Direct Amortized Likelihood Ratio Estimation

Authors: Adam D. Cobb, Brian Matejek, Daniel Elenius, Anirban Roy, Susmit Jha

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

Reproducibility Variable Result LLM Response
Research Type Experimental We benchmark our new ratio estimator and compare it to ratio estimators in the current literature. We show that our new ratio estimator often outperforms these previous approaches. As a further contribution, we introduce a new derivative estimator for likelihood ratio estimators that enables us to compare likelihood-free Hamiltonian Monte Carlo (HMC) with random-walk Metropolis-Hastings (MH). We show that HMC is equally competitive, which has not been previously shown. Finally, we include a novel real-world application of SBI using our neural ratio estimator to design a quadcopter.
Researcher Affiliation Industry Neuro-Symbolic Computing and Intelligence (Nu SCI) Research Group Computer Science Laboratory, SRI International adam.cobb@sri.com
Pseudocode Yes Algorithm 1: Optimization of dφ(x, , 0)
Open Source Code Yes Code is available at https://github.com/SRI-CSL/dnre.
Open Datasets Yes For the next experiment, we work with the two moons dataset, which is a typical low-dimensional benchmark within the SBI literature (Greenberg, Nonnenmacher, and Macke 2019; Lueckmann et al. 2021). ... In this section we perform a quantitative comparison between the three ratio estimators, where we use the SBI Benchmark examples of Lueckmann et al. (2021) (see Appendix B). ... Finally, we include a novel real-world application of SBI using our neural ratio estimator to design a quadcopter. Code is available at https://github.com/SRI-CSL/dnre.
Dataset Splits Yes We use a training set of size 10,000 and a validation set of size 5,000. ... We train all three neural ratio estimators using a five layer fully connected architecture with 128 hidden units per layer. We use a learning rate of 0.001 and train all models for 2,000 epochs using a batch size of 512 with 3,592 training points and 898 validation points.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory, or specific computing cluster names) used for running the experiments.
Software Dependencies No The baseline approaches are implemented using the LAMPE Python package (Rozet 2022). While a package is mentioned, specific version numbers for LAMPE or Python are not provided.
Experiment Setup Yes All models were trained over 1000 epochs with neural network architectures of three layers of 64 units. We use a training set of size 10,000 and a validation set of size 5,000. ... All estimators have an architecture of five layers of 64 units, using the Exponential Linear Unit non-linearity between layers. For all approaches we applied the same grid search over both the learning rate and standard deviation of the proposal distribution for the random-walk MH sampling scheme. ... We train all three neural ratio estimators using a five layer fully connected architecture with 128 hidden units per layer. We use a learning rate of 0.001 and train all models for 2,000 epochs using a batch size of 512 with 3,592 training points and 898 validation points.