https://github.com/pointlander/beeline

Beeline gradient descent
https://github.com/pointlander/beeline

Last synced: 4 months ago
JSON representation

Beeline gradient descent

• Host: GitHub
• URL: https://github.com/pointlander/beeline
• Owner: pointlander
• License: bsd-3-clause
• Created: 2019-06-16T00:59:50.000Z (about 7 years ago)
• Default Branch: master
• Last Pushed: 2019-06-23T20:11:20.000Z (almost 7 years ago)
• Last Synced: 2024-06-20T12:02:48.511Z (almost 2 years ago)
• Language: Go
• Homepage:
• Size: 49.8 KB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Beeline gradient descent

Gradient descent with shared weights.

The number of shared weights is increased with each epoch, so the dimensionality of the search space is increased with each epoch.

Below is a base cost curve for the iris data set:

![epochs of normal gradient descent](epochs_normal.png?raw=true)

Cost curve for beeline gradient descent:

![epochs of beeline gradient descent](epochs_shared.png?raw=true)

As you can see beeline gradient descent still converges, but doesn't work as well as the normal approach.