SYSTEM, METHOD, AND COMPUTER PROGRAM FOR ESTIMATION OF TARGET VALUE

• US 20190027162A1
• Filed: 01/13/2017
• Published: 01/24/2019
• Est. Priority Date: 01/14/2016
• Status: Active Grant

• Current Assignee: National Institute of Advanced Industrial Science and Technology (Government of Japan)
• Original Assignee: National Institute of Advanced Industrial Science and Technology (Government of Japan)
• Inventors: Fukayama, Satoru, Goto, Masataka
• Granted Patent: US 10,614,830 B2
• Time in Patent Office: Days
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• 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 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G10L 25/45 characterised by the type of analysis window
• CPC Class Codes: G10L 25/51 for comparison or discrimination

Abstract

A system, method, and computer program for estimation of a target value, which can change the aggregation of estimation results based on the degrees of confidence, taking the nature of an input observation signal into consideration. An unknown observation signal is input to a plurality of regression models. A plurality of estimated values are respectively obtained by a plurality of regression models corresponding to the plurality of features of the unknown observation signal. A target value of the unknown observation signal is estimated by aggregation of the target values. The estimating section calculates weights to be added the estimation results output from the regression models, based on the degrees of confidence with respect to the inputs into the regression models. A target value of the unknown observation signal is estimated through the aggregation by calculating a weighted sum of the estimation results output from the regression models.

Claims

1. A system for estimation of target value comprising:

an estimator configuring section operable to extract a plurality of features from each of a plurality of observation signals, and including a plurality of regression models corresponding to the plurality of features, whereby the regression models are optimized based on a plurality of training data comprised of a pair of the extracted features and target values to be estimated for the plurality of observation signals; and

an estimating section operable to input an unknown observation signal to the estimator configuring section, to estimate a plurality of target values for corresponding features of the unknown observation signal using the plurality of the regression models, and to estimate a target value for the unknown observation signal through aggregation of estimation results of the plurality of target values for the corresponding features of the unknown observation signal, wherein;

the plurality of regression models are each capable of obtaining a probability distribution of estimation results and a degree of confidence;

the estimating section is operable to obtain respective weights to be added to a plurality of estimation results output from the plurality of regression models based on the respective degrees of confidence for the inputs to the plurality of regression models, and to estimate the target value for the unknown observation signal by performing the aggregation using a weighted sum of the plurality of estimation results output from the plurality of regression models.

2. The system for estimation of target value according to claim 1, wherein:

the estimator configuring section includes;

a feature extracting section operable to extract the plurality of features from each of the plurality of observation signals; and

a regression model creating and storing section operable to create and store the plurality of regression models corresponding to the plurality of features, whereby the regression models are optimized based on the plurality of training data comprised of a pair of the extracted features and target values to be estimated for the plurality of observation signals;

the estimating section includes;

a degree-of-confidence calculating section operable to calculate the degree of confidence for the input to each of the plurality of regression models;

a weight calculating section operable to calculate respective weights to be added to the plurality of estimation results output from the plurality of regression models, based on the calculated degrees of confidence; and

a weighted-sum calculating and target-value estimating section operable to estimate the target value for the unknown observation signal by performing the aggregation using the weighted sum of the plurality of estimation results output from the plurality of regression models.

3. The system for estimation of target value according to claim 2, wherein:

the estimator configuring section and the estimating section are implemented on a computer.

4. The system for estimation of target value according to claim 2, wherein:

the estimator configuring section and the estimating section are each comprised of a plurality of processors and a plurality of memories; and

the feature extracting section, the regression model creating and storing section, the degree-of-confidence calculating section, the weight calculating section, and the weighted-sum calculating and target-value estimating section are each comprised of at least one processor and at least one memory.

5. The system for estimation of target value according to claim 1, wherein:

the observation signal is a music audio signal and the target value for the unknown observation signal is a music emotion value.

6. The system for estimation of target value according to claim 1, wherein:

the regression models are capable of regression by Gaussian process, Kriging method, or Bayesian estimation.

7. The system for estimation of target value according to claim 6, wherein:

the estimation results are calculated for the individual regression models, based on a mean of the probability distributions obtained from the regression models to which the unknown observation signal is input; and

the degree of confidence is calculated for the individual regression models, based on a variance of the probability distributions obtained using the regression models.

