•   When: Friday, April 13, 2018 from 01:30 PM to 02:30 PM
•   Speakers: Lane Schwartz
•   Location: ENGR 4201
•   Export to iCal

Short Bio:

Lane Schwartz works at the intersection of human and machine translation. His research includes work in grammar induction, machine translation, computer-aided translation, and cognitively-motivated language models. Dr. Schwartz holds a B.A. from Luther College with minors in German and Theatre/Dance. He earned an M.Phil in Computer Speech, Text and Internet Technology from the University of Cambridge, and a Ph.D in Computer Science from the University of Minnesota. He is one of the original developers of Joshua, an open source toolkit for tree-based statistical machine translation, and a frequent contributor to Moses, the de-facto standard for phrase-based statistical machine translation.

Abstract:

Deep learning software demands reliability and performance. However, many of the existing deep learning frameworks are software libraries that act as an unsafe DSL in Python and a computation graph interpreter. We present DLVM, a design and implementation of a compiler infrastructure with a linear algebra intermediate representation, algorithmic differentiation by adjoint code generation, domain-specific optimizations and a planned code generator targeting GPU via LLVM. Designed as a modern compiler infrastructure inspired by LLVM, DLVM is more modular and more generic than existing deep learning compiler frameworks, and supports tensor DSLs with high expressivity. With our prototypical staged DSL embedded in Swift, we argue that the DLVM system enables a form of modular, safe and performant frameworks for deep learning.