Skip to content

dorfmanrobert/ICL-Final-Project

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

19 Commits
 
 
 
 
 
 
 
 

Repository files navigation

ICL-Final-Project

This is the code used to run experiments for my final project Learning with Guarantees. Three approaches are compared: SA compression, PAC-Bayes, and test set bounds. The code used to compute the bound in SA Compression is based on MATLAB code from https://marco-campi.unibs.it/pdf-pszip/Wait-and-judge.PDF, using a bisectional numerical algorithm to solve the equation required for the bound.

We build on top of existing code in implementing the PAC-Bayesian and test set methods. For PAC-Bayes we most heavily rely on the existing code at https://github.com/mperezortiz/PBB. Our most important modifications to this code base are

  1. incorporating the code enabling optimization of inverted binary kl from https://github.com/eclerico/CondGauss

  2. incorporating code enabling optimization through the CondGauss approach from https://github.com/eclerico/CondGauss

  3. building in the ability to run training with the averaged posterior approach of https://arxiv.org/pdf/1905.13367.pdf

For the test set bounds we adapt implementations from https://github.com/cambridge-mlg/pac-bayes-tightness-small-data. For each approach, all experiments for can be run using the runexp.py file (with desired arguments) contained in that approach's folder.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages