connection weight given to the connection between units i and j, which is updated by the s (s = 0, 1,...) Iteration. Represents. Here, the update according to the equation (6) is performed the number of repetitions which is degree information. That is, s is a number from 0 to a value obtained by subtracting 1 from the number of repetitions. Δw _{i, j} (s + 1) is the update amount of the s + 1th update of the connection weight w _{i, j} and is obtained by the following equation (7).

In Expression (7), δ _wij represents an error amount of the connection weight w _{i, j} (s) obtained by the BPTT method using the mean square error E.

As described above, the parameter update unit 87, since the back-propagated error by BPTT method using calculated based on the learning output weights .beta.s _k mean square error E, the learning output weights .beta.s _k, backpropagation error In this case, the dimension of the output data to be noticed is adjusted. That is, the dimension of the output data to be noted when updating the parameters is adjusted by the learning output weight βs _k .

  Further, the parameter update unit 87 performs learning based on the learning weight adjusted so that the learning is performed most strongly at the winner node and the degree of learning is reduced as the distance d from the winner node increases, that is, the node 41. In this way, a large number of dynamics can be learned in a self-organizing manner.

  Next, a learning process in which the information processing apparatus 1 in FIG. 1 learns the dynamics storage network will be described with reference to FIG. This learning process is started, for example, when the information processing apparatus 1 in FIG. 1 is turned on.

  First, in step S1, the parameter update unit 87 (FIG. 5) of the learning unit 14 initializes the parameters of the dynamics storage network stored in the network storage unit 15. Specifically, an appropriate value is assigned as an initial value to the parameter of the dynamic system approximation model 51 of each node 41 of the dynamics storage network.

After the processing in step S1, the process proceeds to step S2, and the signal input unit 12 in FIG. 1 acquires the observation signal 11 and supplies it to the feature extraction unit 13, and the process proceeds to step S3. In step S3, the feature extraction unit 13 extracts the feature amount of the observation signal 11 in time series for each dimension, and uses the time series data of each dimension obtained as a result as learning data, as a reliability extraction unit of the learning unit 14. 81, the score calculation unit 84, and the parameter update unit 87. At this time, the feature extraction unit 13 adds the reliability ρ _k to the time-series data for each dimension.

After the processing in step S3, the process proceeds to step S4, where the reliability extraction unit 81 extracts the reliability ρ _k added to the time series data of each dimension from the feature extraction unit 13, and outputs the input weight adjustment unit 82 and the output. The data is supplied to the weight adjustment unit 83 and the process proceeds to step S5.

In step S5, the input weight adjustment unit 82 adjusts the learning input weight αs _k stored in the input weight storage unit 52 of each node 41 based on the reliability ρ _k from the reliability extraction unit 81, and the step Proceed to S6.

In step S _< b> 6, the output weight adjustment unit 83, like the input weight adjustment unit 82, learns stored in the output weight storage unit 53 of each node 41 based on the reliability ρ _k from the reliability extraction unit 81. The output weight βs _k is adjusted, and the process proceeds to step S7.

  In step S <b> 7, the score calculation unit 84 internally calculates the score of the dynamical approximate model 51 of each node 41 included in the dynamics storage network stored in the network storage unit 15 for the learning data from the feature extraction unit 13. This is done while updating the state quantity. The score calculation unit 84 supplies the score given to each node 41 as a result of the score calculation to the winner node determination unit 85.

  After the processing of step S7, the process proceeds to step S8, and the winner node determination unit 85 compares the scores of the respective nodes 41 supplied from the winner node determination unit 85 so that the node having the smallest score value is determined as the winner node. Information that determines and identifies the winner node is supplied to the learning weight determination unit 86.

  After the processing of step S8, the process proceeds to step S9, where the learning weight determination unit 86 refers to the dynamics storage network stored in the network storage unit 15 and performs learning for each node 41 based on the distance d from the winner node. Determine the weight.

  After the processing of step S9, the process proceeds to step S10, where the learning weight determination unit 86 generates the degree information of each node 41 based on the learning weight of each node 41, and the degree information is obtained as the learning degree of each node 41. Each is supplied to and stored in the storage unit 54.

  After the process of step S10, the process proceeds to step S11, and the parameter update unit 87 performs the iterative calculation in the BPTT method for each node 41 based on the learning data and the degree information from the feature extraction unit 13, and the dynamic system approximation model 51 parameters are updated. Thereafter, the process returns to step S2, and the subsequent processes are repeated. As described above, dynamics learning is performed at the degree indicated by the degree information.

  FIG. 8 is a block diagram illustrating another detailed configuration example of the learning unit in FIG. 1.

  The learning unit 100 includes a weight acquisition unit 101, an input weight extraction unit 102, and an output weight adjustment unit 103 instead of the reliability extraction unit 81, the input weight adjustment unit 82, and the output weight adjustment unit 83.

That is, in FIG. 8, the reliability ρ _{k of} each dimension is not added to the time-series data input from the feature extraction unit 13, and the learning input weight αs _k and the learning output weight βs _k are directly input from the outside. Is done.

The weight acquisition unit 101 acquires a learning input weight αs _k and a learning output weight βs _k input from the outside. Then, the weight acquisition unit 101 supplies the learning input weight αs _k to the input weight adjustment unit 102 and supplies the learning output weight βs _k to the output weight adjustment unit 103.

Input weight adjuster 102, the learning input weights .alpha.s _k from the weight obtaining unit 101, by the input-weight storage unit 52, adjusts the learning input weights .alpha.s _k.

Output weight adjuster 103, similarly to the input-weight adjuster 102, the learning output weights .beta.s _k from the weight obtaining unit 101, by storing the output-weight storage unit 53, adjusts the learning output weights .beta.s _k.

  Next, a learning process in which the information processing apparatus having the learning unit 100 in FIG. 8 learns the dynamics storage network will be described with reference to FIG. This learning process is started, for example, when the information processing apparatus is turned on.

  First, in step S21, the parameter update unit 87 (FIG. 8) of the learning unit 100 initializes the parameters of the dynamics storage network stored in the network storage unit 15. Specifically, an appropriate value is assigned as an initial value to the parameter of the dynamic system approximation model 51 of each node 41 of the dynamics storage network.

  After the process of step S21, the process proceeds to step S22, and the signal input unit 12 acquires the observation signal 11 and supplies it to the feature extraction unit 13, and the process proceeds to step S23. In step S23, the feature extraction unit 13 extracts feature quantities of the observed signal 11 in time series for each dimension, and uses the time series data of each dimension obtained as a result as learning data, as a score calculation unit 84 of the learning unit 14. And supplied to the parameter update unit 87.

After the process of step S23, the process proceeds to step S24, and the weight acquisition unit 101 acquires the learning input weight αs _k and the learning output weight βs _k input from the outside. Then, the weight acquisition unit 101 supplies the learning input weight αs _k to the input weight adjustment unit 102 and supplies the learning output weight βs _k to the output weight adjustment unit 103.

After the processing in step S24, the process proceeds to step S25, the input-weight adjuster 102, the learning input weights .alpha.s _k from the weight obtaining unit 101, by the input-weight storage unit 52, adjusts the learning input weights .alpha.s _k Then, the process proceeds to step S26.

In step S26, the output weight adjustment unit 103 stores the learning output weight βs _k from the weight acquisition unit 101 in the output weight storage unit 53 in the same manner as the input weight adjustment unit 102, thereby obtaining the learning output weight βs _k . Adjust and go to step S27.

  The processing in steps S27 to S31 is the same as that in steps S7 to S11 in FIG.

  FIG. 10 is a block diagram showing still another detailed configuration example of the learning unit in FIG.

  The learning unit 120 is configured by further adding a base weight determining unit 121 to each unit of the learning unit 14 in FIG. 6.

The base weight determination unit 121 is supplied with the reliability ρ _{k of} each dimension from the reliability extraction unit 81. Based on the reliability ρ _k , the base weight determination unit 121 determines a base weight ω _k (0 ≦ ω _k ≦ 1) that serves as a basis for input weights and output weights used in recognition generation described later. Specifically, the base weight determination unit 121 sets the base weight ω _k to 1 before starting learning, and based on the reliability ρ _k supplied from the reliability extraction unit 81, the base weight ω _k is based on the equation (8). Update weight ω _k . Accordingly, the base weight ω _k is adjusted with learning.

