The perceptron stops at the first separating hyperplane it finds — but for separable data, there are usually infinitely many. Session 9's margin bound suggests the boundary with the biggest margin should generalize best. That's SVM's entire idea in one sentence.