Distance metrics for universal pattern processing tasks

• US 8,856,002 B2
• Filed: 04/11/2008
• Issued: 10/07/2014
• Est. Priority Date: 04/12/2007
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Kanevsky, Dimitri, Nahamoo, David, Sainath, Tara N
• Primary Examiner(s): GUERRA-ERAZO, EDGAR X
• Application Number:
  US12/101,791
• Publication Number:
  US 20090259471A1
• Time in Patent Office: 2,370 Days
• Field of Search: 704/260, 704/243, 704/244, 704/246, 704/256, 704/256.7, 704/235, 704/236, 704/240, 704/255, 704/277
• US Class Current: 704/235
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 17/04 Training, enrolment or model building

Abstract

A universal pattern processing system receives input data and produces output patterns that are best associated with said data. The system uses input means receiving and processing input data, a universal pattern decoder means transforming models using the input data and associating output patterns with original models that are changed least during transforming, and output means outputting best associated patterns chosen by a pattern decoder means.

Claims

1. A universal pattern processing system receiving input data and producing output patterns that are best associated with said input data, comprising:

an input unit configured to input and process said input data and a set of models, each model in said set of models representing a different event;

a universal pattern decoder unit comprising a transforming unit, said universal pattern decoder unit being configured to transform, in a processor in a computer, models using the input data and to associate output patterns with original models that are changed a least amount during the transforming, said universal pattern decoder unit being configured to select criteria for training said set of models and selecting control parameters, which control how fast and to what degree a model is transformed while the model is exposed to the input data; and

an output unit configured to output best associated patterns chosen by the universal pattern decoder unit,

wherein said universal pattern decoder unit associates one of text data, an audio signal, a video signal, and a biometric signal of said input data with one of decoded words for speech recognition, a translation, and a biometric for a speaker identification of said output patterns.

2. The system as claimed in claim 1, where the universal pattern decoder unit comprises at least one of a speech recognizer, an image recognizer, a segmentation system, a machine translation, a search engine, a text-to-speech, a pattern recognizer, and a language processor.

3. The system as claimed in claim 1, wherein the transformation unit comprises at least one of Maximum Likelihood training, Maximum Entropy training, EM training, Maximum Mutual Information training, Baum-Welch training, Extended Baum-Welch training, Bayesian network training, supervised training, unsupervised training, and Hidden Markov Model training.

4. The system as in claim 1, wherein a measuring of a change during a transformation comprises at least one of the following metrics:

one or more differences scores of transformed and original models;

one or more of slopes through a score curve of model transformations;

one or more tangents to a score curve of model transformations; and

one or more tangent hyperplanes to a score manifold of model transformations.

16. The system as claimed in claim 4, wherein said measuring of said change further comprises a vector support machine separation metric for one or more objects in any one of said differences scores of transformed models and original models, slopes through said score curve of model transformations, tangents to said score curve of model transformations, and tangent hyperplanes to said score manifold of model transformations.

17. The system as claimed in claim 4, wherein said measuring of said change further comprises a Hidden Markov Models metric that is accumulated from local Hidden Markov Model metrics for Hidden-Markov-Model states.

5. A method of universal pattern processing in a universal pattern processor that receives input data and produces output patterns that are best associated with said input data, said method comprising:

receiving and processing input data and a set of models, each model in said set of models representing a different event;

transforming, using a computer processor, universal pattern processing models using the input data, criteria selected for training said set of models and selected control parameters, which control how fast and to what degree a model is transformed while the model is exposed to the input data;

associating output patterns with original models that changed least during the transformation process; and

outputting best associated patterns,

wherein said associating comprises associating one of text data, an audio signal, a video signal, and a biometric signal of said input data with one of decoded words for speech recognition, a translation, and a biometric for a speaker identification of said output patterns.

6. The method as claimed in claim 5, wherein the input data processing comprises at least one of the following:

signal processing, textual processing, audio processing, image processing, video processing, binary data processing, analog data processing, biometrics sensor processing, and language processing.

7. The method as claimed in claim 5, wherein the universal pattern decoder process comprises at least one of the following:

a speech recognition process, an image recognition process, segmentation process, a machine translation, a search, a text-to-speech process, a pattern recognition process, and a summarization or simplification.

8. The method as claimed in claim 5, wherein the transformation process comprises at least one of the following:

Maximum Likelihood training, Maximum Entropy training, EM training, Maximum Mutual Information training, Baum-Welch training, Extended Baum-Welch training, Bayesian network training, supervised training, and unsupervised training.

9. The method as claimed in claim 5, wherein measuring a change of model process during a transformation comprises at least one of the following:

subtracting scores of transformed models from scores of original models;

computing slopes through a score curve of model transformations;

computing tangents to a score curve of model transformations;

computing tangent hyperplanes to a score manifold of model transformations;

computing support machine separation metric for one or more objects in any one of said differences scores of transformed models and original models, slopes through said score curve of model transformations, tangents to said score curve of model transformations, and tangent hyperplanes to said score manifold of model transformations; and

computing a Hidden Markov Models metric that is accumulated from local Hidden Markov Model metrics for Hidden Markov Model states.

10. A non-transitory computer-readable storage medium on which are encoded machine-readable instructions which, when executed, cause a computer to perform the method of claim 5.

11. A digital computer comprising the non-transitory computer-readable storage medium of claim 10.

12. A system, comprising:

an input unit configured to receive input data and a set of models;

a processor comprising a change determination unit configured to apply the input data to the set of models to transform the set of models, and configured to determine which of said models is changed least during the transforming, wherein said to apply and to determine comprises associating one of text data, an audio signal, a video signal, and a biometric signal of said input data with one of decoded words for speech recognition, a translation, and a biometric for a speaker identification of said out-put patterns.

13. The system according to claim 12, wherein said change determination unit is configured to determine a degree to which each of the set of models is changed.

14. The system according to claim 13, wherein said determination unit determines the degree to which each of the set of models is changed by comparing an original model to a transformed model using a plurality of metrics.

15. The system according to claim 12, wherein the set of models are transformed using controlling parameters, which control how fast and to what degree a model is transformed while the model is exposed to the input data.