1: \begin{abstract}

2:   We describe and analyze a new boosting algorithm for deep learning

3:   called SelfieBoost. Unlike other boosting algorithms, like AdaBoost,

4:   which construct ensembles of classifiers, SelfieBoost boosts the

5:   accuracy of a single network. We prove a $\log(1/\epsilon)$

6:   convergence rate for SelfieBoost under some ``SGD success''

7:   assumption which seems to hold in practice.

8: \end{abstract}

9: