Inference Building Blocks
Abstract
We address the problem that probabilistic inference algorithms are difficult and tedious to implement, by expressing them in terms of a small number of building blocks, which are automatic transformations on probabilistic programs. On one hand, our curation of these building blocks reflects the way human practitioners discuss probabilistic inference with each other, so our probabilistic programming language supports modular composition of inference procedures and serves as a medium for collaboration. On the other hand, our implementation of these building blocks combines high-level mathematical reasoning with low-level computational optimization, so the speed and accuracy of the generated solvers are competitive with state-of-the-art systems.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 15, 2018
- Accession Number
- AD1047179
Entities
People
- Chung-chieh Shan
- Geneva Smith
- Jacques Carett
- Norman Ramsey
- Oleg Kiselyov
- Praveen Narayanan
- Rajan Walia
- Robert Zinkov
- Scherrer Chad
- Wazim I. Mohammed
- Wren Romano
- Yuriy Toporovskyy
- Zachary Sullivan
Organizations
- Indiana University
- United States Army Combat Capabilities Development Command