A classifier’s answer is not the whole story: margin-based reasoning considers how strongly the evidence separates the correct class from its competitors. In an SVM, the margin is geometric distance from the decision boundary; in boosting, it reflects the correct class’s vote relative to competing votes. section 8.8.3
Boosting methods such as AdaBoost can keep increasing margins after training accuracy reaches 100%, offering one explanation for continued improvement on new data. Related ideas appear in representation learning and deep networks, while gradient descent can approach a maximum-margin solution on separable data under specific assumptions. These connections do not make deep networks descendants of SVMs or establish margin theory as a complete explanation of their behavior. section 8.8.3
