Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Videos

Study of a fast discriminative training algorithm for pattern recognition.

Qi Li1, Biing-Hwang Juang

  • 1Bell Labs, Lucent Technologies, USA. qili@ieee.org

IEEE Transactions on Neural Networks
|September 28, 2006
PubMed
Summary

This study introduces a fast discriminative training algorithm for pattern recognition, improving speech application accuracy. The novel method converges faster than traditional gradient-descent and expectation-maximization algorithms.

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Innate immunity alone is not sufficient for chronic rejection but predisposes healed allografts to T cell-mediated pathology.

Transplant immunology·2012
Same author

Electroacupuncture improves survival in rats with lethal endotoxemia via the autonomic nervous system.

Anesthesiology·2012
Same author

Single-incision laparoscopic Roux-en-Y hepaticojejunostomy using conventional instruments for children with choledochal cysts.

Surgical endoscopy·2011
Same author

Coherent spin precession via photoinduced antiferromagnetic interactions in La0.67Ca0.33MnO3.

Physical review letters·2011
Same author

Mapping the central effects of chronic ketamine administration in an adolescent primate model by functional magnetic resonance imaging (fMRI).

Neurotoxicology·2011
Same author

Role for Class I histone deacetylases in multidrug resistance.

Experimental cell research·2011

Area of Science:

  • Pattern Recognition
  • Machine Learning
  • Speech Processing

Background:

  • Discriminative training directly minimizes a cost function, differing from traditional probability estimation in Bayes' formulation.
  • Current gradient-descent (GD) methods for discriminative training are slow and struggle with learning rate selection.

Purpose of the Study:

  • To develop a faster discriminative training algorithm for nonlinear classifiers.
  • To improve recognition accuracy and reduce training iterations compared to existing methods.

Main Methods:

  • Initialization of parameters using the expectation-maximization (EM) algorithm.
  • Optimization using novel closed-form formulas to minimize error rate.
  • Application and testing in speech recognition tasks.

Related Experiment Videos

Main Results:

  • The proposed algorithm achieves better recognition accuracy than EM and GD-trained neural networks.
  • The algorithm requires fewer iterations for convergence compared to conventional methods.
  • Demonstrated effectiveness in speech applications.

Conclusions:

  • The developed fast discriminative training algorithm offers improved performance and efficiency.
  • The proposed objective and formulas provide a foundation for future research in discriminative training.
  • Further investigation into the algorithm's convergence properties is warranted.