XPS: FULL: Collaborative Research: Parallel and Distributed Circuit Programming for Structured Prediction

William Marsh Rice University, Houston TX

Investigators

Abstract

This project develops a system for "circuit programming," which allows a programmer to focus on the high-level solution to a problem rather than on the details of how the computation is organized. Circuit programming consists of writing rules that describe how data items depend on one another. The intellectual merits lie in the design of a new programming language for specifying these rules, along with the algorithms whereby the computer automatically finds efficient strategies for managing the necessary computations on available parallel hardware. The project's broader significance and importance lie in its potential to streamline work in areas such as artificial intelligence and machine learning. With the growing complexity of systems in these areas and their need to process big data in depth, research and teaching typically get bogged down in programming details, especially for parallel platforms; this project aims to delegate those details to automatic methods. The research develops a programming system for Dyna, a circuit programming language that enables concise specification of large function graphs that may be cyclic and/or infinite. Dyna employs (1) a pattern-matching notation that augments pure Prolog with evaluation and aggregation and (2) an object-like mechanism for dynamically defining new sub-circuits as modifications of old ones. This project is building an adaptive system that can mix forward and backward chaining to seek a fixpoint of the circuit and to update this fixpoint as the inputs change. The system will perform compile-time and runtime analysis of the Dyna program and will map it to Habanero, a system for scheduling parallel computations on multicore processors, with extensions for task priorities, task cancellation, GPU execution, and distributed execution.