Structured Prediction via Deep Learning for Inverse Problems

Eric Jonas, University of Chicago

Inverse problems are a well-studied branch of applied mathematics where we attempt to invert a well-understood forward model in an attempt to learn the properties of some hidden source. Classical examples include seismography in geoscience and tomography in medical imaging, where boundary-value measurements are used to infer properties of the interior. Many scientific problems can be cast as inverse problems with structured outputs -- graphs, discrete point sources, and more. I will present some of our recent work developing machine learning techniques for structured prediction to inverse problems in neuroscience, computational microscopy, and nuclear spectroscopy. By leveraging advances in scalable compute, deep models, and function approximation, we can solve these problems faster, more accurately, and even compensate for more complicated physical generative processes.

Bio: Eric Jonas is a new Assistant Professor in Computer Science at the University of Chicago, working on developing machine learning techniques to accelerate and improve scientific measurement. He earned his PhD in Computational Neuroscience, M. Eng in Electrical Engineering, BS in Electrical Engineering and Computer Science, and BS in Neurobiology, all from MIT. Prior to his return to academia, he was founder and CEO of Prior Knowledge, a predictive database company which was acquired in 2012 by, where he was Chief Predictive Scientist until 2014. In 2015 he was named one of the top rising stars in bioengineering by the Defense Department’s Advanced Research Projects Agency (DARPA).