Demo for Bi-Tempered Logistic Loss for Training Neural Nets with Noisy Data.

Demonstration of practical properties of "Robust Bi-Tempered Logistic Loss Based on Bregman Divergences", 2019,
Ehsan Amid, Manfred K. Warmuth, Rohan Anil, Tomer Koren. Read manuscript.

This trains a 2-layer neural network of size [10, 5] and visualizes the decision boundary above.
Colors indicate the class label.
You can drag & drop points on the canvas to see how the model handles noisy data points.
You can also auto-generate noise levels using the options on the right.
Please refresh the page when something breaks :)

Select noise type:

No Noise Small Margin Noise Large Margin Noise Random Noise

Noise Level = 10%

Decrease Temperature 1 to handle large margin noise.
Increase Temperature 2 to handle small margin noise.
Adjust Temperature 1 & 2 to handle random noise.

Adjust temperature of loss function:

Temperature 1 = 1.0

Temperature 2 = 1.0

Warmstart model parameters when data/temperature settings change?

Yes, reset the model parameters at the start of training. No, continue training from the current model.