Skip to content

Visualization of different gradient descent methods used in deep learning frameworks

License

Notifications You must be signed in to change notification settings

christianwaldmann/optimizer-visualization

Repository files navigation

Visualization of deep learning optimizers

This simulation visualizes gradient descent methods in different situations. The main motivation behind that is to better understand optimizers which are being used by deep learning frameworks.

Optimizers

All optimizers use the gradient for calculating their next step. In this simulation the gradient is calculated numerically (symmetric derivative).

The optimizers use the following formulas to calculate their next step:

Scenes

Local minimum

3D View

Top-down view

Plateau

3D View

Top-down view

Downwards parabola

3D View

Top-down view

Hills

3D View

Top-down view

Saddle point

3D View

Top-down view

5th degree polynom

3D View

Top-down view

About

Visualization of different gradient descent methods used in deep learning frameworks

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages