Lesson 24 — Logistic regression training

Train a classifier from scratch

Now combine everything: sigmoid, a score w × x + b, and gradient descent. For logistic regression the gradient is beautifully simple: error = p − y (probability minus true label), then w moves by error × x and b by error — the same shape as linear regression.

p = sigmoid(w * x + b)
error = p - y

After training, check accuracy on the data.

Why it matters: you have just built, by hand, the model behind countless spam filters, credit scores and click predictors.