Unlike Naive Bayes, the perceptron learns a boundary directly — no model of how data was generated.

Model: for labels y ∈ {−1,+1},

ŷ = sign(wᵀx + b)

w defines a hyperplane; points on each side get different labels.

Assumption: data is linearly separable (some hyperplane splits it perfectly). What happens when it isn't motivates soft-margin SVM (Session 11).