In Equation (8), ω ′ _k represents the updated base weight, and Δ (0 <Δ <1) represents the step width of the change in the base weight ω _k due to the update, for example, 0.05.

According to equation (8), continues steady value is input as the reliability [rho _k, the update of the Motoomomi omega _k is sufficiently performed, Motoomomi omega _k is converges to the reliability [rho _k. That is, the base weight ω _k is adjusted so that the reliability ρ _k is high and the value for the dimension to be noted at the time of learning is large.

For example, among the state vectors x1 _t and x2 _{t at} the time t in the two-dimensional time series data corresponding to the sensor signal, and the state vectors x3 _t and x4 _{t at} the time t in the two-dimensional time series data corresponding to the motor signal The base weights ω ₁ and ω ₃ corresponding to the node 41 learned by focusing on the state vectors x1 _t and x3 _t are 1, and the base weights ω ₂ and ω ₄ are 0.

The base weight determination unit 121 determines the updated base weight ω ′ _k as a new base weight ω _k and supplies the base weight to a node 141 (FIG. 11 described later) stored in the network storage unit 15 for storage.

FIG. 11 is a diagram illustrating details of the node 141 when the base weight ω _k is stored.

  The node 141 in FIG. 11 includes a dynamical approximate model 51 having an internal state quantity, an input weight storage unit 52, an output weight storage unit 53, a learning degree storage unit 54, and a base weight storage unit 142. 4 that are the same as those in FIG. 4 are denoted by the same reference numerals, and the description thereof will not be repeated.

The base weight storage unit 142 stores the base weight ω _k of each dimension supplied from the base weight determination unit 121 of FIG. The base weight ω _k is used when adjusting an input weight and an output weight used at the time of recognition generation described later.

  Next, a learning process in which the information processing apparatus having the learning unit 120 in FIG. 10 learns the dynamics storage network will be described with reference to FIG. This learning process is started, for example, when the information processing apparatus is turned on.

  The processing in steps S41 to S46 is the same as the processing in steps S1 to S6 in FIG.

After the processing of step S46, the process proceeds to step S47, and the base weight determination unit 121 in FIG. 10 sets the base weight ω _k according to the above-described equation (8) based on the reliability ρ _k from the reliability extraction unit 81. Determine and proceed to step S48.

  The processing in steps S48 to S52 is the same as the processing in steps S7 to S11 in FIG.

As described above, the information processing apparatus in FIG. 1 can learn dynamics efficiently by performing weighting based on the learning input weight αs _k and the learning output weight βs _k, and thus more stable learning. It can be carried out. In addition, more stable recognition and generation can be performed by performing recognition and generation using a dynamics storage network that holds the learned dynamics.

[About recognition generation]
FIG. 13 shows a detailed configuration example of the recognition unit 16 and the generation unit 19 of the information processing apparatus 1 of FIG.

  Here, the recognition and generation of time-series data means that the input time-series data is recognized and new time-series data is generated based on the recognition result.

  According to recognition generation, for example, when a human makes a voice calling to a robot, the robot generates a motor signal for taking a motion corresponding to the call and a synthesized sound responding to the call. A parameter signal or the like is generated.

  As shown in FIG. 13, the recognition unit 16 includes a reliability extraction unit 211, an input weight adjustment unit 212, an output weight adjustment unit 213, an internal state quantity update unit 214, a score calculation unit 215, a determination unit 216, and an output unit 217. Consists of.

It is assumed that the reliability ρ _{k of} each dimension is added to the time series data output from the feature extraction unit 13 as additional information.

Confidence extractor 211 from the reliability [rho _k is added time-series data of each dimension was to extract the reliability [rho _k of each dimension, and supplies the input-weight adjuster 212 and the output-weight adjuster 213.

The input weight adjustment unit 212 is based on the reliability ρ _k from the reliability extraction unit 211 and is stored in the input weight storage unit 52 of FIG. 4 and used for recognition (hereinafter referred to as recognition input weight). αr _k (0 ≦ αr _k ≦ 1) is adjusted. Specifically, the input-weight adjuster 212, based on the reliability [rho _k, value confidence [rho _k is for a high level is increased, so that the value reliability [rho _k is for lower dimensional decreases, recognition input determine the weight .alpha.r _k, stores and supplies the recognition input weights .alpha.r _k to the input-weight storage unit 52.

Similar to the input weight adjustment unit 212, the output weight adjustment unit 213 is based on the reliability ρ _k from the reliability extraction unit 211 and is stored in the output weight storage unit 53 and used for recognition (hereinafter referred to as output weight). Βr _k (0 ≦ βr _k ≦ 1) is adjusted.

  The internal state quantity update unit 214 reads the internal state quantity updated and stored last time from the internal state storage unit 21 into the dynamic system approximation model 51 of each node 41 of the dynamics storage network. Specifically, the internal state quantity update unit 214 reads the internal state quantity from the internal state storage unit 21, and uses the internal state quantity as the internal state quantity of the dynamical approximate model 51 of each node 41, thereby calculating the score calculation unit 215. To supply.

  As a result, in the dynamical approximate model 51, the internal state quantity can be updated using the value read by the internal state quantity updating unit 214 as an initial value based on the input time-series data.

  Also, the internal state quantity update unit 214 is supplied from the score calculation unit 215 and the update value (updated internal state quantity) of the internal state quantity of each node 41 when the winner node is determined by the determination unit 216. The internal state storage unit 21 stores the initial value of the internal state quantity of each node 41 when the winner node is determined.

  Here, the updated value of the internal state quantity stored in the internal state storage unit 21 is read into the dynamic system approximation model 51 and used for the next score calculation. The initial value of the internal state quantity stored in the internal state storage unit 21 is used in the generation unit 19 when generating time-series data.

  Similar to the score calculation unit 84, the score calculation unit 215 applies the time series data from the feature extraction unit 13 to the dynamic system approximation model 51 of each node 41 included in the dynamics storage network stored in the network storage unit 15. The score is calculated while updating the internal state quantity.

That is, the score calculation unit 215 inputs time series data to the addition unit 60 of the dynamic system approximation model 51 of each node 41, and outputs output data corresponding to the input from the output layer 63 of the dynamic system approximation model 51. obtain. Incidentally, the addition unit 60, like the learning time, based on the recognition input weights .alpha.r _k, performs weighting in accordance with the equation a recognition input weights .alpha.r _k learning input weights .alpha.s _k of formula (1) described above. Therefore, by adjusting the recognition input weights .alpha.r _k, it is possible to adjust the dimensions of the time-series data that are to be considered during the recognition.

The score calculation unit 215 also calculates a mean square error as an evaluation value for each node 41 according to an expression in which the learning output weight βs _k of the above-described expression (2) is the recognition output weight βr _k . Therefore, by adjusting the recognition output weight β _k , it is possible to adjust the dimension of the time series data to be noted as the evaluation value at the time of recognition. The score calculation unit 84 assigns the mean square error to each node 41 as a score.

The score calculation unit 215 uses the error amount δ _{co, m} obtained by using the mean square error obtained as the evaluation value at the time of recognition for each node 41 _, and uses the above-described equations (3) and ( According to 4), among the values obtained by updating the predetermined value, the value that makes the score the smallest is the initial value C _{0, m} of the context as the internal state quantity of the dynamical approximate model 51 The score is calculated while updating the context from the initial value C _{0, m} .

  The score calculation unit 215 supplies the score given to each node 41 as a result of the score calculation to the determination unit 216, and the updated value and initial value of the internal state quantity of each node 41 when the score is given Are supplied to the internal state quantity update unit 214 as the update value and initial value of the internal state quantity of each node 41 when the winner node is determined.

As described above, based on the recognition input weights .alpha.r _k, is weighted time-series data input performed from the score calculating unit 215, Evaluation of weighted mean square error based on the recognition output weights .beta.r _k recognition Since it is a value, it is possible to recognize by focusing only on the time-series data to be noted at the time of recognition among the input time-series data. As a result, the recognition processing load can be reduced, and recognition can be performed efficiently based on multidimensional time-series data. In addition, recognition accuracy can be improved.

  Based on the score from the score calculation unit 215, the determination unit 216 determines the node having the smallest score value as the winner node. That is, the determination unit 216 selects the dynamics corresponding to the winner node as the dynamics most suitable for the time series data input from the feature extraction unit 13. The determination unit 216 supplies the output unit 217 with a signal that identifies the winner node corresponding to the dynamics that best matches the input time-series data.

