{1416} revision 0 modified: 10-31-2018 23:49 gmt

Learning data manifolds with a Cutting Plane method

  • Looks approximately like SVM: perform binary classification on a high-dimensional manifold (or sets of manifolds in this case).
  • The general idea behind Mcp_simple is to start with a finite number of training examples, find the maximum margin solution for that training set, augment the draining set by finiding a poing on the manifolds that violates the constraints, iterating the process until a tolerance criteria is met.
  • The more complicated cutting plane SVM uses slack variables to allow solution where classification is not linearly separable.
    • Propose using one slack variable per manifold, plus a manifold center, which strictly obeys the margin (classification) constraint.
  • Much effort put to proving the convergence properties of these algorithms; admittedly I couldn't be bothered to read...