Abstract: 

Although deep learning may often be thought of as "just another tool," it has a number of properties that make it particularly well suited for scientific applications.  I will address common myths about deep learning, such as deep learning being "a black box" and it requiring big data. In fact, deep learning is a natural choice for data-driven analysis of physical systems. I will show how we have used unique properties of deep learning to advance the state of the art in dynamical systems, control theory, chemistry, and other fields. In addition, I will demonstrate that deep learning provides natural interpretable visualizations, and these visualizations show neural networks learn similar features that scientists use to understand systems.  

Bio: 

Dr. Nathan Hodas is a Senior Research Scientist in the Data Science and Analytics group at Pacific Northwest National Lab. He is currently helps to lead the Deep Learning for Scientific Discovery Agile Investment at PNNL, a laboratory-level investment in using deep learning to advance the frontiers of science. Before arriving at PNNL in Spring 2014, he was a post-doc under Kristina Lerman at Information Sciences Institute at USC, studying information propagation on social networks and developing human response dynamics to predict users’ behavior online. He attended the Santa Fe Institute’s Complex Systems Summer School in 2009. At SFI, he transitioned his research from purely statistical physics toward studying complex systems and machine learning. His graduate work at Caltech involved studying nonlinear dynamics of nano-scale systems, conducting nonlinear optical biological experiments, developing advanced theories of nanoscale protein dynamics, and modeling surface chemical reactions under Professor Rudy Marcus and Professor Scott Fraser.