Towards Reliable Simulation-Based Inference with Balanced Neural Ratio Estimation

Authors: Arnaud Delaunoy, Joeri Hermans, François Rozet, Antoine Wehenkel, Gilles Louppe

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate BNRE on a wide variety of tasks and show that it produces conservative posterior surrogates on all tested benchmarks and simulation budgets. We start by providing an extensive validation of BNRE on a broad range of benchmarks demonstrating that the proposed method alleviates the problem.
Researcher Affiliation Academia Arnaud Delaunoy University of Liège a.delaunoy@uliege.be Joeri Hermans* Unaffiliated joeri@peinser.com François Rozet University of Liège francois.rozet@uliege.be Antoine Wehenkel University of Liège antoine.wehenkel@uliege.be Gilles Louppe University of Liège g.louppe@uliege.be
Pseudocode Yes Algorithm 1 Training algorithm for Balanced Neural Ratio Estimation (BNRE). Inputs: Implicit generative model p(x | ϑ) (simulator) and prior p(ϑ) Outputs: Approximate classifier ˆdψ(ϑ, x) parameterized by ψ hyper-parameters: Balancing condition strength λ (default = 100) and batch-size n Sample data from the joint {ϑi, xi p(ϑ, x), yi = 1}n/2 i=1 Sample data from the marginals {ϑi, xi p(ϑ)p(x), yi = 0}n i=n/2+1 L[ ˆdψ] = 1 n Pn i=1 yi log ˆdψ(ϑi, xi) + (1 yi) log(1 ˆdψ(ϑi, xi)) B[ ˆdψ] = 2 n Pn/2 i=1 ˆdψ(ϑi, xi) + 2 n Pn i=n/2+1 ˆdψ(ϑi, xi) ψ = minimizer_step(params=ψ, loss=L[ ˆdψ] + λ(B[ ˆdψ] 1)2) until convergence return ˆdψ(ϑ, x).
Open Source Code Yes Code is available at https://github.com/montefiore-ai/balanced-nre.
Open Datasets Yes We evaluate the expected coverage of posterior estimators produced by both NRE and BNRE on various problems. Those benchmarks cover a diverse set of problems from particle physics (Weinberg), epidemiology (Spatial SIR), queueing theory (M/G/1), population dynamics (Lotka Volterra, and astronomy (Gravitational Waves). They are representative of real scientific applications of simulation-based inference. A more detailed description of the benchmarks can be found in Appendix C.
Dataset Splits No No explicit mention of a separate validation dataset split. The paper mentions "simulation budgets" for training and evaluates on "10000 unseen samples".
Hardware Specification No No specific hardware details like GPU/CPU models or memory specifications are provided. The paper only mentions 'Computational resources have been provided by the Consortium des Équipements de Calcul Intensif (CÉCI)'.
Software Dependencies No The paper describes neural network architectures and training hyperparameters (Appendix D) but does not list specific software dependencies with version numbers (e.g., PyTorch 1.x, Python 3.x).
Experiment Setup Yes The architectures and hyper-parameters used for each problem are defined in Appendix D. (e.g., learning rate 10^-3, batch size 2^10, 500 epochs, Adam optimizer).