8. The system for estimation of target value according to claim 1, wherein:

the plurality of features are categorized into a plurality of feature groups; and

the regression models are created correspondingly to the plurality of feature groups.

9. The system for estimation of target value according to claim 8, wherein:

the observation signal is a music audio signal and the target value for the unknown observation signal is a music emotion value;

the feature groups include at least a temporal feature group, a spectral feature group, a pitch-related feature group, and an amplitude spectral feature group.

10. The system for estimation of target value according to claim 2, wherein:

the degree-of-confidence calculating section defines the degree of confidence such that the degree of confidence is higher in a region with a smaller variance in the estimation results of the feature groups as obtained using the plurality of regression models, and the degree of confidence is lower in a region with a larger variance in the estimation results.

11. The system for estimation of target value according to claim 2, wherein:

the degree-of-confidence calculating section defines, in a linear relation, the plurality of degrees of confidence obtained from the plurality of regression models as pre-normalized weights, and defines, as the weights, normalized results obtained by normalizing the pre-normalizing weights to sum up to one.

12. The system for estimation of target value according to claim 2, wherein:

the degree-of-confidence calculating section defines, in a nonlinear relation, pre-normalized weights such that the higher the degrees of confidence obtained from the plurality of regression models are, the larger the pre-normalized weights are emphasized to be, and that the lower the degrees of confidence obtained from the plurality of regression models are, the smaller the pre-normalized weights are emphasized to be, and defines, as the weights, normalized results obtained by normalizing the pre-normalizing weights to sum up to one.

13. The system for estimation of target value according to claim 1, wherein:

the estimating section has a function of interpolation of the estimation results according to a proximity of an input value to known input values prepared for the feature groups in the plurality of regression models when the input value is different from the known input values.

14. The system for estimation of target value according to claim 5, wherein:

the music emotion value is a two-dimensional value comprised of valence and arousal; and

the estimating section outputs weighted sums of the two-dimensional values estimated by the plurality of regression models as an estimated value for valence and an estimated value for arousal.

15. A method for estimation of target value comprising the steps of:

preparing a plurality of regression models, whereby a plurality of features are extracted from each of a plurality of observation signals; and

the regression models correspond to the plurality of features and are optimized based on a plurality of training data comprised of a pair of the extracted features and target values to be estimated for the plurality of observation signals;

inputting an unknown observation signal to the plurality of regression models; and

estimating a plurality of target values for the corresponding features of the unknown observation signal, using the plurality of regression models, and estimating a target value for the unknown observation signal through aggregation of estimation results of the plurality of target values for the corresponding features, wherein;

the plurality of regression models are each capable of obtaining a probability distribution of the estimation results and a degrees of confidence; and

in the step of estimating, respective weights to be added to a plurality of estimation results output from the plurality of regression models are obtained, based on the respective degrees of confidence for the inputs to the plurality of regression models, and a target value is estimated for the unknown observation signal by performing the aggregation using a weighted sum of the plurality of estimation results output from the plurality of regression models.

16. A computer program for estimation of target value, recorded in a computer-readable non-transitory recording medium, the computer program being implemented on a computer to execute the steps of:

preparing a plurality of regression models, whereby a plurality of features are extracted from each of a plurality of observation signals; and

the regression models correspond to the plurality of features and are optimized based on a plurality of training data comprised of a pair of the extracted features and target values to be estimated for the plurality of observation signals;

inputting an unknown observation signal to the plurality of regression models; and

estimating a plurality of target values for the corresponding features of the unknown observation signal, using the plurality of regression models, and estimating a target value for the unknown observation signal through aggregation of estimation results of the plurality of target values for the corresponding features, wherein;

the plurality of regression models are each capable of obtaining a probability distribution of the estimation results and a degrees of confidence; and

in the step of estimating, respective weights to be added to a plurality of estimation results output from the plurality of regression models are obtained, based on the respective degrees of confidence for the inputs to the plurality of regression models, and a target value is estimated for the unknown observation signal by performing the aggregation using a weighted sum of the plurality of estimation results output from the plurality of regression models.