  The output unit 217 outputs a signal specifying the winner node from the determination unit 216 as the recognition result 17. As described above, the recognition unit 16 recognizes the input time-series data. The recognition result 17 obtained as a result of the recognition is used as a control signal 18 for designating which dynamics the time series data is generated from which node 41 in the generation process in the recognition generation.

  13 includes a reliability extraction unit 311, an input weight adjustment unit 312, an output weight adjustment unit 313, a generation node determination unit 314, an internal state reading unit 315, a time series data generation unit 316, and an output unit 317. Composed.

Confidence extractor 311 from the reliability [rho _k is added time-series data of each dimension was to extract the reliability [rho _k of each dimension, and supplies the input-weight adjuster 312 and the output-weight adjuster 313.

The input weight adjustment unit 312 is based on the reliability ρ _k from the reliability extraction unit 311 and is stored in the input weight storage unit 52 of FIG. 4 and used for generation (hereinafter referred to as generation input weight). αg _k (0 ≦ αg _k ≦ 1) is adjusted. Specifically, the input-weight adjuster 312, based on the reliability [rho _k, the value of the reliability [rho _k is high level is increased, so that the value of the confidence [rho _k low dimensionality is reduced, recognition input The weight αg _k is determined, and the recognized input weight αg _k is supplied to and stored in the input weight storage unit 52.

Similar to the input weight adjustment unit 312, the output weight adjustment unit 313 is based on the reliability ρ _k from the reliability extraction unit 311 and is stored in the output weight storage unit 53 and used for generation (hereinafter referred to as output weight). , adjusted) that generates output weights .beta.g _k a _{(0 ≦ βg k ≦ 1)} .

  The recognition result 17 output from the output unit 217 is input to the generation node determination unit 314 as the control signal 18. Based on the control signal 18, the generation node determination unit 314 determines a generation node that is a node 41 that should generate time-series data. That is, the winner node determined by the determination unit 216 is determined as the generation node. The generation node determination unit 314 supplies information specifying the determined generation node to the internal state reading unit 315.

  Based on the information from the generation node determination unit 314, the internal state reading unit 315 transfers the internal state to the dynamic system approximation model 51 of the generation node among the nodes 41 included in the dynamics storage network stored in the network storage unit 15. The value stored in the storage unit 21 is read as the initial value of the internal state quantity.

  That is, the internal state reading unit 315 reads the initial value of the internal state quantity when the generation node is determined as the winner node in the recognition unit 16 among the values stored in the internal state storage unit 21, The initial value of the state quantity is supplied to the time series data generation unit 316 as the internal state quantity of the dynamic system approximation model 51 of the generation node.

  The time series data generation unit 316 is supplied with the time series data supplied from the feature extraction unit 13. The time series data generation unit 316 generates time series data based on the time series data while updating the internal state quantity.

  Specifically, the time series data generation unit 316 inputs the time series data to the addition unit 60 of the dynamic system approximation model 51 of the generation node, and outputs output data corresponding to the input to the output of the dynamic system approximation model 51. Obtained from layer 63.

Incidentally, the addition unit 60, similarly to the learning and recognition, based on the generation input weights .alpha.g _k, performs weighting in accordance with the equation a generation input weights .alpha.g _k learning input weights .alpha.s _k of formula (1) described above . Therefore, by adjusting the generation input weight αg _k , it is possible to adjust the dimension of the time series data to be noted at the time of generation.

Further, the time series data generation unit 316 calculates a mean square error as an evaluation value for each node 41 in accordance with the equation (2) in which the learning output weight βs _k is the generated output weight βg _k . Therefore, by adjusting the generation output weight βg _k , it is possible to adjust the dimension of the time series data to be noted as the evaluation value at the time of generation.

Then, the time series data generation unit 316 uses the δ _{co, m} obtained by using the mean square error obtained as the evaluation value at the time of generation for each node 41 with reference to a predetermined value as described above. Of the values obtained by updating the predetermined value according to the equations (3) and (4), the value that minimizes the mean square error is the dynamic system approximation model 51 as the internal state quantity. of determining the initial value C _{0, m} the context, the dynamical system approximation model 51, while updating the context from an initial value C _{0, m,} to output the output data.

The time series data generation unit 316 weights the output data of each dimension obtained from the dynamical approximate model 51 corresponding to the generation node based on the generation output weight βg _k and generates the result as time series data. To do. That is, the time series data generation unit 316 generates time series data of each dimension at a ratio corresponding to the generation output weight βg _k . Then, the time series data generation unit 316 supplies the generated time series data to the output unit 317.

As described above, the time series data input from the feature extraction unit 13 is weighted based on the generation input weight αg _k , and the output from the dynamic system approximation model 51 is based on the generation output weight βg _k. Since weighting is performed on the data, it is possible to generate time-series data by focusing on only time-series data that should be focused on at the time of generation among the input time-series data. As a result, it is possible to prevent generation of time series data that contradicts the time series data input to the information processing apparatus 1.

  The output unit 317 outputs the time series data from the time series data generation unit 316 as the generation result 20. As described above, the generation unit 19 generates time series data as the generation result 20 from the winner node determined in the recognition process by the recognition unit 16 and outputs the generated time series data.

  Next, a recognition process in which the recognition unit 16 in FIG. 13 recognizes time-series data will be described with reference to FIG. This recognition process is started, for example, when time-series data is input from the feature extraction unit 13.

In step S71, the confidence extractor 211 from the reliability [rho _k is added time-series data of each dimension was to extract the reliability [rho _k of each dimension, the input-weight adjuster 212 and the output-weight adjuster 213 Supply.

After the process of step S71, the process proceeds to step S72, and the input weight adjustment unit 212 is stored in the input weight storage unit 52 of FIG. 4 based on the reliability ρ _k from the reliability extraction unit 211. adjusting the recognition input weights .alpha.r _k used.

After the process of step S72, the process proceeds to step S73, and the output weight adjustment unit 213 stores the output weight storage unit 53 based on the reliability ρ _k from the reliability extraction unit 211, similarly to the input weight adjustment unit 212. It is to adjust the recognition output weights .beta.r _k used during recognition.

  After the processing of step S73, the process proceeds to step S74, where the internal state quantity update unit 214 uses the internal state quantity that was updated and stored last time from the internal state storage unit 21 as the dynamic system of each node 41 of the dynamics storage network. The approximate model 51 is read.

  After the processing of step S74, the process proceeds to step S75, where the score calculation unit 215 performs the dynamic system of each node 41 included in the dynamics storage network stored in the network storage unit 15 for the time series data from the feature extraction unit 13. The score calculation of the approximate model 51 is performed while updating the internal state quantity.

  The score calculation unit 215 supplies the score given to each node 41 as a result of the score calculation to the determination unit 216, and the updated value and initial value of the internal state quantity of each node 41 when the score is given Are supplied to the internal state quantity update unit 214 as the update value and initial value of the internal state quantity of each node 41 when the winner node is determined.

  After the processing of step S75, the process proceeds to step S76, where the internal state quantity update unit 214 updates the internal state quantity of each node 41 supplied from the score calculation unit 215 when the winner node is determined by the determination unit 216. The internal state storage unit 21 stores the value and the initial value of the internal state quantity of each node 41 when the winner node is determined.

  After the process of step S76, the process proceeds to step S77, and the determination unit 216 determines the node having the smallest score value as the winner node based on the score from the score calculation unit 215. The determination unit 216 supplies a signal specifying the winner node to the output unit 217.

  After the process of step S77, the process proceeds to step S78, where the output unit 217 outputs a signal specifying the winner node from the determination unit 216 as the recognition result 17, and ends the process.

  Next, a generation process in which the generation unit 19 in FIG. 13 generates time-series data will be described with reference to FIG. This generation process is started, for example, when the recognition result 17 output from the output unit 217 is input as the control signal 18.

In step S91, the confidence extractor 311 from the reliability [rho _k is added time-series data of each dimension was to extract the reliability [rho _k of each dimension, the input-weight adjuster 312 and the output-weight adjuster 313 Supply.

