Recognition apparatus using neural network, and learning method therefor

• US 6,151,592 A
• Filed: 01/20/1998
• Issued: 11/21/2000
• Est. Priority Date: 06/07/1995
• Status: Expired due to Term

• Current Assignee: Seiko Epson Corporation (Seiko Group)
• Original Assignee: Seiko Epson Corporation (Seiko Group)
• Inventors: Inazumi, Mitsuhiro
• Primary Examiner(s): Hafiz, Tariq H.
• Assistant Examiner(s): DAVIS, GEORGE B
• Application Number:
  US09/008,321
• Time in Patent Office: 1,036 Days
• Field of Search: 704/249, 704/232, 704/236, 704/246, 706/30, 706/16, 706/25, 706/39
• US Class Current: 706/16
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A recognition apparatus and method using a neural network is provided. A neuron-like element stores a value of its inner condition. The neuron-like element also updates a values of its internal status on the basis of an output from the neuron-like element itself, outputs from other neuron-like elements and an external input, and an output value generator a value of its internal status into an external output. Accordingly, the neuron-like element itself can retain the history of input data. This enables the time series data, such as speech, to be processed without providing any special devices in the neural network.

Claims

1. A recognition apparatus comprising:

feature extracting means for extracting values of an input to be recognized and for inputting extracted values into a recurrent neural network, and

a learning section for causing said recurrent neural network to learn, the learning section comprising;

input data storage means for storing input learning data of a plurality of continuous data streams within a plurality of categories;

input data selection means for selecting input learning data of a plurality of continuous data streams to be learned within a plurality of categories from said input data storage means;

output data storage means comprising a positive output data storage means and a negative output data storage means for storing output learning data of a plurality of continuous data streams within a plurality of categories each of which corresponds to an input learning data category;

output data selection means for selecting output learning data of a plurality of continuous data streams to be learned, each of which corresponds to an input learning data category selected by said input data selection means from said output data storage means;

input data connecting means for connecting the input learning data selected by said input data selection means into a single continuous data stream;

output data connecting means for connecting the output learning data selected by said output data selection means into a single continuous data stream in correlation with the connection of said input learning data; and

learning control means for inputting said connected input learning data stream into said feature extracting means and for changing weightings at connections of neuron elements on the basis of outputs of said recurrent neural network and said connected output learning data stream.

2. The recognition apparatus as in claim 1, the recurrent neural network comprising layered and partially recurrent components comprising output limiting type, dynamic neuron elements, each of the neuron elements comprising:

internal state value storage means, for storing a current internal state value, said storage means comprising an integrator for integrating accumulated values;

internal state value updating means for inputting accumulated values into said internal state value storage means comprising a weighted accumulator for performing weighted accumulations of input values to said dynamic neuron element and the current internal state value; and

an output value limiter for converting said internal state value into an external output value between an upper preset limit and a lower preset limit, said integrator and accumulator forming a feedback loop in said dynamic neuron element.

3. The recognition apparatus as in claim 1, wherein the number of said categories is equal to 2.

4. The recognition apparatus as in claim 1, wherein said learning section further comprises:

noise data storage means for storing noise data;

noise overlaying means for overlaying said selected input learning data with the noise data from noise data storage means, the selected input learning data overlaid with the noise data by said noise overlying means causing said recurrent neural network to learn; and

means for shifting said noise data to different overlaying positions on said selected input learning data to repeat learning.

5. The recognition apparatus as in claim 4, further comprising:

means for training the recurrent neural network to learn selected input learning data not overlaid with the noise data, and selected portions of the input learning data overlaid with the noise data.

6. The recognition apparatus as in claim 1, further comprising a limiter for converting an internal state value into an external output value between an upper preset limit and a lower preset limit and outputting a symmetrical range of values in both positive and negative directions.

7. The recognition apparatus as in claim 1, wherein said recurrent neural network has a plurality of outputs each of which is corresponding to a said input learning data category.

8. A method for recognizing signals, comprising:

extracting values of an input to be recognized;

inputting the extracted values into a recurrent neural network;

storing input learning data of a plurality of continuous data streams within a plurality of categories;

selecting input learning data of a plurality of continuous data streams to be learned within a plurality of categories;

storing positive output learning data of a plurality of continuous data streams within a plurality of categories corresponding to an input learning data category;

storing negative output learning data of a plurality of continuous data streams within a plurality of categories corresponding to an input learning data category;

selecting output learning data of a plurality of continuous data streams to be learned, each of which corresponds to an input learning data category;

connecting the selected input learning data into a single continuous data stream;

connecting the selected output learning data into a single continuous data stream in correlation with the connection of said input learning data;

inputting said connected input learning data stream to the extraction step; and

changing weightings at connections of neuron elements on the basis of outputs of said recurrent neural network and said connected output learning data streams.

9. The method for recognizing signals as in claim 8, wherein the number of said input learning data categories is equal to two.

10. The method for recognizing signals as in claim 8, further comprising:

storing noise data;

overlaying said selected input learning data with noise data selected from the stored noise data, the selected input learning data overlaid with the noise data causing said recurrent neural network to learn so that said noise data contained in said input learning data can be differentiated.

11. The method for recognizing signals as in claim 10, further comprising shifting said noise data to different overlaying positions on said selected input learning data to repeat learning.

12. The method for recognizing signals as in claim 10, wherein selected input learning data not overlaid with the noise data and selected portions of the input learning data overlaid with the noise data are used for learning.