Deep Amortized Inference for Probabilistic Programs

Probabilistic programming languages (PPLs) are a powerful modeling tool, ableto represent any computable probability distribution. Unfortunately, probabilisticprogram inference is often intractable, and existing PPLs mostly rely on expensive,approximate sampling-based methods. To alleviate this problem, one could tryto learn from past inferences, so that future inferences run faster. This strategyis known as amortized inference; it has recently been applied to Bayesian net-works [28, 22] and deep generative models [20, 15, 24]. This paper proposes asystem for amortized inference in PPLs. In our system, amortization comes in theform of a parameterized guide program. Guide programs have similar structureto the original program, but can have richer data flow, including neural networkcomponents. These networks can be optimized so that the guide approximatelysamples from the posterior distribution defined by the original program. We presenta flexible interface for defining guide programs and a stochastic gradient-basedscheme for optimizing guide parameters, as well as some preliminary results onautomatically deriving guide programs. We explore in detail the common machinelearning pattern in which a ‘local’ model is specified by ‘global’ random valuesand used to generate independent observed data points; this gives rise to amortizedlocal inference supporting global model learning.