Using second-order information in machine learning

This thesis proposes second-order optimization algorithms for support vector machines (SVM) and autoencoders. In the first part of this thesis, we propose a smoothing algorithm for solving the soft-margin SVM optimization problem with an $\ell^1$\,penalty. In the second part, we present a stochastic Hessian-free optimization method for training autoencoders by reducing the amount of data used in training.