Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Mean-Field Langevin Dynamics : Exponential Convergence and Annealing

Authors: Lénaïc Chizat

TMLR 2022 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, we show that our results apply to noisy gradient descent on infinitely wide two-layer neural networks and we provide numerical experiments for G being a kernel Maximum Mean Discrepancy (MMD). We conclude this paper with numerical experiments exploring the behavior of NPGD1 defined in (3). Figure 1a shows an example of a large-time particle configuration, with the atoms of ν is red and the atoms of ˆµt in black (with t large), with a noise temperature τ = 0.1. Figure 1b shows the evolution of the objective Fτ = G+τH (up to a constant, adjusted for ease of comparison) along the iterations, where the entropy H is estimated using the 1-nearest-neighbor estimator (Kozachenko and Leonenko, 1987; Singh et al., 2003). Finally, Figure 1c shows the advantage of NPGD with simulated annealing vs. PGD to minimize the unregularized function G.
Researcher Affiliation Academia Lénaïc Chizat lenaic.chizat@epfl.ch EPFL
Pseudocode No The paper describes algorithms through mathematical equations and definitions, such as Equation (3) for NPGD and Equation (11) for SDE, but does not include a distinct block or section labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code Yes 1Link to Julia code to reproduce the experiments: https://github.com/lchizat/2022-mean-field-langevin-rate.
Open Datasets No We take ν as a random empirical distribution of m = 10 samples from the uniform distribution on X.
Dataset Splits No The paper describes generating a random empirical distribution for ν and running simulations with a fixed number of particles, but does not provide specific training/test/validation splits for any dataset.
Hardware Specification No The paper does not specify any particular hardware (CPU, GPU, etc.) used for running the numerical experiments.
Software Dependencies No The paper mentions 'Julia code' in a footnote for reproducing experiments, but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes We run NPGD with m = 50 particles, a step-size η = 0.08 and µ0 being the uniform distribution on X. We used a noise temperature that decays polynomially as τt = 20(t + 1)−1 where t is the iteration count. At iteration 800, we stopped the noise in order to observe the quality of the configuration of particles.