Towards Probabilistic Programming for Reliable Machine Perception - Marco Cusumano-Towner

Workshop on Dependable and Secure Software Systems 2019

Applications of Bayesian inference in fields like robotics and computer vision often require specialized real-time approximate inference implementations, and practitioners in these fields must make trade offs between the approximation error and performance characteristics of their implementation. This talk will introduce two probabilistic programming tools aimed at supporting performance-constrained inference applications. The first part of the talk will introduce Gen, a general-purpose probabilistic programming system that achieves flexibility and efficiency via several novel language constructs, including an interface for encapsulating hand-optimized inference code, and interoperable modeling languages that strike different flexibility/efficiency trade-offs. The second part of the talk will introduce the Auxiliary Inference Divergence Estimator (AIDE), which measures the KL divergence between the output distributions of two sampling programs. AIDE can be used offline to evaluate the approximation error of a fast real-time inference algorithm against gold-standard inference algorithm on specific user-selected inference problems.