Sampling in Unit Time with Kernel Fisher-Rao Flow

Authors: Aimee Maurais, Youssef Marzouk

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We introduce a new mean-field ODE and corresponding interacting particle systems (IPS) for sampling from an unnormalized target density. The IPS are gradient-free, available in closed form, and only require the ability to sample from a reference density and compute the (unnormalized) target-to-reference density ratio. The mean-field ODE is obtained by solving a Poisson equation for a velocity field that transports samples along the geometric mixture of the two densities, π1 t 0 πt 1, which is the path of a particular Fisher Rao gradient flow. We employ a RKHS ansatz for the velocity field, which makes the Poisson equation tractable and enables discretization of the resulting mean-field ODE over finite samples. The mean-field ODE can be additionally be derived from a discrete-time perspective as the limit of successive linearizations of the Monge Amp ere equations within a framework known as sampledriven optimal transport. We introduce a stochastic variant of our approach and demonstrate empirically that our IPS can produce high-quality samples from varied target distributions, outperforming comparable gradient-free particle systems and competitive with gradient-based alternatives.
Researcher Affiliation Academia 1Center for Computational Science and Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. The methods are described using mathematical equations and textual explanations.
Open Source Code Yes Code for the experiments is available at at https://github.com/amaurais/KFRFlow.jl.
Open Datasets No The paper mentions "The reference distribution π0 is always standard Gaussian." and describes problem setups like "Two-Dimensional Bayesian Posteriors" with specific likelihood definitions, but it does not provide concrete access information (link, DOI, citation) for a publicly available dataset used for training or evaluation. The datasets are constructed based on specific likelihood functions rather than being pre-existing public datasets.
Dataset Splits No The paper does not explicitly provide training/test/validation dataset splits. It describes generating "target ensembles" for evaluation but does not detail how data is partitioned into these specific splits for reproduction purposes.
Hardware Specification Yes Benchmarks were performed in Julia using Benchmark Tools.jl (Chen & Revels, 2016) on a 2020 Mac Book Air with Apple M1 processor.
Software Dependencies No The paper states: "We perform all experiments in Julia using the package Differential Equations.jl (Rackauckas & Nie, 2017)". However, it does not provide specific version numbers for Julia or the DifferentialEquations.jl package, which are necessary for full reproducibility.
Experiment Setup Yes Table 2: Funnels: Selected hyperparameters for each algorithm and target dimension d. lambda (KFRFlow-I) 0.01 0.001 0.001 0.001 epsilon (KFRD) 5 5 5 2.5 lambda (KFRD) 0.001 0.01 0.1 0.1 T (CBS) 25 12.5 25 25 beta (CBS) 0.125 0.5 0.25 0.25 T (SVGD) 100 100 100 100 T (ULA) 12.5 12.5 25 12.5