Improving Gradient-Guided Nested Sampling for Posterior Inference

Authors: Pablo Lemos, Nikolay Malkin, Will Handley, Yoshua Bengio, Yashar Hezaveh, Laurence Perreault-Levasseur

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare the performance of gradient-guided nested sampling with two popular nested sampling algorithms, already introduced in 2: Poly Chord and dynesty. We use the same likelihood function for all algorithms, which is a Gaussian likelihood with a diagonal covariance matrix, and therefore has DKL(P|Π) 𝑑.
Researcher Affiliation Collaboration Pablo Lemos 1 2 3 4 5 Nikolay Malkin 1 3 Will Handley 6 7 8 Yoshua Bengio 1 3 9 Yashar Hezaveh 1 2 3 4 10 11 Laurence Perreault-Levasseur 1 2 3 4 10 11 1Mila Qu ebec Artificial Intelligence Institute 2Ciela Institute 3Universit e de Montr eal 4CCA Flatiron Institute 5Dreamfold 6Cavendish Laboratory 7Kavli Institute for Cosmology 8Gonville and Caius College, University of Cambridge 9Canadian Institute for Advanced Research (CIFAR) 10Perimeter Institute 11Trottier Space Institute.
Pseudocode Yes Algorithm 1 A summarized version of the GGNS algorithm. (Full version in F.)
Open Source Code Yes An implementation of GGNS in Py Torch (Paszke et al., 2019), along with notebooks to reproduce the results from the experiments, is available at https://github.com/Pablo-Lemos/GGNS.
Open Datasets No The paper uses synthetic datasets like 'Gaussian mixture', 'Neal’s funnel distribution', 'Many Well problem', 'Image Generation', and 'Lennard-Jones potential', but does not provide concrete access information (e.g., links, DOIs, or specific citations to publicly available versions) for these datasets.
Dataset Splits No The paper describes experiments on synthetic distributions and does not specify train, validation, or test dataset splits (e.g., percentages or sample counts) typically found in machine learning experiments.
Hardware Specification No The paper mentions running experiments on 'a single CPU' and discusses 'massive parallelization on GPUs', but does not specify any particular CPU or GPU models, or other detailed hardware specifications.
Software Dependencies No The paper mentions using 'Py Torch' for implementation and refers to 'jAX and torch' in the context of differentiable programming, but it does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We summarize the hyperparameters in C, and provide extended ablation studies in E.