Chapter 2 · Machine learning basics
Gradient descent
Gradient descent walks downhill on the loss surface, step by step, to find better parameters.
Preparing gradient descent
Chapter 2 · Machine learning basics
Gradient descent walks downhill on the loss surface, step by step, to find better parameters.
The curve is the loss. The ball is the model's current setting. It looks at the slope under it and steps the other way.
Training begins with a guess for the parameters, often random. That is the ball's starting spot.
Choose where the ball starts and the step size, then press Step or Run. The numbers are the real gradient steps.
Bottom of the bowl is at about 1.89. Watch the steps shrink as the ground flattens.
Which way does gradient descent move?
Near the bottom of the bowl, the steps get…
What does the loss do as the ball descends?