Data Processing Device, Data Processing Method, and Program

• US 20070250461A1
• Filed: 04/05/2007
• Published: 10/25/2007
• Est. Priority Date: 04/06/2006
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: Yokono, Jun, Sabe, Kohtaro, Aoyama, Kazumi, Noda, Kuniaki, Hidai, Kenichi
• Granted Patent: US 7,783,585 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: B25J 9/161 Hardware, e.g. neural networks, fuzzy logic, interfaces, processor
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning

Abstract

A data processing device for processing time-sequence data includes a learning unit for performing self-organizing learning of a SOM (self-organization map) making up a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure, using, as SOM input data which is input to the SOM, a time-sequence of node information representing a winning node of a lower-order SOM which is at a lower hierarchical level from the SOM.

Claims

1. A data processing device for processing time-sequence data, said data processing device comprising:

learning means for performing self-organizing learning of a SOM (self-organization map) making up a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure, using, as SOM input data which is input to said SOM, a time-sequence of node information representing a winning node of a lower-order SOM which is at a lower hierarchical level from said SOM.

2. The data processing device according to claim 1,

wherein said learning means perform

self-organizing learning of a SOM making up said hierarchical SOM using SOM input data, and

prediction learning which updates so as to intensify a weighting which represents the degree of jointing between

a winning node of a SOM as to SOM input data of a predetermined frame length and

a winning node of a SOM as to SOM input data of said predetermined frame length at the next point-in-time;

and wherein self-organizing learning and prediction learning of a SOM of interest of the plurality of SOMs making up said hierarchical SOM is performed using a time-sequence of node information representing a winning node of a lower-order SOM which is at a lower hierarchical level from said SOM.

3. The data processing device according to claim 2, wherein a plurality of lower-order SOMs are connected to said SOM of interest;

and wherein said learning means perform self-organizing learning and prediction learning of the SOM of interest using, as SOM input data, a time-sequence wherein is integrated said node information obtained from each of the plurality of lower-order SOMs connected to said SOM of interest.

4. The data processing device according to claim 2, further comprising

recognition generating means for generating prediction time-sequence data predicting time-sequence data of a point-in-time subsequent to time-sequence data at a certain point-in-time, using said hierarchical SOM regarding which self-organizing learning and prediction learning has been performed.

5. The data processing device according to claim 4, wherein said recognition generating means

in an order from a lower hierarchical level toward a higher hierarchical level

obtain a node which is a winning node for each SOM of said hierarchical SOM;

in an order from a higher hierarchical level toward a lower hierarchical level

obtain a prediction node, which is a node predicted to serve as the winning node in the subsequent point-in-time, based on said weight of said winning node for SOMs regarding which there is no higher order SOM which is a higher order hierarchical level SOM, and

obtain a prediction node based on said weight of said winning node and a prediction node of the higher order SOM for SOMs regarding which there is a higher order SOM; and

generate said prediction time-sequence data using, of the plurality of SOMs making up said hierarchical SOM, the prediction node of the lowest hierarchical level.

6. The data processing device according to claim 4, wherein said recognition generating means obtain a node to be a winning node in a SOM, in an order from a lower hierarchical level toward a higher hierarchical level, and take a time-sequence of node information representing said winning node as SOM input data to a higher order SOM which is a SOM of a higher hierarchical level;

and obtain a prediction node which is a node of a SOM predicted to be a winning node at the next point-in-time, in an order from a higher hierarchical level toward a lower hierarchical level, based on a prediction node of a higher order SOM of said SOM.

7. The data processing device according to claim 6, wherein said recognition generating means further obtain a transition probability which is a probability that a winning node makes transition in a SOM, based on said weight;

and obtain a prediction node of the SOM of interest, based on said transition probability of the SOM of interest and a prediction node of a higher order SOM of said SOM.

8. The data processing device according to claim 7, wherein said recognition generating means

in an order from a lower hierarchical level toward a higher hierarchical level

obtain an observation likelihood that SOM input data will be observed in a node of a SOM making up said hierarchical SOM,

obtain a state probability at the current point-in-time, which represents the probability that a node of a SOM is a winning node, based on said observation likelihood,

obtain the winning node of a SOM, based on said state probability at the current point-in-time, and

take a time-sequence of node information representing a winning node of the SOM of interest as SOM input data to a higher order SOM of the SOM of interest;

in an order from a higher hierarchical level toward a lower hierarchical level obtain a transition probability of the SOM of interest using said weight,

obtain a transition probability of the next point-in-time, based on the transition probability of the SOM of interest and a prediction node of a higher order SOM to said SOM of interest, and

obtain a prediction node of said SOM of interest based on said state probability at the subsequent point-in-time; and

generate said prediction time-sequence data using, of the plurality of SOMs making up said hierarchical SOM, the prediction node of the lowest hierarchical level.

9. The data processing device according to claim 8, wherein said recognition generating means obtain said state probability by applying a particle filter to a SOM making up said hierarchical SOM.

10. The data processing device according to claim 2, wherein the higher the hierarchical level SOM that said SOM input data of a predetermined length is used at is, the longer the time-wise length of said input data is.

11. The data processing device according to claim 4, wherein said recognition generating means obtain a prediction node of the SOM of the lowest hierarchical level of the plurality of SOMs making up said hierarchical SOM, by applying to said hierarchical SOM a Bayesian network solution using a λ

message and a π

message, and using the prediction node to generate said prediction time-sequence data.

12. A data processing method for processing time-sequence data, said method comprising the step of:

performing self-organizing learning of a SOM (self-organization map) making up a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure, using, as SOM input data which is input to said SOM, a time-sequence of node information representing a winning node of a lower-order SOM which is at a lower hierarchical level from said SOM.

