Draw a line through the plane and call one side class 1. Logistic regression softens that into a probability: p = σ(w·x + b) rises smoothly from 0 to 1 across the line. Gradient descent on the cross-entropy turns and slides the line until it separates the classes as well as a line can.