Composable probabilistic inference with BLAISE

Abstract

If we are to understand human-level cognition, we must understand how the mind finds the patterns that underlie the incomplete, noisy, and ambiguous data from our senses and that allow us to generalize our experiences to new situations. A wide variety of commercial applications face similar issues: industries from health services to business intelligence to oil field exploration critically depend on their ability to find patterns in vast amounts of data and use those patterns to make accurate predictions. Probabilistic inference provides a unified, systematic framework for specifying and solving these problems. Recent work has demonstrated the great value of probabilistic models defined over complex, structured domains. However, our ability to imagine probabilistic models has far outstripped our ability to programmatically manipulate them and to effectively implement inference, limiting the complexity of the problems that we can solve in practice. This thesis presents BLAISE, a novel framework for composable probabilistic modeling and inference, designed to address these limitations. BLAISE has three components: * The BLAISE State-Density-Kernel (SDK) graphical modeling language that generalizes factor graphs by: (1) explicitly representing inference algorithms (and their locality) using a new type of graph node, (2) representing hierarchical composition and repeated substructures in the state space, the interest distribution, and the inference procedure, and (3) permitting the structure of the model to change during algorithm execution. * A suite of SDK graph transformations that may be used to extend a model (e.g. to construct a mixture model from a model of a mixture component), or to make inference more effective (e.g. by automatically constructing a parallel tempered version of an algorithm or by exploiting conjugacy in a model).

(cont.) * The BLAISE Virtual Machine, a runtime environment that can efficiently execute the stochastic automata represented by BLAISE SDK graphs. BLAISE encourages the construction of sophisticated models by composing simpler models, allowing the designer to implement and verify small portions of the model and inference method, and to reuse mode components from one task to another. BLAISE decouples the implementation of the inference algorithm from the specification of the interest distribution, even in cases (such as Gibbs sampling) where the shape of the interest distribution guides the inference. This gives modelers the freedom to explore alternate models without slow, error-prone reimplementation. The compositional nature of BLAISE enables novel reinterpretations of advanced Monte Carlo inference techniques (such as parallel tempering) as simple transformations of BLAISE SDK graphs. In this thesis, I describe each of the components of the BLAISE modeling framework, as well as validating the BLAISE framework by highlighting a variety of contemporary sophisticated models that have been developed by the BLAISE user community. I also present several surprising findings stemming from the BLAISE modeling framework, including that an Infinite Relational Model can be built using exactly the same inference methods as a simple mixture model, that constructing a parallel tempered inference algorithm should be a point-and-click/one-line-of-code operation, and that Markov chain Monte Carlo for probabilistic models with complicated long-distance dependencies, such as a stochastic version of Scheme, can be managed using standard BLAISE mechanisms.