Skip to content

Code for the ICCV 2019 paper "Sampling-free Epistemic Uncertainty Estimation Using Approximated Variance Propagation"

Notifications You must be signed in to change notification settings

janisgp/Sampling-free-Epistemic-Uncertainty

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

30 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Sampling-free Epistemic Uncertainty Estimation Using Approximated Variance Propagation

This repository provides the code for the ICCV'19 publication "Sampling-free Epistemic Uncertainty Estimation Using Approximated Variance Propagation". We provide a sampling-free approach for estimating epistemic uncertainty when applying methods based on noise injection (e.g. stochastic regularization). Our approach is motivated by error propagation. We primarily compare our approach with Monte-Carlo (MC) dropout by approximating the sampling procdeure of the latter.

Following the experiment section in our paper, this repository is divided into three sections:

  • UCI_regression: Comparison of predictive performance between MC dropout and our variance propagation approach. The code further implements Mixture Density Networks and learning the dropout parameter using our sampling-free approximation, which is not included in our publication.
  • Bayesian_SegNet: Sampling-free approximation of Bayesian SegNet. We investigate the quality of our approximation on a high-dimensional semantic segmentation task.
  • Monocular_Depth_Regression: Applies our approximation to Unsupervised Monocular Depth Regression. We enhance the original work by inserting dropout at the final layers and analyze the quality of our approximation.

About

Code for the ICCV 2019 paper "Sampling-free Epistemic Uncertainty Estimation Using Approximated Variance Propagation"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published