Name: Paul Hovland
Pronouns: he/him/his
Biography:
Paul Hovland’s research focuses on software development tools for high-performance scientific computing. He holds a B.S. in computer engineering and an M.S. in computer science from Michigan State University. He received his Ph.D. in computer science with a computational science and engineering option from the University of Illinois at Urbana-Champaign. Research Interests include automatic differentiation (autodial or AD), automatic empirical performance tuning (autotuning), program verification, and quantum computing.
Institution/Lab: Argonne National Laboratory
Website:
SRP Collaboration Topic/Title: ML-based compression for derivative computation
Field or research area: Automatic differentiation
Please select all the topical areas that apply to your project:
Computer Science (i.e., architectures, compilers/languages, networks, workflow/edge, experiment automation, containers, neuromorphic computing, programming models, operating systems, sustainable software); Machine Learning and AI
Brief Abstract:
Efficient computation of derivatives for both computational science and machine learning applications often relies on the so-called reverse mode of automatic differentiation (autodiff). Unfortunately, reverse-mode autodiff requires saving many intermediate states, which can lead to substantial memory or storage requirements. This project will investigate whether lossy compression techniques based on machine learning can be used to compress this data and reduce memory requirements while maintaining suitable levels of accuracy in the derivative computations.
Desired relevant skills, background, or interests:
Basic understanding of derivatives (first semester calculus) Programming in python or a related language Interest in machine learning
Other comments:
Do any special requirements apply? Permanent Resident OK; International OK
Other, specify:
Keywords:
Automatic differentiation; autodiff; algorithmic differentiation; autoencoder; lossy compression; machine learning
Lightning Talk Title: AutoDiff @ Argonne