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

Bias reduction in skewed binary classification with Bayesian neural networks.

P J Lisboa1, A Vellido, H Wong

  • 1School of Computing and Mathematical Sciences, Liverpool John Moores University, UK.

Neural Networks : the Official Journal of the International Neural Network Society
|August 18, 2000
PubMed
Summary
This summary is machine-generated.

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

Listening to and learning from patients: A review of the results from the 2023 national radiotherapy patient experience survey in England.

Radiography (London, England : 1995)·2026
Same author

Psychometric validation of comprehensive coordination scale to assess the advanced coordination ability in people with stroke.

European journal of physical and rehabilitation medicine·2025
Same author

Primary hyperparathyroidism in a domestic shorthair cat following I<sup>131</sup> radioiodine therapy.

The Journal of small animal practice·2025
Same author

Methylation Mesa define functional regulatory elements for targeted gene activation.

Research square·2024
Same author

Early onset metastatic colorectal cancer in Australia.

Cancer treatment and research communications·2024
Same author

Corrigendum to "A comparison of the efficacy of trastuzumab deruxtecan in advanced HER2-positive breast cancer: active brain metastasis versus progressive extracranial disease alone": [ESMO Open 8 (2023) 102033].

ESMO open·2024

Marginalizing neural network outputs to the prior distribution improves calibration. Conversely, marginalizing to the midrange can bias conditional probability estimates, particularly in censored data modeling.

Area of Science:

  • Machine Learning
  • Bayesian Inference
  • Neural Networks

Background:

  • The Bayesian evidence framework is standard for neural network estimation of class conditional probabilities.
  • This framework typically marginalizes over network weights using approximations that moderate output towards the midrange.

Purpose of the Study:

  • To investigate the impact of different marginalization strategies on neural network calibration.
  • To assess the potential biases introduced by midrange marginalization in probability estimation.

Main Methods:

  • Compared marginalization to the prior distribution versus marginalization to the midrange for neural network weight distributions.
  • Evaluated network calibration and probability estimates under both marginalization approaches.
  • Focused analysis on the specific challenges posed by modeling censored data.

Related Experiment Videos

Main Results:

  • Marginalizing to the prior distribution significantly improves neural network calibration.
  • Marginalization to the midrange introduces considerable bias into conditional probability estimates.
  • The bias from midrange marginalization is particularly pronounced when modeling censored data.

Conclusions:

  • Marginalizing neural network outputs to the prior distribution is a superior approach for accurate probability estimation.
  • Midrange marginalization should be avoided due to its potential to seriously bias results, especially with censored data.
  • The findings highlight the importance of appropriate marginalization techniques within the Bayesian evidence framework for neural networks.