We propose an amortized inference strategy for probabilistic programs, one that learns from the past inferences to speed up the future inferences. Our proposed inference strategy is to train neural guidance programs via a minimax game, with the probabilistic program … Continue reading →