Abstract

Memory efficiency is an important issue in Sequential Monte Carlo (SMC) algorithms, arising for example in inference of high-dimensional latent variables via Rao-Blackwellized SMC algorithms, where the size of individual particles combined with the required number of particles can stress the main memory. Standard SMC methods have a memory requirement that scales linearly in the number of particles present at all stage of the algorithm. Our contribution is a simple scheme that makes the memory cost of SMC methods depends on the number of distinct particles that survive resampling. We show that this difference has a large empirical impact on the quality of the approximation in realistic scenarios, and also—since memory access is generally slow—on the running time. The method is based on a two pass generation of the particles, which are represented implicitly in the first pass. We parameterize the accuracy of our algorithm with a memory budget rather than with a fixed number of particles. Our algorithm adaptively selects an optimal number of particle to exploit this fixed memory budget. We show that this adaptation does not interfere with the usual consistency guarantees that come with SMC algorithms.