Build a Learning Habit
Download Class Central's free printable study calendar
Download for Free
This 48-minute talk by Elizabeth Polgreen from the University of Edinburgh explores innovative approaches to automatically lifting code to domain-specific languages (DSLs). Learn how the proliferation of specialized accelerators has created numerous DSLs whose high-level nature enables efficient code generation for heterogeneous hardware. Discover two novel approaches that combine language models with program synthesis: one generating probabilistic context-free grammars to represent likely solutions followed by enumerative synthesis, and another using small language models to create initial solutions refined through measurement oracles and edit rule searches. See practical applications of these techniques for lifting legacy code to tensor DSLs, achieving impressive speed-ups of up to 38x over unlifted code. Part of the Theoretical Aspects of Trustworthy AI series at the Simons Institute.
