Amortized Inference for Latent Feature Models Using Variational Russian Roulette

The Indian buffet process (IBP) provides a principled prior distribution for inferring the number of latent features in a dataset. Traditionally, inference for these models is slow when applied to large datasets, which motivates the use of amortized neural inference methods. However, previous works on variational inference for these models require the use of a truncated approximation, in which the maximum number of features is predetermined. To address this problem, we present a new dynamic variational posterior by introducing auxiliary variables to the stick-breaking construction of IBP. We describe how to estimate the evidence lower bound, which contains summations of infinite terms, using Russian roulette sampling.