Maren Mahsereci will present her work on efficient step-size selection for deep learning in front of the 3000-strong plenum at Neural Information Processing Systems in Montréal this week. NIPS is the flagship conference of machine learning, and full orals such as this are only awarded to a handful of papers, less than 1% of the submitted works, each year.
Maren will talk about her work, with Philipp Hennig, constructing a probabilistic version of classic line search algorithms (full paper on the proceedings page). Line search methods are basic tools in nonlinear optimization; but so far they were not applicable to the large-scale, stochastic problems encountered in big data and deep learning. The new probabilistic line search is robust to these settings, and provides an efficient, computationally lightweight way to remove the need for practitioners to choose a specific ``learning rate'' when training, for example, a deep neural network.