4 / 7 Logistic Regression

Punish the Confident Wrong

Why is binary cross-entropy loss used to train logistic regression instead of mean squared error on the predicted probabilities?

A

Cross-entropy loss eliminates the need to compute gradients

B

Combining MSE with a sigmoid output produces a non-convex loss surface, while cross-entropy stays convex in the weights, so gradient descent reliably reaches the global minimum

C

Cross-entropy is simply faster to compute on modern hardware

D

MSE cannot mathematically be applied to values between 0 and 1