After the processing of step S91, the process proceeds to step S92, where the input weight adjustment unit 312 is stored in the input weight storage unit 52 of FIG. 4 based on the reliability ρ _k from the reliability extraction unit 311. Adjust the generated input weight αg _k used.

After the processing of step S92, the process proceeds to step S93, and the output weight adjustment unit 313 stores the output weight storage unit 53 based on the reliability ρ _k from the reliability extraction unit 311 as in the case of the input weight adjustment unit 312. The generated output weight βg _k used at the time of generation is adjusted.

  After the processing of step S93, the process proceeds to step S94, and the generation node determination unit 314 determines a generation node that is a node 41 that should generate time-series data based on the control signal 18. The generation node determination unit 314 supplies information specifying the determined generation node to the internal state reading unit 315.

  After the process of step S94, the process proceeds to step S95, and the internal state reading unit 315 is based on the information from the generation node determination unit 314, and the node 41 is included in the dynamics storage network stored in the network storage unit 15. The initial value of the internal state quantity stored in the internal state storage unit 21 is read into the dynamic system approximation model 51 of the generation node.

  After the processing of step S95, the process proceeds to step S96, and the time series data generation unit 316 generates time series data while updating the internal state quantity based on the time series data supplied from the feature extraction unit 13. The time series data generation unit 316 supplies the generated time series data to the output unit 317.

  After the process of step S96, the process proceeds to step S97, where the output unit 317 outputs the time series data from the time series data generation unit 316 as the generation result 20, and ends the process.

  FIG. 16 is a block diagram illustrating another detailed configuration example of the recognition unit and the generation unit in FIG.

  The recognition unit 401 in FIG. 16 includes a reliability extraction unit 411, an input weight extraction unit 412, and an output weight adjustment unit 413 instead of the reliability extraction unit 211, the input weight adjustment unit 212, and the output weight adjustment unit 213. ing.

That is, in FIG. 16, the time-series data input from the feature extraction unit 13, not being added reliability [rho _k for each dimension, recognition input weights .alpha.r _k and recognition output weights .beta.r _k is input directly from the outside Is done.

Weight obtaining unit 411 obtains the recognition input weights .alpha.r _k input from outside the recognition output weights .beta.r _k. Then, the weight obtaining unit 411 supplies the recognition input weights .alpha.r _k to the input-weight adjuster 412 supplies the recognition output weights .beta.r _k to the output-weight adjuster 413.

Input weight adjuster 412, the recognition input weights .alpha.r _k from the weight obtaining unit 411, by the input-weight storage unit 52, adjusts the recognition input weights .alpha.r _k.

Output weight adjuster 413, similarly to the input-weight adjuster 412, the recognition output weights .beta.r _k from the weight obtaining unit 411, by storing the output-weight storage unit 53, adjusts the recognition output weights .beta.r _k.

In addition, the generation unit 402 in FIG. 16 includes a reliability extraction unit 421, an input weight extraction unit 422, and an output weight adjustment unit 423 instead of the reliability extraction unit 311, the input weight adjustment unit 312, and the output weight adjustment unit 313. Is provided. That is, in FIG. 16, the generation input weight αr _k and the generation output weight βr _k are directly input from the outside.

Weight obtaining unit 421 obtains the generation input weights .alpha.r _k and generation output weights .beta.r _k input from the outside. Then, the weight obtaining unit 421 supplies the generated input weights .alpha.r _k to the input-weight adjuster 422 supplies the generated output weights .beta.r _k to the output-weight adjuster 423.

Input weight adjuster 422, the generation input weights .alpha.r _k from the weight obtaining unit 421, by the input-weight storage unit 52, adjusts the generation input weights .alpha.r _k.

Output-weight adjuster 423, similarly to the input-weight adjuster 422, the generation output weights .beta.r _k from the weight obtaining unit 421, by storing the output-weight storage unit 53, adjusts the generation output weights .beta.r _k.

  Next, a recognition process in which the recognition unit 401 in FIG. 16 recognizes time-series data will be described with reference to FIG. This recognition process is started, for example, when time series data is supplied from the feature extraction unit 13.

In step S101, the weight obtaining unit 411 obtains the recognition input weights .alpha.r _k input from outside the recognition output weights .beta.r _k. Then, the weight obtaining unit 411 supplies the recognition input weights .alpha.r _k to the input-weight adjuster 412 supplies the recognition output weights .beta.r _k to the output-weight adjuster 413.

After the step S101, the process proceeds to step S102, input-weight adjuster 412, the recognition input weights .alpha.r _k from the weight obtaining unit 411, by the input-weight storage unit 52, adjusts the recognition input weights .alpha.r _k To do.

After the processing in step S102, the process proceeds to step S103, and the output weight adjustment unit 413 stores the recognized output weight βr _k from the weight acquisition unit 411 in the output weight storage unit 53, similarly to the input weight adjustment unit 412. Accordingly, by adjusting the recognition output weights .beta.r _k, the process proceeds to step S104.

  Note that the processing in steps S104 to S108 is the same as that in steps S74 to S78 in FIG.

  Next, a generation process in which the generation unit 402 in FIG. 16 generates time-series data will be described with reference to FIG. This generation process is started, for example, when the recognition result 17 output from the output unit 217 is input as the control signal 18.

In step S111, the weight obtaining unit 421 obtains the generation input weights .alpha.r _k and generation output weights .beta.r _k input from the outside. Then, the weight obtaining unit 421 supplies the generated input weights .alpha.r _k to the input-weight adjuster 422 supplies the generated output weights .beta.r _k to the output-weight adjuster 423.

After the step S111, the process proceeds to step S112, input-weight adjuster 422 adjusts the generation input weights .alpha.r _k from the weight obtaining unit 421, by the input-weight storage unit 52, a generation input weights .alpha.r _k To do.

After the step S112, the process proceeds to step S113, the output-weight adjuster 423, similarly to the input-weight adjuster 422, to the generation output weights .beta.r _k from the weight obtaining unit 421 causes the output-weight storage unit 53 Accordingly, by adjusting the generation output weights .beta.r _k, the process proceeds to step S114.

  Note that the processing in steps S114 to S117 is the same as the processing in steps S94 to S97 in FIG.

  FIG. 19 is a block diagram showing still another detailed configuration example of the recognition unit and the generation unit in FIG.

  The recognition unit 501 in FIG. 19 includes a reliability calculation unit 511 and an internal state quantity update unit 512 instead of the reliability extraction unit 211 and the internal state quantity update unit 214 in FIG.

That is, in FIG. 19, the time-series data input from the feature extraction unit 13, not being added reliability [rho _k for each dimension, the recognition unit 501, the reliability of the time series data that is input [rho _k Calculate

The reliability calculation unit 511 is supplied with time-series data of each dimension supplied from the feature extraction unit 13. The reliability calculation unit 511 calculates the reliability ρ _k of each dimension based on the time-series data of each dimension and the initial value of the internal state quantity supplied from the internal state quantity update unit 512.

  Specifically, the reliability calculation unit 511 inputs the time series data of each dimension to the addition unit 60 of the dynamical system approximate model 51 of each node 41 of the dynamics storage network stored in the network storage unit 15. Then, the context as the initial value of the internal state quantity is input to the context unit of the input layer 61, and output data that is output in response to the input is obtained from the output layer 63 of the dynamical system approximation model 51. At this time, the adding unit 60 supplies the input time-series data of each dimension to each unit of the input layer 61 as it is.

The reliability calculation unit 511 calculates time-dependent data for each dimension based on the time series data input to the dynamic system approximation model 51, the output data obtained from the dynamic system approximation model 51 immediately before, and the time series data. The true value of the output to be obtained for the series data and the prediction error ε _k (0 <ε ₁ <1), which is an error for each dimension of the output value from the dynamic system approximation model 51 for the time series data, is _represented by a node 41. Calculate every.

Then, the reliability calculation unit 511 calculates the reliability ρ _k of each dimension according to the following equation (9) using the prediction error ε _k for each node 41.

According to equation (9), for the greater dimension of the prediction error epsilon _k, less reliability [rho _k is given, for the small dimension of the prediction error epsilon _k, it is given greater reliability [rho _k.

The reliability calculation unit 511 supplies the reliability ρ _k obtained as a result of the calculation according to Equation (9) to the input weight adjustment unit 212 and the output weight adjustment unit 213.

