Feature Scaling and Gradient Descent

Click anywhere on the loss landscape to choose the starting point.

Features

Model

y^ = w1x1 + w2x2
Loss = MSE(y^, y)

Gradient descent

Learning rate 0.00001

Dataset

80 observations

x1 is roughly 1�10.
x2 is roughly 100�1000.

The target is generated from a linear model with noise.

The data does not change when switching between the two views.

Click anywhere to choose starting point

Loss

Mean squared error

Iteration

Gradient descent steps

Gradient norm

Distance from stationary point