This animation contrasts a supervised SVM, trained only on a handful of labeled points, with a semi-supervised version that also uses many unlabeled points to find a low-density decision boundary. It visualizes labeled and unlabeled hinge-loss terms, explains why unknown labels make the optimization nonconvex, and shows class-balance constraints and failure cases from misleading structure. It distinguishes inductive from transductive evaluation and introduces must-link/cannot-link constraints, ending with a practical guide for choosing a learning paradigm. Useful for first-year AI students comparing supervised and semi-supervised methods.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# Semi-supervised learning depends on assumptions, not just more data A small set of labeled points determines a supervised SVM boundary, while many gray unlabeled points reveal two dense regions. Show how a semi-supervised SVM may move the boundary toward a low-density gap by jointly considering model parameters and candidate labels for the gray points. Display labeled and unlabeled hinge-loss terms with separate weights, and explain why unknown labels make the optimization generally nonconvex. Keep candidate labels visually distinct from known truth. Introduce class-balance constraints and explain that misleading unlabeled structure can hurt performance. Contrast inductive evaluation with a transductive setting that can use target inputs but never their hidden evaluation labels. Briefly show must-link and cannot-link constraints as a different form of partial supervision for clustering. Finish with a decision guide based on label availability and assumptions. Use black backgrounds, white unlabeled points, and restrained cyan-orange class colors. Useful for choosing learning paradigms responsibly. Audience: 一年级人工智能专业学生