  The internal state quantity update unit 512 reads the internal state quantity updated and stored last time from the internal state storage unit 21 into the dynamic system approximate model 51 of each node 41 of the dynamics storage network.

  Specifically, the internal state quantity update unit 512 reads the internal state quantity that has been updated and stored last time from the internal state storage unit 21 and stores the internal state quantity in the dynamic system approximation model 51 of each node 41. The initial value of the state quantity is supplied to the reliability calculation unit 511 or supplied to the score calculation unit 215.

  Similarly to the internal state quantity update unit 214 of FIG. 13, the internal state quantity update unit 512 is supplied from the score calculation unit 215 and the internal state of each node 41 when the winner node is determined by the determination unit 216. The update value of the amount and the initial value of the internal state amount of each node 41 when the winner node is determined are stored in the internal state storage unit 21.

In addition, the generation unit 502 in FIG. 19 includes a reliability calculation unit 521 and an internal state reading unit 522 instead of the reliability extraction unit 311 and the internal state reading unit 315 in FIG. That is, in FIG. 19, the generation unit 502 calculates the reliability ρ _k of the input time-series data.

Like the reliability calculation unit 501, the reliability calculation unit 521 is supplied with time-series data of each dimension supplied from the feature extraction unit 13. Similar to the reliability calculation unit 501, the reliability calculation unit 521 determines the reliability ρ of each dimension based on the time-series data of each dimension and the initial value of the internal state quantity supplied from the internal state reading unit 522. _k is calculated. The reliability calculation unit 511 supplies the reliability ρ _k obtained as a result of the calculation to the input weight adjustment unit 312 and the output weight adjustment unit 313.

  Based on the information from the generation node determination unit 314, the internal state reading unit 522 transmits the internal state to the dynamic system approximation model 51 of the generation node among the nodes 41 included in the dynamics storage network stored in the network storage unit 15. The value stored in the storage unit 21 is read as the initial value of the internal state quantity.

  That is, the internal state reading unit 522 reads the initial value of the internal state quantity when the generation node is determined as the winner node in the recognition unit 16 among the values stored in the internal state storage unit 21, The initial value of the state quantity is supplied to the reliability calculation unit 521 and the time series data generation unit 316 as the initial value of the internal state quantity of the dynamic system approximation model 51 of the generation node.

  Next, a recognition process in which the recognition unit 501 in FIG. 19 recognizes time-series data will be described with reference to FIG. This recognition process is started, for example, when time-series data of each dimension is supplied from the feature extraction unit 13.

  In step S121, the internal state quantity update unit 512 reads the internal state quantity updated and stored last time from the internal state storage unit 21 into the dynamic system approximation model 51 of each node 41 of the dynamics storage network.

After the processing of step S121, the process proceeds to step S122, in which the reliability calculation unit 511 initializes the time series data of each dimension supplied from the feature extraction unit 13 and the internal state quantity supplied from the internal state quantity update unit 512. Based on the value, the reliability ρ _k of each dimension is calculated according to the above-described equation (9). The reliability calculation unit 511 supplies the reliability ρ _k to the input weight adjustment unit 212 and the output weight adjustment unit 213.

After the process of step S122, the process proceeds to step S123, and the input weight adjustment unit 212 stores the input weight storage unit 52 in FIG. 4 based on the reliability ρ _k from the reliability calculation unit 511 during recognition. adjusting the recognition input weights .alpha.r _k used.

After the processing of step S123, the process proceeds to step S124, and the output weight adjustment unit 213 stores the output weight storage unit 53 based on the reliability ρ _k from the reliability calculation unit 511, similarly to the input weight adjustment unit 212. is to adjust the recognition output weights .beta.r _k used during recognition, the process proceeds to step S125.

  Note that the processing of steps S125 to S128 is the same as steps S75 to S78 of FIG.

  Next, a generation process in which the generation unit 502 of FIG. 19 generates time-series data will be described with reference to FIG. This generation process is started, for example, when the recognition result 17 output from the output unit 217 is input as the control signal 18.

  In step S131, the generation node determination unit 314 determines a generation node, which is the node 41 that should generate time-series data, based on the control signal 18. The generation node determination unit 314 supplies information specifying the determined generation node to the internal state reading unit 522.

  After the processing of step S131, the process proceeds to step S132, and the internal state reading unit 522 is based on the information from the generation node determination unit 314, among the nodes 41 included in the dynamics storage network stored in the network storage unit 15. The initial value of the internal state quantity stored in the internal state storage unit 21 is read into the dynamic system approximation model 51 of the generation node.

After the process of step S132, the process proceeds to step S133, and the reliability calculation unit 521 supplies the time series data of each dimension supplied from the feature extraction unit 13 and the internal state reading unit 522, like the reliability calculation unit 501. Based on the initial value of the internal state quantity, the reliability ρ _k of each dimension is calculated according to the above-described equation (9). The reliability calculation unit 511 supplies the reliability ρ _k obtained as a result of the calculation to the input weight adjustment unit 312 and the output weight adjustment unit 313.

After the processing of step S133, the process proceeds to step S134, and the input weight adjustment unit 312 is stored in the input weight storage unit 52 of FIG. 4 based on the reliability ρ _k from the reliability calculation unit 511. Adjust the generated input weight αg _k used.

After the processing of step S134, the process proceeds to step S135, and the output weight adjustment unit 313 stores the output weight storage unit 53 in the same manner as the input weight adjustment unit 312 based on the reliability ρ _k from the reliability extraction unit 311. The generated output weight βg _k used at the time of generation is adjusted, and the process proceeds to step S136.

  The processing in steps S136 and S137 is the same as the processing in steps S96 and S97 in FIG.

  FIG. 22 is a block diagram illustrating still another detailed configuration example of the recognition unit and the generation unit in FIG. 1.

  Note that the recognition unit 601 and the generation unit 602 in FIG. 22 correspond to the learning unit 120 in FIG. 10 and perform recognition generation using the dynamics storage network learned by the learning unit 120.

That is, the recognition unit 601 in FIG. 22 includes a base weight acquisition unit 611, an input weight adjustment unit 612, and an output weight adjustment instead of the reliability extraction unit 211, the input weight adjustment unit 212, and the output weight adjustment unit 213 in FIG. The unit 613 is provided, and the recognition input weight αr _k and the recognition output weight βr _k are adjusted based on the base weight ω _k calculated by the learning unit 120 of FIG. 10 during learning.

The base weight acquisition unit 611 acquires the base weight ω _k from the base weight coefficient storage unit 142 of each node 141 of the dynamics storage network stored in the network storage unit 15. The base weight acquisition unit 611 supplies the base weight ω _k to the input weight adjustment unit 612 and the output weight adjustment unit 613.

Input-weight adjuster 612, based on the basis weight omega _k from the base-weight acquisition unit 611, as Motoomomi omega _k recognition input weights .alpha.r _k relative dimensions larger increases, determining the recognition input weights .alpha.r _k And stored in the input weight storage unit 52. For example, the input weight adjustment unit 612 stores the base weight ω _k in the input weight storage unit 52 as the recognition input weight αr _k . Thus, recognition input weights .alpha.r _k stored in the input-weight storage unit 52 is adjusted.

Similar to the input weight adjustment unit 612, the output weight adjustment unit 613 determines the recognition output weight βr _k based on the base weight ω _k from the base weight acquisition unit 611 and stores it in the output weight storage unit 53. adjusts the recognition output weights .beta.r _k. For example, the output weight adjustment unit 612 stores the base weight ω _k as the recognition output weight βr _k in the output weight storage unit 53.

As described above, the output-weight adjuster 613 to the input-weight adjuster 612, based on Motoomomi omega _k, recognition input weights .alpha.r _k and recognition output weights .beta.r _k increases relative dimensions Motoomomi omega _k is greater as such, because it determines the recognition input weights .alpha.r _k and recognition output weights .beta.r _k, the dimension which is focused on the learning is also noted at the time of recognition.

For example, among the state vectors x1 _t and x2 _{t at} the time t in the two-dimensional time series data corresponding to the sensor signal, and the state vectors x3 _t and x4 _{t at} the time t in the two-dimensional time series data corresponding to the motor signal The base weights ω ₁ and ω ₃ corresponding to the node 141 learned by focusing on the state vectors x1 _t and x3 _t are 1, and the base weights ω ₂ and ω ₄ are 0. In this case, for example, the base weight ω _k is determined as it is as the recognition input weight αr _k and the recognition output weight βr _k , and among the four dimensions corresponding to the time-series data supplied from the feature extraction unit 13 at the time of recognition, The first and third dimensions corresponding to the vectors x1 _t and x3 _t are noted.

