1: \begin{abstract}

2: \textcolor{black}{We consider the problem of classifying data manifolds

3: where each manifold represents invariances that are parameterized

4: by continuous degrees of freedom. Conventional data augmentation methods

5: rely upon sampling large numbers of training examples from these manifolds;

6: instead, we propose an iterative algorithm called $M_{CP}$ based

7: upon a cutting-plane approach that efficiently solves a quadratic

8: semi-infinite programming problem to find the maximum margin solution.

9: We provide a proof of convergence as well as a polynomial bound on

10: the number of iterations required for a desired tolerance in the objective

11: function. The efficiency and performance of $M_{CP}$ are demonstrated

12: in high-dimensional simulations and on image manifolds generated from

13: the ImageNet dataset. Our results indicate that $M_{CP}$ is able

14: to rapidly learn good classifiers and shows superior generalization

15: performance compared with conventional maximum margin methods using

16: data augmentation methods.}

17: \end{abstract}

18: