Momentum Particle Maximum Likelihood

Authors: Jen Ning Lim, Juan Kuntz, Samuel Power, Adam Michael Johansen

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

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 6, we demonstrate empirically the efficacy of our method and study the effects of various design choices. As a large-scale experiment, we benchmark our proposed method for training Variational Autoencoders (Kingma and Welling, 2014) against current methods for training latent variable models.
Researcher Affiliation Collaboration 1University of Warwick 2Polygeist 3University of Bristol.
Pseudocode Yes Algorithm 1 A single subsampled step. In pink, we indicate the existing subsampling scheme of Kuntz et al. (2023). We indicate our proposed additions in teal.
Open Source Code Yes The code for reproducibility is available online: https://github.com/jenninglim/mpgd.
Open Datasets Yes For this task, we consider two datasets: MNIST (Le Cun et al., 1998) and CIFAR-10 (Krizhevsky and Hinton, 2009).
Dataset Splits No The paper states 'We use N = 5000 images for training and 3000 for testing' but does not explicitly mention or detail a validation split.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU models, or cloud computing instance types used for experiments.
Software Dependencies No The paper does not specify versions for any software dependencies or libraries used, such as Python, PyTorch, or CUDA.
Experiment Setup Yes MPD. For step sizes, we have hθ = hx = 10 4. The number of particles is set to M = 5. For the momentum parameters, we use γθ = γx = 0.9 with the momentum coefficient µθ = 0.95, µx = 0 (or equivalently, ηθ 556, ηx 11, 111 )for MNIST and µθ = 0.95, µx = 0.5 (or equivalently, ηθ 556, , ηx 55, 555 ) for CIFAR. We use the RMSProp preconditioner (see Appendix J.1) with β = 0.9.