Scalable Deep Learning

Abstract: Although deep learning has received much acclaim due to its widespread empirical success, fundamental bottlenecks exist when attempting to develop deep learning applications at scale. One bottleneck involves exploring the design space of a model family, which typically requires training tens to thousands of models with different hyperparameters. Model training itself is a second major bottleneck, as classical learning algorithms are often infeasible for the petabyte datasets that are fast becoming the norm. In this talk, I present my research addressing these two core bottlenecks. I first introduce Hyperband, a novel algorithm for hyperparameter optimization that is simple, flexible, theoretically sound, and an order of magnitude faster than leading competitors. I then present work aimed at understanding the underlying landscape of training deep learning models in parallel and distributed environments. Specifically, I introduce an analytical performance model called Paleo, which can quickly and accurately model the expected scalability and performance of putative parallel and distributed deep learning systems.

Bio: Ameet Talwalkar is co-founder and Chief Scientist at Determined AI. Starting in January, he will also be an assistant professor in the School of Computer Science at Carnegie Mellon University. His interests include problems related to scalability and ease-of-use in the field of statistical machine learning, with applications in computational genomics. He led the initial development of the MLlib project in Apache Spark, is a co-author of the graduate-level textbook 'Foundations of Machine Learning' (2012, MIT Press), and teaches an award-winning MOOC on edX called 'Distributed Machine Learning with Apache Spark.'