Analysis and Optimization of Approximate Programs

Many modern applications (such as multimedia processing, machine learning, and big-data analytics) exhibit an inherent tradeoff between the accuracy of the results they produce and the execution time or energy consumption. These applications allow us to investigate new optimization approaches that exploit approximation opportunities at every level of the computing stack and therefore have the potential to provide savings beyond the reach of standard semantics-preserving program optimizations.

In this talk, I will describe a novel approximate optimization framework based on accuracy-aware program transformations. These transformations trade accuracy in return for improved performance, energy efficiency, and/or resilience. The optimization framework includes program analyses that characterize the accuracy of transformed programs and search techniques that navigate the tradeoff space induced by transformations to find approximate programs with profitable tradeoffs. I will particularly focus on how we (i) automatically generate computations that execute on approximate hardware platforms, while ensuring that they satisfy the developer's accuracy specifications and (ii) apply probabilistic reasoning to quantify uncertainty coming from inputs or caused by program transformations, and analyze the accuracy of approximate computations.

Bio

Sasa Misailovic graduated with a Ph.D. from MIT in 2015. He will start as an Assistant Professor in the Computer Science Department at the University of Illinois at Urbana-Champaign in Fall 2016. During this academic year he is visiting Software Reliability Lab at ETH Zurich. His research interests include programming languages, software engineering, and computer systems, with an emphasis on improving performance, energy efficiency, and resilience in the face of software errors and approximation opportunities.