This animation shows why a kernel function must be more than a convenient similarity measure. Inner and outer rings that resist linear separation are lifted into three dimensions using a simple feature map, making a separating plane possible. The video then contrasts linear, polynomial, and RBF kernels, and uses the matrix [[1,2],[2,1]] to show that symmetry alone does not guarantee a valid Gram matrix, since its eigenvalues include a negative value. Useful for first-year AI students learning the formal requirements behind the kernel trick.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# A kernel must be more than a similarity score Begin with inner and outer rings that a straight line cannot separate. Lift points using the illustrative feature map (x1,x2,x1 squared + x2 squared), allowing a plane to separate the rings. Explain that this is a concrete teaching map, not the full feature map of every kernel. Introduce K(x,z) as an inner product in a feature space, then compare linear, polynomial, and Gaussian RBF formulas with their parameter conditions. Build the symmetric matrix [[1,2],[2,1]] and reveal its eigenvalues 3 and -1: symmetry alone does not make a valid Gram matrix. For a=(1,-1), its quadratic form is -2. Explain that a valid kernel must produce positive-semidefinite Gram matrices for every finite input set; one successful finite check is not a universal proof. Use black backgrounds, cyan geometry, orange counterexamples, and short English labels. Useful for understanding the kernel trick and its mathematical constraints. Audience: 一年级人工智能专业学生