Adaptive Annealed Importance Sampling with Constant Rate Progress

Authors: Shirin Goshtasbpour, Victor Cohen, Fernando Perez-Cruz

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we run a number of experiments to illustrate the performance of CR-AIS on support coverage and adaptivity with 2d distributions and we asses its efficiency and accuracy with estimation of the log normal-ization constant of high dimensional synthetic targets and the posterior of Bayesian models.
Researcher Affiliation Academia 1Computer Science Department, ETH Zurich, Zurich, Switzerland 2Swiss Data Science Center, Zurich, Switzerland.
Pseudocode Yes Algorithm 1 CR-AIS tuning for α-divergences
Open Source Code Yes Code is available at https://github.com/shgoshtasb/cr_ais.
Open Datasets Yes We use two UCI datasets, Pima Indians diabetes dataset (N = 768 and d = 8) and Sonar dataset (N = 207 and d = 60) with bi-nary labels and a setup similar to (Chopin et al., 2020) with AIS. We estimate the log marginal likelihood of a Variational Auto Encoder trained on binarized MNIST dataset (Salakhutdinov & Murray, 2008).
Dataset Splits No The paper mentions using 'cross-validation' and describes tuning processes for schedules, but it does not provide explicit details of training/validation/test dataset splits (e.g., percentages or sample counts) for its experiments.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., CPU/GPU models, memory specifications).
Software Dependencies No The paper does not list specific software dependencies with their version numbers required for reproducibility.
Experiment Setup Yes We initialize CR-AIS with a standard normal distribution for q0 and we use the same value of δ = 1/32 for all the targets. Each AIS transition is a single Hamiltonian Monte Carlo (HMC) step with step size 0.5 and N = 1024 particles are used to approximate the constant rate schedule. We set N = 4096 and chose α from a grid in { 0.5, 0.0, ..., 2.}, δ from {1/256, ..., 4096} for CR-AIS depending on the target... We vary δ between {64, ..., 2048} for CR-AIS and maximum step size in the range {1/65536, ..., 1/2048} for Adaptive AIS with N = 256 particles.