Fwd: "Automatic self-configurable machine learning to enable `small-data' science for non-experts" job talk Thursday noon

Please join us for our third tenure-track faculty job talk Thursday noon, the position focus is on Data Science in the Department of Computer Science.

Nick Cheney

University of Vermont

Thurs Feb 27, 12 - 1pm, John Dewey Lounge (325 Old Mill)

Sandwiches and drinks to be served.

Automatic self-configurable machine learning 

to enable "small-data" science for non-experts

The dramatic rise and popularity of machine learning, and especially deep learning, has accelerated automation, innovation, and scale in the many industries that seek to monetize "big data".  In other domains, such as in academia, the adoption of deep learning has been tempered, limited largely by the scarcity of experienced data scientists and the lack of large labeled datasets. I conjecture that the development of machine learning methods that are increasingly robust and self-configurable will help to remove the hard-earned machine-learning expertise and intuition that is currently required to configure deep learning architectures, hyperparameters, and data-processing pipelines – helping to open the use of these techniques to non-machine-learning-experts.  These automatic machine learning (AutoML) pipelines often rely on meta-learning – which employs a machine learning process to automatically configure other machine learning processes (i.e. learning how to learn). In this talk, I will demonstrate a meta-learning technique that is robust to the ordering of data fed into the model (enabling online continual learning) and automatically adjusts hyperparameters such as effective learning rates with unprecedented granularity. I'll further show how this technique enables "small data" or few-shot learning – successfully training a deep neural network on just 15 examples of never-before-seen classes.  Finally, I'll give examples of how deep learning can impact and accelerate fields like environmental science and medicine across the UVM campus, and explore other ongoing and potential areas of future work in both theoretical and applied AutoML.  

Bio: Nick Cheney is a Research Assistant Professor of Computer Science at the University of Vermont.   He directs the UVM Neurobotics Lab, is a core faculty in the Vermont Complex Systems Center, an Affiliate of the Gund Institute for Environment, and a Participating Faculty in the Quantitative and Evolutionary STEM Traineeship (QuEST).  Nick earned a PhD in Computational Biology from Cornell University – co-advised by Hod Lipson and Steve Strogratz – following a BS in Mathematics from the University of Vermont. He has held visiting researcher and faculty positions at Columbia University, NASA Ames, the Santa Fe Institute, and the University of Wyoming.  Nick's research in machine learning focuses on multi-scale learning-to-learn processes (meta-learning) and automatic machine learning (AutoML) to design self-configurable and robust machine learning pipelines for a wide range of applications. Nick's work has also won numerous awards for scientific visualization and public communication, and been featured in popular media venues such as Wired, Popular Science, The New Yorker, NBC News, and TED.