Nonlinear function approximation over high-dimensional domains

• US 8,521,488 B2
• Filed: 06/10/2011
• Issued: 08/27/2013
• Est. Priority Date: 09/05/2006
• Status: Active Grant

• Current Assignee: Colorado State University Research Foundation (Colorado State University System)
• Original Assignee: National Science Foundation (Government of the United States of America)
• Inventors: Kirby, Michael J., Jamshidi, Arthur A.
• Primary Examiner(s): Jacob, Mary C
• Application Number:
  US13/134,597
• Publication Number:
  US 20110257950A1
• Time in Patent Office: 809 Days
• Field of Search: 703/2
• US Class Current: 703/2
• CPC Class Codes: G06F 17/175 of multidimensional data

Abstract

An algorithm is disclosed for constructing nonlinear models from high-dimensional scattered data. The algorithm progresses iteratively adding a new basis function at each step to refine the model. The placement of the basis functions is driven by a statistical hypothesis test that reveals geometric structure when it fails. At each step the added function is fit to data contained in a spatio-temporally defined local region to determine the parameters, in particular, the scale of the local model. The proposed method requires no ad hoc parameters. Thus, the number of basis functions required for an accurate fit is determined automatically by the algorithm. The approach may be applied to problems including modeling data on manifolds and the prediction of financial time-series. The algorithm is presented in the context of radial basis functions but in principle can be employed with other methods for function approximation such as multi-layer perceptrons.

Claims

1. A method for facilitating voice recognition using a modeling relationship between first and second data collections, wherein for each member of the first data collection, there is a corresponding member of the second collection, comprising:

determining residuals between the first and second collections;

determining, using the residuals, a position of one of the members of the first data collection;

determining proximity data relative to a value (V) for the one member of the first data collection at the position;

determining a subcollection of the first collection, wherein each member of the subcollection has a value that is in proximity to the value V according to the proximity data;

generating a basis function by a computer from the subcollection for obtaining a model of the relationship; and

outputting model information for presentation to a user or a predetermined process for affecting or identifying a physical event, wherein the model information includes at least one of;

(a) data indicative of a correspondence between the model and the relationship, (b) data indicative of a variance between the model and the relationship, (c) an extrapolation of the relationship, (d) an interpolation of the relationship, (e) multi-dimensional model output, and (f) notification of the physical event and wherein an approximation function is modulated, or skewed, by a shape function; and

facilitating voice recognition by using the modeling relationship between the first and second data collections.

2. The method of claim 1, wherein the approximation function is selected form the group consisting of basis functions, multi-layer perceptrons, and feed-forward neural networks.

3. The method of claim 1, wherein the steps of claim 1 are iteratively performed for generating a model of the relationship, and with at least most iterations, a step of determining structure or information content in the residuals.

4. The method of claim 3, wherein for each of at least most of the iterations, the corresponding instance of the subcollection is determined accordingly to an instance of the position determined as a function of a plurality of autocorrelation components.

5. The method of claim 3, wherein the second collection has a dimensionality greater than one, and the step of determining the position includes, for each of at least most of the iterations, performing a test on the residuals for determining whether or not the steps of claim 1 are to be performed again.

6. The method of claim 5, wherein the test includes using the residuals to determine an error that is a function of more than one of the dimensions of the second collection.

7. The method of claim 1, wherein the step of determining the position includes performing an autocorrelation for independent identically and distributed in the residuals.

8. The method of claim 1, wherein the approximation function is a basis function, wherein the step of generating includes determining at least one parameter value for the basis function by iteratively adjusting the parameter value for reducing a result of a function dependent upon a difference between an instance of the basis function, and members of the second data collection that correspond to the members of the subcollection.

9. The method of claim 1, wherein the approximation function is a basis function, wherein the basis function is generated using a mollifier, circle or bump function.

