State and parameter learning with PARIS particle Gibbs

Authors: Gabriel Cardoso, Yazid Janati El Idrissi, Sylvain Le Corff, Eric Moulines, Jimmy Olsson

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate our theoretical results with numerical experiments supporting our claims. We provide numerical simulations to support our claims, and we show that our algorithm outperforms the current competitors in the two different examples analysed.
Researcher Affiliation Academia Gabriel Cardoso 1 2 3 Yazid Janati El Idrissi 4 Sylvain Le Corff 5 Eric Moulines 1 Jimmy Olsson 6 1CMAP, Ecole Polytechnique, Institut Polytechnique de Paris, Palaiseau. 2Universit e de Bordeaux, CRCTB U4045, INSERM, Bordeaux, France 3IHU Liryc, fondation Bordeaux Universit e, Pessac, France 4Samovar, T el ecom Sud Paris, d epartement CITI, TIPIC, Institut Polytechnique de Paris, Palaiseau 5LPSM, Sorbonne Universit e, UMR CNRS 8001, 4 Place Jussieu, 75005 Paris. 6Department of Mathematics, KTH Royal Institute of Technology, Stockholm, Sweden.
Pseudocode Yes Algorithm 1 One conditional PARIS update (c Pa RIS)... Algorithm 2 One iteration of PPG... Algorithm 3 Gradient estimation with roll-out PPG (c Gd)... Algorithm 4 Score ascent with PPG... Algorithm 5 Score ascent with particle Gibbs kernel... Algorithm 6 One conditional particle filter step CPFs+
Open Source Code Yes The code used in this section is available 1. [Footnote: 1https://anonymous.4open.science/r/ppg/]
Open Datasets No The paper uses simulated data for its experiments: "By simulation, a record of t = 999 observations." and describes a "chaotic recurrent neural network" model. It does not provide access information for a pre-existing public dataset.
Dataset Splits No The paper describes generating data through simulation for its experiments (e.g., "a record of t = 999 observations"), but it does not specify any training, validation, or test dataset splits or percentages.
Hardware Specification Yes All the experiments were performed on a server equipped with 7 A40 Nvidia GPUs.
Software Dependencies No The algorithms were implemented in Python with the JAX Python package (Bradbury et al., 2018) and run on GPU. While JAX is mentioned, specific version numbers for Python or JAX are not provided, which is necessary for reproducibility.
Experiment Setup Yes all the parameters were initialized by sampling from a centered multivariate gaussian distribution with covariance matrix of 0.01I. We have used the ADAM optimizer (Kingma & Ba, 2014) with a learning rate decay of 1/ℓ where ℓ is the iteration index, with a starting learning rate of 0.2. We rescale the gradients by T.