An SVM is purely a linear model from the right perspective, and if you're being really reductive, RELU neural networks are piecewise linear. I think this may be

An SVM is purely a linear model from the right perspective, and if you're being really reductive, RELU neural networks are piecewise linear. I think this may be obscuring more than it helps; picking the right transformation for your particular case is a highly nontrivial problem; why sin(x) and x^2, rather than, say, tanh(x) and x^(1/2).