Gradient Descent Feature Scaling Lab

The source assigns curvature coefficients sixty eight point four and one in raw mode, four point two and one in min-max mode, and one and one in z-score mode. Loss is one half times the weighted sum of the squared coordinates. These curvature fixtures are selected by mode; they are not computed from the displayed dataset’s Hessian. The bars compare that explicitly chosen stiffness ratio. For raw curvature sixty eight point four, a gradient descent update multiplies the first coordinate by one minus learning rate times sixty eight point four. At learning rate zero point zero five the factor is minus two point four two: a coordinate one becomes minus two point four two and its magnitude grows. With z-score fixture curvature one, the factor is zero point nine five and the magnitude shrinks. This explains the fixture’s contrast without claiming a measured training result. The actual dataset transformation subtracts the minimum and divides by the range for min-max, or subtracts the mean and divides by population standard deviation for z-score. In the illustrative values ten twenty thirty, min-max becomes zero one half one. The mean is twenty and the population standard deviation square root two hundred thirds, so z-scores are about minus one point two two five, zero and plus one point two two five. Exported scaled data and optimizer traces preserve these distinct calculations.