Notice: This website is continuously updated as new conference proceedings and journal volumes are published. Consequently, the figures presented here may differ from those reported in our publications.
A brief overview of our methodology is provided on the Methods page, along with references to publications documenting our approach and links to the code and data used in each study.
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 and potential bias; scores should be interpreted as estimates.
On Sampling with Approximate Transport Maps
Authors: Louis Grenioux, Alain Oliviero Durmus, Eric Moulines, Marylou Gabrié
ICML 2023 | Venue PDF | LLM Run Details | Input Tokens: 39,591 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 6,711 Total number of tokens produced by the LLM (including reasoning/thinking tokens) for this paper's analysis.
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our study concludes that multimodal targets can be reliably handled with flow-based proposals up to moderately high dimensions. In contrast, methods relying on reparametrization struggle with multimodality but are more robust otherwise in high-dimensional settings and under poor training. To further illustrate the influence of targetproposal adequacy, we also derive a new quantitative bound for the mixing time of the Independent Metropolis-Hastings sampler.In all of our experiments, we selected the samplers hyperparameters by optimizing case-specific performance metrics. The length of chains was chosen to be twice the number of steps required to satisfy the ˆR diagnosis (Gelman & Rubin, 1992) at the 1.1-threshold for the fastest converging algorithm. We used MALA as local sampler as it is suited for the log-concave distributions considered, faster and easier to tune than HMC. Evaluation metrics are fully descibed in App. D.1. |
| Researcher Affiliation | Academia | 1 Ecole Polytechnique. Correspondence to: Louis Grenioux <EMAIL>, Alain Oliviero Durmus <EMAIL>, Eric Moulines <EMAIL>, Marylou Gabri e <EMAIL>. |
| Pseudocode | No | The paper does not contain any explicitly labeled “Pseudocode” or “Algorithm” blocks. |
| Open Source Code | Yes | The code to reproduce the experiments is available at https://github.com/h2o64/flow mcmc. |
| Open Datasets | Yes | Our second experiment is a sparse Bayesian hierarchical logistic regression on the German credit dataset (Dua & Graff, 2017), which has been used as a benchmark in recent papers (Hoffman et al., 2019; Grumitt et al., 2022; Cabezas & Nemeth, 2022). The ground truth was obtained through parallel tempering MD simulations from the same paper. We trained a SN-GAN (Miyato et al., 2018) on the CIFAR10 dataset for 100.000 epochs (implementation from https://github.com/kwotsin/mimicry). |
| Dataset Splits | No | The paper mentions training on datasets like German credit and CIFAR10, but it does not explicitly provide percentages, sample counts, or specific methodology for training, validation, and test splits for reproducibility. While standard datasets often have predefined splits, the paper does not state how these were applied or referenced for its experiments. |
| Hardware Specification | Yes | We used a single type of GPU, the Nvidia A100. Each experiment of Section 3 took about 2 hours to run when distributed on 4 GPUs. |
| Software Dependencies | No | The paper mentions using Adam (Kingma & Ba, 2014) optimizer and Pyro’s (Bingham et al., 2019) implementation of HMC. However, it does not specify version numbers for these or other software libraries (e.g., Python, PyTorch/TensorFlow, CUDA) required for exact reproducibility. |
| Experiment Setup | Yes | In all of our experiments, we selected the samplers hyperparameters by optimizing case-specific performance metrics. The length of chains was chosen to be twice the number of steps required to satisfy the ˆR diagnosis (Gelman & Rubin, 1992) at the 1.1-threshold for the fastest converging algorithm. We used MALA as local sampler as it is suited for the log-concave distributions considered, faster and easier to tune than HMC. Evaluation metrics are fully descibed in App. D.1. (Tables 3, 4, 5, 6, 7, 8, 9 provide specific hyperparameters like number of iterations, learning rate, hidden layer sizes, MCMC steps, and particle counts for various experiments.) |