Leveraging Self-Consistency for Data-Efficient Amortized Bayesian Inference

Authors: Marvin Schmitt, Desi R. Ivanova, Daniel Habermann, Ullrich Koethe, Paul-Christian Bürkner, Stefan T. Radev

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Empirical Evaluation We evaluate our self-consistent estimator across a range of synthetic tasks and real-world problems.
Researcher Affiliation Academia 1University of Stuttgart, Germany 2University of Oxford, UK 3TU Dortmund University, Germany 4Heidelberg University, Germany 5Rensselaer Polytechnic Institute, USA.
Pseudocode Yes Algorithm 1 Self-consistency loss for finite training. {I}: likelihood-based with analytic likelihood {II}: simulation-based with approximate likelihood
Open Source Code Yes Code Availability We provide reproducible code in the open repository at https://github.com/marvinschmitt/ self-consistency-abi
Open Datasets Yes As a scientific real-world example, we apply our method to an experimental data set in biology (Silk et al., 2011).
Dataset Splits No No specific details on dataset split percentages, counts, or methodology for training/validation/test sets were provided. While 'posterior loss on a separate validation dataset' is mentioned in Appendix E, the exact details of this split (e.g., size, methodology) are not specified.
Hardware Specification No The paper does not provide specific details on the hardware used for experiments, such as GPU/CPU models, memory, or cloud instance specifications.
Software Dependencies No The paper mentions software like 'Stan', 'neural spline flow', 'Deep Set', 'tensorflow_probability', and 'scipy.stats', but does not provide specific version numbers for these software dependencies, only citations to their original papers.
Experiment Setup Yes The neural networks are trained for a total of 35 epochs with a batch size of 32 and an initial learning rate of 10-3. We choose a stepwise constant annealing schedule for the self-consistency weight λ such that λ = 0 for the first 5 epochs, and λ = 1 for the remaining 30 epochs.