The input weight adjuster 612 and the output-weight adjuster 613, Motoomomi omega _k instead of directly to input weights .alpha.r _k and output weights .beta.r _k a, multiplied by the confidence [rho _k during recognition Motoomomi omega _k it may be the input weights .alpha.r _k and output weights .beta.r _k what was. In this case, as the reliability [rho _k recognition time, it is possible to use and reliability [rho _k, which is added to the time-series data supplied from the feature extraction unit 13, the calculated confidence [rho _k by the prediction error epsilon _k .

22 includes a base weight acquisition unit 621, an input weight adjustment unit 622, and an output weight adjustment instead of the reliability extraction unit 311, the input weight adjustment unit 312, and the output weight adjustment unit 313 in FIG. A unit 623 is provided, and the generation input weight αg _k and the generation output weight βg _k are adjusted based on the base weight ω _k calculated by the learning unit 120 of FIG. 10 during learning.

The base weight acquisition unit 621 acquires the base weight ω _k from the base weight coefficient storage unit 142 of each node 141 of the dynamics storage network stored in the network storage unit 15. The base weight acquisition unit 621 supplies the base weight ω _k to the input weight adjustment unit 622 and the output weight adjustment unit 623.

Input-weight adjuster 622, like the input-weight adjuster 612, on the basis of the base weights omega _k from base weight obtainer 621, Motoomomi omega _k generation input weights .alpha.g _k to increase relative to dimensions greater Then, the generated input weight αg _k is determined and stored in the input weight storage unit 52. As a result, the generated input weight αg _k stored in the input weight storage unit 52 is adjusted.

Similarly to the output weight adjustment unit 613, the output weight adjustment unit 623 determines the generated output weight βg _k based on the base weight ω _k from the base weight acquisition unit 621 and stores it in the output weight storage unit 53. The generated output weight βg _k is adjusted.

As described above, the input weight adjustment unit 622 and the output weight adjustment unit 623 are similar to the input weight adjustment unit 612 and the output weight adjustment unit 613 based on the base weight ω _k and the generated input weight αg _k and the generated output weight. Since βg _k is determined, the dimension focused on at the time of learning is also focused on at the time of generation.

For example, among the state vectors x1 _t and x2 _{t at} the time t in the two-dimensional time series data corresponding to the sensor signal, and the state vectors x3 _t and x4 _{t at} the time t in the two-dimensional time series data corresponding to the motor signal The base weights ω ₁ and ω ₃ corresponding to the node 41 learned by focusing on the state vectors x1 _t and x3 _t are 1, and the base weights ω ₂ and ω ₄ are 0.

In this case, for example, Motoomomi omega _k is determined to generate input weights .alpha.g _k and generation output weights .beta.g _k as is, in the generation time, of the four-dimensional corresponding to time-series data supplied from the feature extraction unit 13, the state The first and third dimensions corresponding to the vectors x1 _t and x3 _t are noted. As a result, the time-series data generating unit 316 generates only the first-dimensional and third-dimensional output data as the time-series data among the four-dimensional output data obtained from the dynamic system approximate model 51.

In this way, by determining the generation input weight αg _k based on the base weight ω _k , it is possible to prevent output data of a dimension not focused at the time of learning from being output as the generation result 20. Thereby, for example, only the time-series data corresponding to the two-hand signal is generated by the node 141 as the generation node, or only the time-series data corresponding to the both-leg signal is generated.

  Next, a recognition process in which the recognition unit 601 in FIG. 22 recognizes time-series data will be described with reference to FIG. This recognition process is started, for example, when time series data is supplied from the feature extraction unit 13.

In step S <b> 141, the base weight acquisition unit 611 acquires the base weight ω _k from the base weight coefficient storage unit 142 of each node 141 of the dynamics storage network stored in the network storage unit 15. The base weight acquisition unit 611 supplies the base weight ω _k to the input weight adjustment unit 612 and the output weight adjustment unit 613.

After the processing of step S141, the process proceeds to step S142, and the input weight adjustment unit 612 stores the recognized input weight αr _k in the input weight storage unit 52 based on the base weight ω _k from the base weight acquisition unit 611. Accordingly, adjusting the recognition input weights .alpha.r _k.

After the processing of step S142, the process proceeds to step S143, and the output weight adjustment unit 613 determines the recognition output weight βr _k based on the base weight ω _k from the base weight acquisition unit 611, similarly to the input weight adjustment unit 612. by the output-weight storage unit 53, adjusts the recognition output weights .beta.r _k, the process proceeds to step S144.

  The processing in steps S144 to S148 is the same as that in steps S74 to S78 in FIG.

  Next, a generation process in which the generation unit 602 in FIG. 22 generates time-series data will be described with reference to FIG. This generation process is started, for example, when the recognition result 17 output from the output unit 217 is input as the control signal 18.

In step S151, the base weight acquisition unit 621 acquires the base weight ω _k from the base weight coefficient storage unit 142 of each node 141 of the dynamics storage network stored in the network storage unit 15. The base weight acquisition unit 621 supplies the base weight ω _k to the input weight adjustment unit 622 and the output weight adjustment unit 623.

After the processing of step S151, the process proceeds to step S152, and the input weight adjustment unit 622 inputs the generated input weight ω _k based on the base weight ω _k from the base weight acquisition unit 621, similarly to the input weight adjustment unit 612. The generation input weight αg _k is adjusted by storing it in the weight storage unit 52.

After the processing of step S152, the process proceeds to step S153, and the output weight adjustment unit 623, like the output weight adjustment unit 613, generates the generated output weight βg _k based on the base weight ω _k from the base weight acquisition unit 621. By making it memorize | store in the output weight memory | storage part 53, production | generation output weight (beta) _gk is adjusted and it progresses to step S154.

  Note that the processing in steps S154 to S157 is the same as the processing in steps S94 to S97 in FIG.

  FIG. 25 is a block diagram illustrating still another detailed configuration example of the recognition unit in FIG.

  The recognition unit 701 in FIG. 25 is provided with a determination unit 711 instead of the determination unit 216 in FIG. 22, and all of the nodes 141 whose scores are within a predetermined range are determined as winner nodes.

  Based on the score from the score calculation unit 215, the determination unit 711 determines all the nodes corresponding to the score of a value within a predetermined range including the smallest score value as a winner node. That is, the determination unit 711 determines, based on the score, the node having the smallest score value, that is, the node that holds the dynamics most suitable for the time-series data supplied from the feature amount extraction unit 13, as the best winner node. The winner node is determined as the winner node and the node corresponding to the score having a value within a predetermined range from the score value corresponding to the winner node.

  The predetermined range includes a predetermined constant range, a range within A (A> 1) times the smallest score value, and the like. The determination unit 711 supplies a signal specifying the winner node to the output unit 217.

As a result, the generation unit 602 is supplied with a control signal 18 that specifies all nodes 141 corresponding to the score of values within a predetermined range including the smallest score value as winner nodes. As a result, all of the winner nodes are determined as generation nodes, and the time-series data generation unit 316 outputs node 41 for the output data for each dimension obtained from the dynamical approximate model 51 corresponding to each node 141 of the winner node. Each time, weighting is performed based on the generated output weight βg _k, and the output data of each node 41 obtained as a result is combined and generated as time series data.

  For example, the time-series data corresponding to the two-hand signal in the output data of one winner node and the time-series data corresponding to the both-leg signal in the output data of the other winner node Is generated as

  Thus, since the determination unit 711 sets the score of the node 141 determined as the winner node to a score within a predetermined range including the smallest score value, the generation result 20 that contradicts the input observation signal 11 is obtained. Generation | occurrence | production can be suppressed.

In addition, the determination unit 711 determines a node 141 corresponding to a score within a predetermined range including the smallest score value as a winner node, and the time-series data generation unit 316 outputs each dimension output from the winner node. Since the output data is weighted based on the generated output weight βg _k , the time series data output from the node 141 having the highest degree of attention to the dimension among the winner nodes is selected for each dimension. The Thereby, the generation result 20 more suitable for the observation signal 11 can be generated. As a result, for example, it is possible to cause the robot to perform an action in accordance with the actual situation.

