Serialization and Applications for the Gen Probabilistic Programming Language

Abstract

Probabilistic programming has emerged as a powerful framework for building expressive models that can handle uncertainty in a wide range of applications. Serialization, the process of converting data structures or objects into a format suitable for storage or transmission, plays a crucial role in the development and execution of probabilistic programs. Efficient serialization techniques are essential for tasks such as data persistence, distributed computation, and data exchange between different programs or machines. We delve into specific challenges and considerations unique to probabilistic programming for serialization. Probabilistic models often involve complex structures, including nested random variables, hierarchical dependencies, and potentially infinite or unbounded dimensions. Serializing samples from such models requires careful handling of these complexities, including strategies for preserving model fidelity, dealing with modeling dependencies, and specializing for disk representations. In this thesis, we discuss twofold objectives for the Gen probabilistic programming model. The first establishes a formalism for serializing (and deserializing) traces as an interface that respects the existing Gen interfaces and faithfully reconstructs data objects from disk. We highlight challenges for efficient serialization for Gen’s DSLs. The second objective is to show how serialization routines common in other areas of computing transfer well to Gen. We show how serialization provides easiers means for visualizations, remote computing, and training inference approximators.