Learning apparatus, learning method, and program for efficiently learning dynamics

• US 7,953,683 B2
• Filed: 04/04/2007
• Issued: 05/31/2011
• Est. Priority Date: 04/06/2006
• Status: Expired due to Fees

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: Yoshiike, Yukiko, Minamino, Katsuki, Suzuki, Hirotaka, Ito, Masato, Kawamoto, Kenta
• Primary Examiner(s): Sparks; Donald
• Assistant Examiner(s): GONZALES, VINCENT
• Application Number:
  US11/732,773
• Publication Number:
  US 20070239644A1
• Time in Patent Office: 1,518 Days
• Field of Search: 706/15, 706/16, 706/25
• US Class Current: 706/25
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent

Abstract

A learning apparatus includes a storage unit configured to store a network formed by a plurality of nodes each holding dynamics; a learning unit configured to learn the dynamics of the network in a self-organizing manner on the basis of observed time-series data; a winner-node determiner configured to determine a winner node, the winner node being a node having dynamics that best match the time-series data; and a weight determiner configured to determine learning weights for the dynamics held by the individual nodes according to distances of the individual nodes from the winner node. The learning unit is configured to learn the dynamics of the network in a self-organizing manner by degrees corresponding to the learning weights.

Claims

1. A learning apparatus including a central processing unit, comprising:

storage means for storing a network formed by a plurality of nodes each holding dynamics;

learning means for learning the dynamics of the network in a self-organizing manner on the basis of observed time-series data;

winner-node determining means for determining a winner node, the winner node being a node having dynamics that best match the time-series data, the winner node being determined according to a prediction error corresponding to an average square error between the time-series data and learning data; and

weight determining means for determining learning weights for the dynamics held by the individual nodes according to distances of the individual nodes from the winner node;

wherein the learning weight is calculated by raising an attenuation coefficient to a power determined by a ratio between the distance of the individual nodes from the winner node and a variable that is adjusted to decrease as time elapses, and

wherein the learning means learns the dynamics of the network in a self-organizing manner by degrees corresponding to the learning weights.

2. The learning apparatus according to claim 1, wherein the dynamics are modeled by a dynamical-system approximation model having internal-state quantities.

3. The learning apparatus according to claim 1, wherein the dynamics are modeled by a recurrent neural network.

4. The learning apparatus according to claim 1, wherein the learning means learns the dynamics of the network by a gradient method based on repeated calculation in which the magnitude of a gradient or the number of iterations of the repeated calculation is determined in accordance with the learning weights.

5. A learning method of a learning apparatus including a central processing unit, comprising the steps of:

determining a winner node among a plurality of nodes holding dynamics and forming a network, the winner node being a node having dynamics that best match observed time-series data;

determining learning weights for the dynamics held by the individual nodes according to distances of the individual nodes from the winner node, the winner node being determined according to a prediction error corresponding to an average square error between the time-series data and learning data; and

learning the dynamics of the network in a self-organizing manner on the basis of the time-series data by degrees corresponding to the learning weights,

wherein the learning weight is calculated by raising an attenuation coefficient to a power determined by a ratio between the distance of the individual nodes from the winner node and a variable that is adjusted to decrease as time elapses.

6. A recording medium storing an executable program that, when executed, controls a computer to execute processing comprising the steps of:

determining a winner node among a plurality of nodes holding dynamics and forming a network, the winner node being a node having dynamics that best match observed time-series data;

determining learning weights for the dynamics held by the individual nodes according to distances of the individual nodes from the winner node, the winner node being determined according to a prediction error corresponding to an average square error between the time-series data and learning data; and

learning the dynamics of the network in a self-organizing manner on the basis of the time-series data by degrees corresponding to the learning weighs,

wherein the learning weight is calculated by raising an attenuation coefficient to a power determined by a ratio between the distance of the individual nodes from the winner node and a variable that is adjusted to decrease as time elapses.

7. A learning apparatus including a central processing unit, comprising:

a storage unit configured to store a network formed by a plurality of nodes each holding dynamics;

a learning unit configured to learn the dynamics of the network in a self-organizing manner on the basis of observed time-series data;

a winner-node determiner configured to determine a winner node, the winner node being a node having dynamics that best match the time-series data, the winner node being determined according to a prediction error corresponding to an average square error between the time-series data and learning data; and

a weight determiner configured to determine learning weights for the dynamics held by the individual nodes according to distances of the individual nodes from the winner node;

wherein the learning weight is calculated by raising an attenuation coefficient to a power determined by the distance of the individual nodes from the winner node and a variable, and

wherein the learning weight is calculated by raising an attenuation coefficient to a power determined by a ratio between the distance of the individual nodes from the winner node and a variable that is adjusted to decrease as time elapses.