Automatically marginalized MCMC in probabilistic programming
Authors: Jinlin Lai, Javier Burroni, Hui Guan, Daniel Sheldon
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our methods can substantially improve the effectiveness of samples from hierarchical partial pooling models and hierarchical linear regression models and significantly outperforms model reparameterization (Betancourt & Girolami, 2015) in those models where both apply. |
| Researcher Affiliation | Academia | 1University of Massachusetts Amherst. Correspondence to: Jinlin Lai <jinlinlai@cs.umass.edu>. |
| Pseudocode | Yes | Algorithm 1 Marginalize and recover unobserved nodes; Algorithm 2 Reversing an edge (normal-normal case); Algorithm 3 Determining dependency of a variable on an input.; Algorithm 4 Determining affinity and linearity of a variable on an input.; Algorithm 5 Getting the coefficients of affine relationship between a variable wj on an input x.; Algorithm 6 The full version of Algorithm 2: reversing an edge. |
| Open Source Code | Yes | Our code is available at https://github.com/lll6924/automatically-marginalized MCMC. |
| Open Datasets | Yes | The eight schools model (Gelman et al., 1995) is an important demonstration model for PPLs (Gorinova, 2022) and reparameterization (Papaspiliopoulos et al., 2007).; Applications include the rat tumors dataset (Tarone, 1982), the baseball hits 1970 dataset (Efron & Morris, 1975) and the baseball hit 1996 AL dataset (Carpenter et al., 2017).; The electric company model (Gelman & Hill, 2006) studies the effect of an educational TV program on children s reading abilities.; The Pulmonary fibrosis dataset (Shahin et al., 2020) has patient observation records over time of forced vital capacity (FVC), a disease indicator. |
| Dataset Splits | No | No specific dataset splits (e.g., percentages or counts for training, validation, and testing) were explicitly provided in the paper. The paper mentions '10,000 warm up samples to tune the sampler' and '100,000 samples for evaluation' which refer to MCMC chains, not dataset partitioning. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory specifications, or cloud instance types) used for running experiments were mentioned. |
| Software Dependencies | No | The paper mentions software like 'JAX' and 'Num Pyro' but does not provide specific version numbers for reproducibility. |
| Experiment Setup | Yes | For all experiments, we use 10,000 warm up samples to tune the sampler, 100,000 samples for evaluation, and evaluate performance via effective sample size (ESS) and time (inclusive of JAX compilation time).; As a workaround, we manually prevented µi from being marginalized. |