Mean-Field Langevin Dynamics for Signed Measures via a Bilevel Approach

Authors: Guillaume Wang, Alireza Mousavi-Hosseini, Lénaïc Chizat

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

Reproducibility Variable Result LLM Response
Research Type Experimental See Fig. 1 for an illustrative numerical experiment. Note that our simulations of Brownian motion are not exact. The code to reproduce this experiment can be found at https://github.com/mousavih/ 2024-MFLD-bilevel. Neur IPS Paper Checklist: The contributions of this work are theoretical. A numerical illustration is given in Fig. 1
Researcher Affiliation Academia Guillaume Wang 1 Alireza Mousavi-Hosseini 2 Lénaïc Chizat1 1École polytechnique fédérale de Lausanne 2University of Toronto and Vector Institute
Pseudocode Yes Algorithm 1 Annealing of the MFLD. Require: Functional J : P(W) R. Initialization η0, β0 > 0. Schedule K, (Tk)K k=0. 1: η0 0 = η0 2: for k = 0, . . . , K do 3: βk = 2kβ0 4: Run the MFLD with βk initialized from ηk 0 up to Tk, tηk t = div(ηk t J [ηk t ]) + 1 βk ηk t 5: ηk+1 0 = ηk Tk 6: end for 7: return ηK TK.
Open Source Code Yes The code to reproduce this experiment can be found at https://github.com/mousavih/ 2024-MFLD-bilevel.
Open Datasets No ρ is the empirical distribution of a (covariate) dataset (xi)i n of n = 100 training samples, sampled i.i.d. from N 0d 1 , with the last coordinate representing bias. The paper describes the generation of a synthetic dataset but does not provide access information (link, citation, or repository) to a publicly available version of this specific dataset.
Dataset Splits No The paper mentions "training samples" but does not specify any training/validation/test dataset splits (e.g., percentages or counts for each split).
Hardware Specification No The paper states that "any standard laptop or desktop computer can be used to reproduce it in, with a runtime of a few minutes." This is a general statement and does not provide specific hardware details (e.g., CPU/GPU models, memory amounts).
Software Dependencies No The paper mentions that code is available at a GitHub link, but it does not explicitly list specific software dependencies with their version numbers within the text itself (e.g., Python version, PyTorch version, etc.).
Experiment Setup Yes We consider the problem (1.1) where W = Sd and G is defined as in Assumption 2, where d = 10, λ = 10 3... For the algorithms using MFLD, we used β 1 = 10 3. We ran the Euler-Maruyama discretization of the noisy particle gradient flow SDE described in Sec. 2... using N = 1000 particles... and a step size of 10 2 for (1a) and 10 3 for (1b)... the wi 0 are drawn i.i.d. uniformly on Sd, and for the algorithms using the lifting formulation, the ri 0 are drawn i.i.d. from N(0, 1).