10. The method of claim 1 further including a step of testing the model on additional data indicative of the relationship, wherein the additional data includes at least one of a first member related to a second member according to the relationship, and wherein at least one of:

the first member is not a member of the first data collection, and the second member is not a member of the second data collection.

11. The method of claim 1, wherein the physical event includes one of:

an airflow around a physical object, a water flow over an object, a radio signal, and recognition of an object, maneuvering of an object.

12. The method of claim 1, wherein a confidence level for a plurality of functions of the residuals is determined at least one of simultaneously and sequentially.

13. The method of claim 1, wherein the data collections are high-dimensional data collections.

14. The method of claim 1, wherein the data collections are spatial and without time parameterization.

15. The method of claim 1, wherein the approximation function is selected from the group consisting of radial basis functions and modulated asymmetric radial basis functions.

16. The method of claim 1, wherein at least one of the first data collection and second data collection is substantially noise free.

17. The method of claim 1, wherein the approximation function is a non-symmetrical radial basis function with at least one of compact support and non-compact support.

18. A method for failure prediction using a modeling relationship between first and second data collections, wherein for each member of the first data collection, there is a corresponding member of the second collection, comprising:

determining residuals between the first and second collections;

determining, using the residuals, a position of one of the members of the first data collection;

determining proximity data relative to a value (V) for the one member of the first data collection at the position;

determining a subcollection of the first collection, wherein each member of the subcollection has a value that is in proximity to the value V according to the proximity data;

generating a basis function by a computer from the subcollection for obtaining a model of the relationship; and

outputting model information for presentation to a user or a predetermined process for affecting or identifying a physical event, wherein the model information includes at least one of;

(a) data indicative of a correspondence between the model and the relationship, (b) data indicative of a variance between the model and the relationship, (c) an extrapolation of the relationship, (d) an interpolation of the relationship, (e) multi-dimensional model output, and (f) notification of the physical event and wherein an approximation function is modulated, or skewed, by a shape function; and

facilitating failure prediction by using the modeling relationship between the first and second data collections.

19. A method for image processing using a modeling relationship between first and second data collections, wherein for each member of the first data collection, there is a corresponding member of the second collection, comprising:

determining residuals between the first and second collections;

determining, using the residuals, a position of one of the members of the first data collection;

determining proximity data relative to a value (V) for the one member of the first data collection at the position;

determining a subcollection of the first collection, wherein each member of the subcollection has a value that is in proximity to the value V according to the proximity data;

generating a basis function by a computer from the subcollection for obtaining a model of the relationship; and

outputting model information for presentation to a user or a predetermined process for affecting or identifying a physical event, wherein the model information includes at least one of;

(a) data indicative of a correspondence between the model and the relationship, (b) data indicative of a variance between the model and the relationship, (c) an extrapolation of the relationship, (d) an interpolation of the relationship, (e) multi-dimensional model output, and (f) notification of the physical event and wherein an approximation function is modulated, or skewed, by a shape function; and

facilitating image processing by using the modeling relationship between the first and second data collections.

20. A method for financial time series analysis using a modeling relationship between first and second data collections, wherein for each member of the first data collection, there is a corresponding member of the second collection, comprising:

determining residuals between the first and second collections;

determining, using the residuals, a position of one of the members of the first data collection;

determining proximity data relative to a value (V) for the one member of the first data collection at the position;

determining a subcollection of the first collection, wherein each member of the subcollection has a value that is in proximity to the value V according to the proximity data;

generating a basis function by a computer from the subcollection for obtaining a model of the relationship; and

outputting model information for presentation to a user or a predetermined process for affecting or identifying a physical event, wherein the model information includes at least one of;

(a) data indicative of a correspondence between the model and the relationship, (b) data indicative of a variance between the model and the relationship, (c) an extrapolation of the relationship, (d) an interpolation of the relationship, (e) multi-dimensional model output, and (f) notification of the physical event and wherein an approximation function is modulated, or skewed, by a shape function; and

facilitating financial time series analysis by using the modeling relationship between the first and second data collections.