Adaptive Perturbation-Based Gradient Estimation for Discrete Latent Variable Models

The integration of discrete algorithmic components in deep learning architectures has numerous applications. Recently, Implicit Maximum Likelihood Estimation, a class gradient estimators for exponential family distributions, was proposed by combining implicit differentiation through perturbation with the path-wise estimator. However, due to finite difference approximation gradients, it is especially sensitive choice step size, which needs be specified user. In this work, we present Adaptive IMLE (AIMLE), first adaptive estimator complex distributions: adaptively identifies target distribution trading off density information degree bias estimates. We empirically evaluate our on synthetic examples, as well Learning Explain, Discrete Variational Auto-Encoders, and Neural Relational Inference tasks. experiments, show that can produce faithful estimates while requiring orders magnitude fewer samples than other estimators.