Local-Global MCMC kernels: the best of both worlds

Authors: Sergey Samsonov, Evgeny Lagutin, Marylou Gabrié, Alain Durmus, Alexey Naumov, Eric Moulines

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the efficiency of Ex2MCMC and its adaptive version on classical sampling benchmarks as well as in sampling high-dimensional distributions defined by Generative Adversarial Networks seen as Energy Based Models. We perform a numerical evaluation of Ex2MCMC and Fl Ex2MCMC for various sampling problems, including sampling GANs as energy-based models.
Researcher Affiliation Academia Sergey Samsonov1 Evgeny Lagutin1 Marylou Gabrié2 Alain Durmus3 Alexey Naumov1 Eric Moulines2 1HSE University 2Ecole Polytechnique 3ENS Paris-Saclay
Pseudocode Yes Algorithm 1: Single stage of i-SIR algorithm with independent proposals; Algorithm 2: Single stage of Ex2MCMC algorithm with independent proposals; Algorithm 3: Single stage of Fl Ex2MCMC.
Open Source Code Yes We provide the code to reproduce the experiments below at https://github.com/svsamsonov/ex2mcmc_new.
Open Datasets Yes We consider sampling from a mixture of 3 equally weighted Gaussians in dimension d = 2. ... Following [52] and [32], we consider the funnel and the banana-shape distributions. ... MNIST results. We consider a simple Jensen-Shannon GAN model trained on the MNIST dataset... Cifar-10 results. We consider two popular architectures trained on Cifar-10...
Dataset Splits No The paper describes experiments on MCMC sampling from various distributions and pre-trained GANs. It does not involve training their proposed Ex2MCMC method in a supervised learning context that would require train/validation/test splits for their method's evaluation.
Hardware Specification No The paper states: '3. If you ran experiments... (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] We provide this information in the supplement paper.' This indicates the hardware specifications are not provided in the main paper.
Software Dependencies No The paper does not provide specific software names with version numbers in the main text. It mentions that training details (which might include software dependencies) are provided in the supplement: '3. If you ran experiments... (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] The hyperparameters are provided in the supplement paper.'
Experiment Setup No The paper states: '3. If you ran experiments... (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] The hyperparameters are provided in the supplement paper.' This indicates that comprehensive experimental setup details are not in the main paper.