Analysis and Testing for Probabilistic Programming - Sasa Misailovic скачать в хорошем качестве

Analysis and Testing for Probabilistic Programming - Sasa Misailovic

Workshop on Dependable and Secure Software Systems 2019 Probabilistic programming languages offer an intuitive way to model uncertainty by representing complex probability models as simple probabilistic programs. Thus, even a programmer with a limited exposure to statistical machine learning can benefit from powerful probabilistic inference. In effect, the programmer only needs to write a high-level model in a programming language with support for random sampling and conditioning on data. The underlying probabilistic programming system (which executes the high-level probabilistic programs) hides the complexity of inference algorithms away from the program developer. While the research community has focused primarily on improving expressiveness of probabilistic languages and efficiency of inference algorithms, analyses of probabilistic programs and techniques for debugging probabilistic software will have a key role in improving programmer productivity in this emerging domain. In this talk, I will present our recent work on testing probabilistic programming systems and tools for analyzing robustness of probabilistic programs.