  Next, a recognition process in which the recognition unit 701 in FIG. 25 recognizes time-series data will be described with reference to FIG. This recognition process is started, for example, when time series data is supplied from the feature extraction unit 13.

  The processing in steps S161 through S166 is the same as the processing in steps S91 through S96 in FIG.

  After the processing of step S166, the process proceeds to step S167, and the determination unit 711 determines all the nodes corresponding to the score of a value within a predetermined range including the value of the smallest score based on the score from the score calculation unit 215. Determine as the winner node. The determination unit 711 supplies a signal specifying the determined winner node to the output unit 217.

  After the process of step S167, it progresses to step S168, the output part 217 outputs the signal which identifies the winner node from the determination part 216 as the recognition result 17, and complete | finishes a process.

FIGS. 27A and 27B and FIGS. 28A and 28B are graphs showing the results of experiments for verifying the effects of weighting based on the learning input weight αs _k and the learning output weight βs _k during learning.

  In FIGS. 27A and 27B, and FIGS. 28A and 28B, using time-series data corresponding to task A of playing with a red car and time-series data corresponding to task B of raising and lowering a green ball, Learning is done at one RNN.

  27A and 27B, and FIGS. 28A and 28B are time series data obtained by mixing the time series data corresponding to the visual information of task A and task B, that is, a pseudo car created in a pseudo manner. The generation result 20 when the recognition generation is performed by inputting time-series data corresponding to visual information when both the green ball and the green ball are visible is shown.

  In FIGS. 27A and 27B, and FIGS. 28A and 28B, the horizontal axis represents a time step, and the vertical axis represents a value obtained by normalizing the joint angle of the robot for RNN. This also applies to FIGS. 29A and 29B and FIGS. 30A and 30B described later. FIG. 27A and FIG. 28A show the target generation results.

  FIG. 27B is a graph showing a generation result 20 when recognition generation is performed using an RNN that is not weighted during learning. Comparing FIG. 27A and FIG. 27B, when recognition generation is performed with an RNN that is not weighted at the time of learning, a generation result of a time series pattern that is close to the target generation result time series pattern shown in FIG. 27A is generated. It can be seen that is not done. In the experiment, task B could not be recognized and generated.

On the other hand, FIG. 28B shows the generation result 20 when the recognition generation is performed by the RNN weighted by increasing the learning input weight αs _k and the learning output weight βs _k for the dimension relating to vision at the time of learning compared to other dimensions. It is a graph to show. Comparing FIG. 28A and FIG. 28B, when the recognition generation is performed with the RNN weighted at the time of learning, the generation result of the time series pattern close to the target generation time series pattern shown in FIG. 28A may be generated. You can see that it is made. That is, in this case, task A and task B are recognized and generated accurately.

As described above, by performing weighting based on the learning input weight αs _k and the learning output weight βs _k during learning, not only the observation signal 11 that is the same as that during learning but also the observation signal 11 that is different from that during learning is recognized. It can be seen that generation can be performed. That is, it can be seen that robustness is improved not only in the same environment as in learning but also in an unknown environment.

Next, FIGS. 29A and FIG. 29B and FIGS. 30A and FIG. 30B, performs weighting on the basis of the time of recognition generating recognition input weights .alpha.r _k and recognition output weights .beta.r _k, and the generation input weights .alpha.g _k and generation output weights .beta.g _k It is a graph which shows the result of the experiment for verifying the effect by.

  29A and 29B and FIGS. 30A and 30B are graphs showing the results of experiments similar to those shown in FIGS. 27A and 27B and FIGS. 28A and 28B.

  Note that FIG. 29A and FIG. 30A show the target generation results.

  FIG. 29B is a graph showing a generation result 20 when recognition generation is performed with an RNN that is not weighted at the time of recognition generation. Comparing FIG. 29A and FIG. 29B, when recognition generation is performed with an RNN that is not weighted at the time of recognition generation, a generation result of a time series pattern close to the target generation result time series pattern shown in FIG. 29A is generated. You can see that it was not possible. That is, in this case, recognition generation is not performed accurately.

On the other hand, FIG. 30B is carried out a recognition input weights .alpha.r _k for dimensions on Visual upon recognition generating recognition output weights .beta.r _k, and weighting the generated input weights .alpha.g _k and generation output weights .beta.g _k and larger than the other dimension It is a graph which shows the production | generation result 20 when performing recognition production | generation with RRN. Comparing FIG. 30A and FIG. 30B, when recognition generation is performed with the RNN weighted at the time of recognition generation, a generation result of a time series pattern close to the target generation time series pattern shown in FIG. 30A is generated. You can see that That is, in this case, task A and task B are recognized and generated accurately.

Thus, recognition input weights .alpha.r _k and recognition output weights .beta.r _k during recognition generating and by performing weighting on the basis of the generation input weights .alpha.g _k and generation output weights .beta.g _k, to stabilize the capacity of the recognition generating, improve performance You can see that

  As described above, the information processing apparatus 1 in FIG. 1 adjusts the input weight, which is a weighting factor for each dimension, with respect to the input data input to each unit of the input layer 61 of the dynamical approximate model 51 for each dimension. Since the output weighting coefficient, which is the weighting coefficient for each dimension for the output data of a plurality of dimensions, output from each unit of the output layer 63 is adjusted for each dimension, multi-dimensional time-series data can be efficiently converted in the RNN. Can be handled.

  In the above description, the mean square error between the true value of the output to be obtained for the time series data and the output value from the dynamic system approximation model 51 for the time series data is used as a score. It is not limited to this, For example, the distance of a true value and an output value, a probability, etc. may be sufficient. When the score is a distance, the node with the smaller score value is determined as the winner node as in the case of the mean square error, but when the score is the probability, the node with the higher score value is determined as the winner node. Is done.

  In the above description, learning, recognition, and generation are performed using a dynamics storage network including a plurality of nodes, but learning and generation may be performed using one node.

  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.

  Therefore, FIG. 31 shows a configuration example of an embodiment of a computer in which a program for executing the series of processes described above is installed.

  The program can be recorded in advance in a hard disk 805 or a ROM 803 as a recording medium built in the computer.

  Alternatively, the program is temporarily stored in a removable recording medium 811 such as a flexible disk, a CD-ROM (Compact Disc Read Only Memory), an MO (Magneto Optical) disk, a DVD (Digital Versatile Disc), a magnetic disk, or a semiconductor memory. It can be stored permanently (recorded). Such a removable recording medium 811 can be provided as so-called package software.

  The program is installed in the computer from the removable recording medium 811 as described above, or transferred from the download site to the computer wirelessly via a digital satellite broadcasting artificial satellite, or a LAN (Local Area Network), The program can be transferred to a computer via a network such as the Internet. The computer can receive the program transferred in this way by the communication unit 808 and install it in the built-in hard disk 805.

  The computer includes a CPU (Central Processing Unit) 802. An input / output interface 810 is connected to the CPU 802 via a bus 801, and the CPU 802 operates an input unit 807 including a keyboard, a mouse, a microphone, and the like by the user via the input / output interface 810. When a command is input as a result of this, a program stored in a ROM (Read Only Memory) 803 is executed accordingly. Alternatively, the CPU 802 can also read from a program stored in the hard disk 805, a program transferred from a satellite or a network, received by the communication unit 808 and installed in the hard disk 805, or a removable recording medium 811 attached to the drive 809. The program read and installed in the hard disk 805 is loaded into a RAM (Random Access Memory) 804 and executed. Thereby, the CPU 802 performs processing according to the above-described flowchart or processing performed by the configuration of the above-described block diagram. Then, the CPU 802 outputs the processing result from the output unit 806 configured with an LCD (Liquid Crystal Display), a speaker, or the like, for example, via the input / output interface 810, or from the communication unit 808 as necessary. Transmission and further recording on the hard disk 805 are performed.

  Here, in this specification, the processing steps for describing a program for causing a computer to perform various types of processing do not necessarily have to be processed in time series according to the order described in the flowchart, but in parallel or individually. This includes processing to be executed (for example, parallel processing or processing by an object).

  Further, the program may be processed by a single computer, or may be processed in a distributed manner by a plurality of computers. Furthermore, the program may be transferred to a remote computer and executed.

