Predicate Exchange: Inference with Declarative Knowledge
Zenna Tavares · Javier Burroni · Edgar Minasyan · Armando Solar-Lezama · Rajesh Ranganath

Wed Jun 12th 04:20 -- 04:25 PM @ Hall B

We address the problem of conditioning probabilistic models on predicates, as a means to express declarative knowledge. Models conditioned on predicates rarely have a tractable likelihood; sampling from them requires likelihood-free inference. Existing likelihood-free inference methods focus on predicates which express observations. To address a broader class of predicates, we develop an inference procedure called predicate exchange, which \emph{softens} predicates. Soft predicates return values in a continuous Boolean algebra and can serve as a proxy likelihood function in inference. However, softening introduces an approximation error which depends on a temperature parameter. At zero-temperature predicates are identical, but are often intractable to condition on. At higher temperatures, soft predicates are easier to sample from, but introduce more error. To mitigate this trade-off, we simulate Markov chains at different temperatures and use replica exchange to swap between chains. We implement predicate exchange through a nonstandard execution of a simulation based model, and provide a light-weight tool that can be supplanted on top of existing probabilistic programming formalisms. We demonstrate the approach on sequence models of health and inverse rendering.

Author Information

zenna Tavares (MIT)
Javier Burroni (UMass Amherst)
Edgar Minasyan (Princeton University)
Armando Solar-Lezama (MIT)
Rajesh Ranganath (New York University)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors