Preconditioned Neural Posterior Estimation for Likelihood-free Inference
Authors: Xiaoyu Wang, Ryan P. Kelly, David J Warne, Christopher Drovandi
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present comprehensive empirical evidence that this melding of neural and statistical SBI methods improves performance over a range of examples including a motivating example involving a complex agent-based model applied to real tumour growth data. We present five benchmarking tasks (two from Lueckmann et al. (2021) and three from the SBI literature) and two additional examples, including our motivating example, where SNPE, perhaps surprisingly, does not produce highly accurate posterior distributions. |
| Researcher Affiliation | Academia | Xiaoyu Wang EMAIL School of Mathematical Sciences Centre for Data Science Queensland University of Technology Ryan P. Kelly EMAIL School of Mathematical Sciences Centre for Data Science Queensland University of Technology David J. Warne EMAIL School of Mathematical Sciences Centre for Data Science Queensland University of Technology Christopher Drovandi EMAIL School of Mathematical Sciences Centre for Data Science Queensland University of Technology |
| Pseudocode | Yes | Algorithm 1 Precondition SNPE. Algorithm 2 Adaptive SMC ABC |
| Open Source Code | Yes | The code to reproduce the results has been included as supplementary material. |
| Open Datasets | Yes | We use the sbibm package (Lueckmann et al., 2021) for the two-moon and SLCP simulators, and the ELFI package (Lintusaari et al., 2018) for the other benchmark simulators. We calibrate to two real-world pancreatic cancer datasets Wade (2019), which describe tumor growth as time series data. The toad movement model proposed in (Marchand et al., 2017) is an individual-based model for simulating the dispersal of Fowler s toads (Anaxyrus fowleri). |
| Dataset Splits | No | The paper discusses the number of simulations generated for training (e.g., 'For SNPE, we use 10k samples for each round of training', 'around 30k model simulations' for preconditioning), but does not specify explicit training/validation/test splits of a fixed dataset with percentages, counts, or references to predefined splits. |
| Hardware Specification | No | We thank the computational resources provided by QUT’s High Performance Computing and Research Support Group (HPC). |
| Software Dependencies | No | For the unconditional density estimator, we employ unconditional normalizing flows using the Pyro package Bingham et al. (2019)... as implemented in the Scikit-learn package Pedregosa et al. (2011). For APT and TSNPE, we use the implementation of the sbi package Tejero-Cantero et al. (2020) with default settings. (Specific version numbers for these packages are not provided.) |
| Experiment Setup | Yes | We set the tuning parameters as a = 0.5, c = 0.01 and use 1k particles for the algorithm. As the stopping rule, we set the target MCMC acceptance rate at 10%, unless otherwise specified. Regarding the neural networks, we use four fully-connected layers and set the count bins to 16. We use five coupling layers, with each coupling layer using a multilayer perceptron of two layers with 50 hidden units. The flow is trained using the Adam optimizer with a learning rate of 5 10 4 and a batch size of 256. Flow training is stopped when either the validation loss, calculated on 10% of the samples, has not improved over 50 epochs or when the limit of 500 epochs is reached. |