Relaxation of hard classification targets for LSE minimization

Kar Ann Toh, Xudong Jiang, Wei Yun Yau

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

In the spirit of stabilizing a solution to handle possible over-fitting of data which is especially common for high order models, we propose a relaxed target training method for regression models which are linear in parameters. This relaxation of training target from the conventional binary values to disjoint classification spaces provides good classification fidelity according to a threshold treatment during the decision process. A particular design to relax the training target is provided under practical consideration. Extension to multiple class problems is formulated before the method is applied to a plug-in full multivariate polynomial model and a reduced model on synthetic data sets to illustrate the idea. Additional experiments were performed using real-world data from the UCI[1] data repository to derive certain empirical evidence.

Original languageEnglish
Title of host publicationEnergy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings
Pages187-202
Number of pages16
DOIs
Publication statusPublished - 2005 Dec 1
Event5th International Workshop on Energy Minimization Methods in Computer Vision and Pattern Recognition, EMMCVPR 2005 - St. Augustine, FL, United States
Duration: 2005 Nov 92005 Nov 11

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume3757 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Other

Other5th International Workshop on Energy Minimization Methods in Computer Vision and Pattern Recognition, EMMCVPR 2005
CountryUnited States
CitySt. Augustine, FL
Period05/11/905/11/11

Fingerprint

Target
Polynomial Model
Multivariate Polynomials
Overfitting
Reduced Model
Multivariate Models
Plug-in
Synthetic Data
Repository
Fidelity
Regression Model
Disjoint
Higher Order
Binary
Experiment
Training
Experiments
Model
Class
Design

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • Computer Science(all)

Cite this

Toh, K. A., Jiang, X., & Yau, W. Y. (2005). Relaxation of hard classification targets for LSE minimization. In Energy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings (pp. 187-202). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 3757 LNCS). https://doi.org/10.1007/11585978_13
Toh, Kar Ann ; Jiang, Xudong ; Yau, Wei Yun. / Relaxation of hard classification targets for LSE minimization. Energy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings. 2005. pp. 187-202 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
@inproceedings{5c41ccd72ece42daad18f943f977885a,
title = "Relaxation of hard classification targets for LSE minimization",
abstract = "In the spirit of stabilizing a solution to handle possible over-fitting of data which is especially common for high order models, we propose a relaxed target training method for regression models which are linear in parameters. This relaxation of training target from the conventional binary values to disjoint classification spaces provides good classification fidelity according to a threshold treatment during the decision process. A particular design to relax the training target is provided under practical consideration. Extension to multiple class problems is formulated before the method is applied to a plug-in full multivariate polynomial model and a reduced model on synthetic data sets to illustrate the idea. Additional experiments were performed using real-world data from the UCI[1] data repository to derive certain empirical evidence.",
author = "Toh, {Kar Ann} and Xudong Jiang and Yau, {Wei Yun}",
year = "2005",
month = "12",
day = "1",
doi = "10.1007/11585978_13",
language = "English",
isbn = "3540302875",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
pages = "187--202",
booktitle = "Energy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings",

}

Toh, KA, Jiang, X & Yau, WY 2005, Relaxation of hard classification targets for LSE minimization. in Energy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 3757 LNCS, pp. 187-202, 5th International Workshop on Energy Minimization Methods in Computer Vision and Pattern Recognition, EMMCVPR 2005, St. Augustine, FL, United States, 05/11/9. https://doi.org/10.1007/11585978_13

Relaxation of hard classification targets for LSE minimization. / Toh, Kar Ann; Jiang, Xudong; Yau, Wei Yun.

Energy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings. 2005. p. 187-202 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 3757 LNCS).

Research output: Chapter in Book/Report/Conference proceedingConference contribution

TY - GEN

T1 - Relaxation of hard classification targets for LSE minimization

AU - Toh, Kar Ann

AU - Jiang, Xudong

AU - Yau, Wei Yun

PY - 2005/12/1

Y1 - 2005/12/1

N2 - In the spirit of stabilizing a solution to handle possible over-fitting of data which is especially common for high order models, we propose a relaxed target training method for regression models which are linear in parameters. This relaxation of training target from the conventional binary values to disjoint classification spaces provides good classification fidelity according to a threshold treatment during the decision process. A particular design to relax the training target is provided under practical consideration. Extension to multiple class problems is formulated before the method is applied to a plug-in full multivariate polynomial model and a reduced model on synthetic data sets to illustrate the idea. Additional experiments were performed using real-world data from the UCI[1] data repository to derive certain empirical evidence.

AB - In the spirit of stabilizing a solution to handle possible over-fitting of data which is especially common for high order models, we propose a relaxed target training method for regression models which are linear in parameters. This relaxation of training target from the conventional binary values to disjoint classification spaces provides good classification fidelity according to a threshold treatment during the decision process. A particular design to relax the training target is provided under practical consideration. Extension to multiple class problems is formulated before the method is applied to a plug-in full multivariate polynomial model and a reduced model on synthetic data sets to illustrate the idea. Additional experiments were performed using real-world data from the UCI[1] data repository to derive certain empirical evidence.

UR - http://www.scopus.com/inward/record.url?scp=33646540426&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=33646540426&partnerID=8YFLogxK

U2 - 10.1007/11585978_13

DO - 10.1007/11585978_13

M3 - Conference contribution

AN - SCOPUS:33646540426

SN - 3540302875

SN - 9783540302872

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 187

EP - 202

BT - Energy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings

ER -

Toh KA, Jiang X, Yau WY. Relaxation of hard classification targets for LSE minimization. In Energy Minimization Methods in Computer Vision and Pattern Recognition - 5th International Workshop, EMMCVPR 2005, Proceedings. 2005. p. 187-202. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). https://doi.org/10.1007/11585978_13