  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.

It is a block diagram which shows the structural example of one Embodiment of the information processing apparatus to which this invention is applied. It is a figure which shows the example of the dynamical system approximation model in the case of learning using two-dimensional time series data. It is a figure which shows the example of the dynamical system approximation model in the case of learning using 6-dimensional time series data. It is a figure which shows the example of a node. It is a figure explaining weighting. It is a block diagram which shows the detailed structural example of a learning part. It is a flowchart explaining a learning process. It is a block diagram which shows the other detailed structural example of a learning part. It is a flowchart explaining another learning process. It is a block diagram which shows the further detailed structural example of a learning part. It is a figure which shows the other example of a node. It is a flowchart explaining other learning processing. It is a block diagram which shows the detailed structural example of a recognition part and a production | generation part. It is a flowchart explaining a recognition process. It is a flowchart explaining a production | generation process. It is a block diagram which shows the other detailed structural example of a recognition part and a production | generation part. It is a flowchart explaining another recognition process. It is a flowchart explaining another production | generation process. It is a block diagram which shows the other detailed structural example of a recognition part and a production | generation part. It is a flowchart explaining another recognition process. It is a flowchart explaining another production | generation process. It is a block diagram which shows the further another detailed structural example of a recognition part and a production | generation part. It is a flowchart explaining other recognition processing. It is a flowchart explaining other generation processing. It is a block diagram which shows the further another detailed structural example of a recognition part and a production | generation part. It is a flowchart explaining other recognition processing. It is a graph which shows the experimental result at the time of performing recognition production | generation using RNN which was not weighted at the time of learning. It is a graph which shows the experimental result at the time of performing recognition production | generation using RNN which weighted at the time of learning. It is a graph which shows the experimental result at the time of not performing weighting at the time of recognition production | generation. It is a graph which shows the experimental result at the time of performing weighting at the time of recognition production | generation. 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

  14 learning units, 15 network storage units, 16 recognition units, 19 generation units, 81 reliability extraction units, 82 input weight adjustment units, 83 output weight adjustment units, 84 score calculation units, 87 parameter update units, 100 learning units, 101 Weight acquisition unit, 120 learning unit, 121 base weight determination unit, 211 reliability extraction unit, 212 input weight adjustment unit, 213 output weight adjustment unit, 215 score calculation unit, 311 reliability extraction unit, 312 input weight adjustment unit, 313 Output weight adjustment unit, 316 time series data generation unit, 401 recognition unit, 411 weight acquisition unit, 402 generation unit, 421 weight acquisition unit, 501 recognition unit, 511 reliability calculation unit, 502 generation unit, 521 reliability calculation unit, 601 recognition unit, 611 base weight acquisition unit, 602 generation unit, 6 1 group weight obtaining unit, 802 CPU, 803 ROM, 804 RAM, 805 hard disk, 811 removable recording medium

Claims (16)

Storage means for storing dynamics in one node and storing a network constituted by a plurality of the nodes;
Input weight coefficient adjusting means for adjusting, for each dimension, an input weight coefficient that is a weight coefficient for each dimension with respect to input data that is time-series data of a plurality of dimensions that is input to the input unit of the node;
An information processing apparatus comprising: output weight coefficient adjustment means for adjusting, for each dimension, an output weight coefficient that is a weight coefficient for each dimension for output data of a plurality of dimensions output from the output unit of the node. In the time series data, reliability is added for each dimension,
The input weight coefficient adjustment means adjusts the input weight coefficient for each dimension based on the reliability for each dimension,
The information processing apparatus according to claim 1, wherein the output weight coefficient adjustment unit adjusts the output weight coefficient for each dimension based on the reliability for each dimension. An acquisition means for acquiring the input weighting coefficient and the output weighting coefficient input from outside;
The input weight coefficient adjustment means adjusts the input weight coefficient for each dimension based on the input weight coefficient acquired by the acquisition means,
The information processing apparatus according to claim 1, wherein the output weight coefficient adjustment unit adjusts the output weight coefficient for each dimension based on the output weight coefficient acquired by the acquisition unit. The output data output from the output unit one time before and the newly observed time-series data are added for each dimension at a ratio corresponding to the input weighting factor, and the resulting data is The information processing apparatus according to claim 1, further comprising: an input unit that inputs the input data to the input unit. The error for each dimension between the input data and the output data output from the output unit immediately before is added at a ratio corresponding to the output weighting coefficient to obtain an output error, and based on the output error. The information processing apparatus according to claim 1, further comprising: an internal state update unit that updates an initial value of the internal state quantity of the node. The error of each dimension between the input data and the output data output from the output unit immediately before is added at a ratio corresponding to the output weighting factor for each node, and an output error of each node is obtained. And determining means for determining a winner node which is a node corresponding to the dynamics most suitable for the input data, based on the output error of each node;
The information processing apparatus according to claim 1, further comprising: a weight updating unit configured to update a weight attached to the connection of each node to a degree corresponding to a distance from the winner node. Based on the reliability for each dimension, a base weight coefficient used as a basis for an input weight coefficient and an output weight coefficient used when recognizing the input data or generating time-series data is calculated for each dimension. The information processing apparatus according to claim 2, further comprising base weight calculation means. By inputting the input data and the previous input data to the input units of the plurality of nodes corresponding to the plurality of nodes, the internal state quantity of the node is updated for each node. The error for each dimension with respect to the output data obtained is added for each node at a ratio corresponding to the output weighting coefficient to obtain an output error for each node. Based on the output error for each node, A determination means for determining a winner node which is a node having a matching dynamics;
The information processing apparatus according to claim 1, further comprising: a recognition unit that outputs information representing the winner node as a recognition result of the time-series data. The input weight adjustment means sets the input weight coefficient for each dimension based on an error for each dimension between the input data and the output data output from the output unit immediately before the input data for each node. Adjust to
The information processing apparatus according to claim 8, wherein the output weight adjustment unit adjusts the output weight coefficient for each dimension based on an error for each dimension for each node. An acquisition means for acquiring a base weight coefficient that is a basis of the input weight coefficient and the output weight coefficient, calculated for each dimension when learning dynamics based on time-series data;
The input weight coefficient adjustment means adjusts the input weight coefficient for each dimension based on the base weight coefficient,
The information processing apparatus according to claim 8, wherein the output weight coefficient adjustment unit adjusts the output weight coefficient for each dimension based on the base weight coefficient. Determining means for determining a generation node that is a node used for generating time-series data of a plurality of dimensions among the plurality of nodes constituting the network;
The generation unit according to claim 1, further comprising: generation means for generating time-series data of a plurality of dimensions while updating an internal state quantity of the generation node by inputting the input data to an input unit of the generation node. Information processing device. The input weight adjustment means adjusts the input weight coefficient for each dimension based on an error for each dimension between the input data and the output data output from the output unit immediately before the input data,
The information processing apparatus according to claim 11, wherein the output weight adjustment unit adjusts the output weight coefficient for each dimension based on an error for each dimension. An acquisition means for acquiring a base weight coefficient that is a basis of the input weight coefficient and the output weight coefficient, calculated for each dimension when learning dynamics based on time-series data;
The input weight coefficient adjustment means adjusts the input weight coefficient for each dimension based on the base weight coefficient,
The information processing apparatus according to claim 11, wherein the output weight coefficient adjustment unit adjusts the output weight coefficient for each dimension based on the base weight coefficient. The information processing apparatus according to claim 11, wherein the generation unit generates the time-series data at a ratio corresponding to the output weight coefficient for each dimension. A weighting factor for each dimension with respect to input data that is time-series data of a plurality of dimensions that is input to the input unit of the node of the network that holds the dynamics in one node and is configured by a plurality of the nodes. Adjust an input weighting factor for each dimension,
An information processing method including a step of adjusting, for each dimension, an output weighting factor that is a weighting factor for each dimension for output data of a plurality of dimensions output from the output unit of the node. A weighting factor for each dimension with respect to input data that is time-series data of a plurality of dimensions that is input to the input unit of the node of the network that holds the dynamics in one node and is configured by a plurality of the nodes. Adjust an input weighting factor for each dimension,
A program that causes a computer to execute a process including a step of adjusting, for each dimension, an output weighting coefficient that is a weighting coefficient for each dimension of output data of a plurality of dimensions that is output from an output unit of the node.

Images