• at a certain w and b value, take a 360 turn to look for the steepest descent.
• following the steepest descent after descent takes to the lowest valleys.
• there could be multiples of valleys: known as local minima.

Gradient Descent Formula

• alpha: learning rate, a number from 0 to 1.
• updates w and b value little by little

Deriving Linear Regression Cost Function Derivatives

• Squared error cost for linear regression is always a convex function.
  • it always has one global minimum.

• Looks at all training examples:
  • Each gradient descent step computes J(w,b) value, which is an average of sum of squared errors. This is calcuated by using every training examples by formula.