Fast, accurate and lightweight sequential simulation-based inference using Gaussian locally linear mappings
Authors: Henrik Häggström, Pedro L. C. Rodrigues, Geoffroy Oudoumanessah, Florence Forbes, Umberto Picchini
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We illustrate our results on several benchmark models from the SBI literature and on a biological model of the translation kinetics after m RNA transfection. 5 Numerical illustrations: We illustrate Se MPLE on several simulation studies, see also the Appendices for extra results and additional models. |
| Researcher Affiliation | Academia | Henrik Häggström EMAIL Dept. Mathematical Sciences Chalmers and University of Gothenburg; Pedro L. C. Rodrigues EMAIL Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, Inserm U1216, CHU Grenoble Alpes, Grenoble Institut des Neurosciences, LJK; Geoffroy Oudoumanessah EMAIL Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, Inserm U1216, CHU Grenoble Alpes, Grenoble Institut des Neurosciences, LJK Univ. Lyon, CNRS, Inserm, INSA Lyon, UCBL, CREATIS; Florence Forbes EMAIL Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, Inserm U1216, CHU Grenoble Alpes, Grenoble Institut des Neurosciences, LJK; Umberto Picchini EMAIL Dept. Mathematical Sciences, Chalmers and University of Gothenburg |
| Pseudocode | Yes | Algorithm 1 Se MPLE; Algorithm 2 Independence MH with GLLi M surrogate posterior as proposal distribution |
| Open Source Code | Yes | Code is available at https://github.com/henhagg/semple. |
| Open Datasets | Yes | We illustrate our results on several benchmark models from the SBI literature and on a biological model of the translation kinetics after m RNA transfection. Our experiments are run on the same 10 datasets provided in SBIBM, each of them being a vector y0 of length two. (Lueckmann et al., 2021) |
| Dataset Splits | No | No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) is provided for reproducing the data partitioning. The paper states that experiments are run on existing datasets (e.g., from SBIBM) or that data is generated for specific models, but doesn't detail how these are divided into training, validation, or testing sets within their experimental framework beyond referring to 'datasets provided' or 'generated data'. |
| Hardware Specification | Yes | The experiments were run on a machine with a CPU Intel Core TM i7-4790 CPU @ 3.60GHz and 16GB of DRAM. |
| Software Dependencies | Yes | Se MPLE is written in R and uses the GLLi M implementation from the x LLi M package (Perthame et al., 2022). Comparisons with neural-based likelihood and posterior estimation methods (SNL and SNPEC) use the Python implementations in the SBIBM package for SBI benchmarking (Lueckmann et al., 2021). Solutions to the SDE (30) are simulated using an Euler-Maruyama scheme implemented in Rcpp (Eddelbuettel et al., 2024)... R package version 1.0.12. |
| Experiment Setup | Yes | In most cases we find that setting a starting value of K = 20 or 30 produces good results, with a threshold for ηk typically set to 0.005 (multiple hyperboloid, Ornstein-Uhlenbeck, Lotka-Volterra, biological SDE model), 0.03 (Bernoulli GLM model) or simply 0 for no deletion (Two Moons model). For SNL and SNPE-C the SBIBM default of distributing these simulations uniformly across R = 10 rounds is used, but to ease comparison we also run SNL and SNPE-C with the same number of rounds used for Se MPLE, which is R = 4, except for the Bernoulli GLM model where R = 2 suffices. Also, note that the N retained draws via MH are N post-burnin draws, where the burnin consists of 100 iterations. |