13. A program for causing a computer to execute data processing for processing time-sequence data, said program comprising the step of:

performing self-organizing learning of a SOM (self-organization map) making up a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure, using, as SOM input data which is input to said SOM, a time-sequence of node information representing a winning node of a lower-order SOM which is at a lower hierarchical level from said SOM.

14. A data processing device for processing time-sequence data, said data processing device comprising:

storage means for storing a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure; and

recognition generating means for generating prediction time-sequence data wherein time-sequence data at a point-in-time subsequent to time-sequence data at a certain point in time is predicted, using said hierarchical SOM.

15. The data processing device according to claim 14,

wherein said hierarchical SOM stored in said storage means is a hierarchical SOM obtained by performing

self-organizing learning of a SOM making up said hierarchical SOM, using SOM input data which is input to the SOM, and

prediction learning which updates so as to intensify a weighting which represents the degree of jointing between

a winning node of a SOM as to SOM input data of a predetermined frame length and

a winning node of a SOM as to SOM input data of said predetermined frame length at the next point-in-time;

and wherein self-organizing learning and prediction learning of a SOM of interest of the plurality of SOMs making up said hierarchical SOM is performed using a time-sequence of node information representing a winning node of a lower-order SOM which is at a lower hierarchical level from said SOM.

16. The data processing device according to claim 15, wherein said recognition generating means

in an order from a lower hierarchical level toward a higher hierarchical level

obtain a node which is a winning node for each SOM of said hierarchical SOM;

in an order from a higher hierarchical level toward a lower hierarchical level

obtain a prediction node, which is a node predicted to serve as the winning node in the subsequent point-in-time, based on said weight of said winning node for SOMs regarding which there is no higher order SOM which is a higher order hierarchical level SOM, and

obtain a prediction node based on said weight of said winning node and a prediction node of the higher order SOM for SOMs regarding which there is a higher order SOM in hierarchical SOM; and

generate said prediction time-sequence data using, of the plurality of SOMs making up said hierarchical SOM, the prediction node of the lowest hierarchical level.

17. The data processing device according to claim 15, wherein said recognition generating means obtain a node to be a winning node in a SOM, in an order from a lower hierarchical level toward a higher hierarchical level, and take a time-sequence of node information representing said winning node as SOM input data to a higher order SOM which is a SOM of a higher hierarchical level;

and obtain a prediction node which is a node of a SOM predicted to be a winning node at the next point-in-time, in an order from a higher hierarchical level toward a lower hierarchical level, based on a prediction node of a higher order SOM of said SOM.

18. The data processing device according to claim 17, wherein said recognition generating means further obtain a transition probability which is a probability that a winning node makes transition in a SOM, based on said weight;

and obtain a prediction node of the SOM of interest, based on said transition probability of the SOM of interest and a prediction node of a higher order SOM of said SOM.

19. The data processing device according to claim 18, wherein said recognition generating means

in an order from a lower hierarchical level toward a higher hierarchical level obtain an observation likelihood that SOM input data will be observed in a node of a SOM making up said hierarchical SOM,

obtain a state probability at the current point-in-time, which represents the probability that a node of a SOM is a winning node, based on said observation likelihood,

obtain the winning node of a SOM, based on said state probability at the current point-in-time, and

take a time-sequence of node information representing a winning node of the SOM of interest as SOM input data to a higher order SOM of the SOM of interest;

in an order from a higher hierarchical level toward a lower hierarchical level obtain a transition probability of the SOM of interest using said weight,

obtain a state probability of the next point-in-time, based on the transition probability of the SOM of interest and a prediction node of a higher order SOM to said SOM of interest, and

obtain a prediction node of said SOM of interest based on said state probability at the subsequent point-in-time; and

generate said prediction time-sequence data using, of the plurality of SOMs making up said hierarchical SOM, the prediction node of the lowest hierarchical level.

20. The data processing device according to claim 19, wherein said recognition generating means obtain said state probability by applying a particle filter to a SOM making up said hierarchical SOM.

21. The data processing device according to claim 15, wherein the higher the hierarchical level SOM that said SOM input data of a predetermined length is used at is, the longer the time-wise length of said input data is.

22. The data processing device according to claim 15, wherein said recognition generating means obtain a prediction node of the SOM of the lowest hierarchical level of the plurality of SOMs making up said hierarchical SOM, by applying to said hierarchical SOM a Bayesian network solution using a λ

message and a π

message, and using the prediction node to generate said prediction time-sequence data.

23. A data processing method for processing time-sequence data, said method comprising the step of:

generating prediction time-sequence data wherein time-sequence data at a point-in-time subsequent to time-sequence data at a certain point in time is predicted, using a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure.

24. A program for causing a computer to execute data processing for processing time-sequence data, said program comprising the step of:

generating prediction time-sequence data wherein time-sequence data at a point-in-time subsequent to time-sequence data at a certain point in time is predicted, using a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure.

25. A data processing device for processing time-sequence data, said data processing device comprising:

a learning unit for performing self-organizing learning of a SOM (self-organization map) making up a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure, using, as SOM input data which is input to said SOM, a time-sequence of node information representing a winning node of a lower-order SOM which is at a lower hierarchical level from said SOM.

26. A data processing device for processing time-sequence data, said data processing device comprising:

a storage unit for storing a hierarchical SOM in which a plurality of SOMs are connected so as to construct a hierarchical structure; and

a recognition generating unit for generating prediction time-sequence data wherein time-sequence data at a point-in-time subsequent to time-sequence data at a certain point in time is predicted, using said hierarchical SOM.