Skip to content

wgopar/Coordinate-Descent

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 

Repository files navigation

Coordinate Descent Experiment

This notebook showcases two different approaches to implementing coordinate descent on model defined with cross entropy loss. Coordinate descent does have problems if implemented on non-smooth multivariate functions as the algorithm may get stuck at a non-stationary points. This is because the level curves of the function are not smooth. Besides this, the algorithm does have increased interest in large-scale problems where it has been shown to be competitive to other algorithms when training support vector machines and performing non-negative matrix factorization.

In this experiment I implement coordinate descent using the following approaches:

  • Randomly optimize along a random coordinate.
  • Optimize along coordinate with the largest gradient in magnitude

To benchmark the solutions I implemented a standard logistic regression model with out focusing on validation and was just used to obtain Scikit-Learns implementation of gradient descent and final loss. The model is trained with the wine.data dataset provided by UCI machine learning repository.

About

Notebook of experiments and comparison of two different approaches to Coordinate Descent

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors