Classification under Nuisance Parameters and Generalized Label Shift in Likelihood-Free Inference

Authors: Luca Masserano, Alexander Shen, Michele Doro, Tommaso Dorigo, Rafael Izbicki, Ann B. Lee

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate its performance on two challenging scientific problems in biology and astroparticle physics with data from realistic mechanistic models.
Researcher Affiliation Academia 1Department of Statistics and Data Science, Carnegie Mellon University, Pittsburgh, USA 2Machine Learning Department, Carnegie Mellon University, Pittsburgh, USA 3Department of Physics and Astronomy, Universit a di Padova, Padova, Italy 4Istituto Nazionale di Fisica Nucleare, Sezione di Padova, Italy 5Lulea Techniska Universitet, Lulea, Sweden 6Universal Scientific Education and Research Network, Italy 7Department of Statistics, Universidade Federal de S ao Carlos, S ao Paulo, Brazil.
Pseudocode Yes Algorithm 1 Nuisance-aware prediction sets; Algorithm 2 Learning the Rejection Probability Function
Open Source Code Yes A flexible implementation of NAPS is available at https://github.com/lee-group-cmu/lf2i.
Open Datasets Yes We use data from the recently proposed sc Design3 simulator (Song et al., 2023), with reference data taken from the PBMC Systematic Comparative Analysis (Ding et al., 2019).
Dataset Splits Yes In total, we have available 80,000 samples which we divide into train (60%), calibration (35%) and test (5%) sets.
Hardware Specification Yes All computations were performed on a Mac Book Pro M1Pro with 16 GB of RAM.
Software Dependencies No The paper mentions using 'Cat Boost (Prokhorenkova et al., 2018)' and 'Sci Py (Virtanen et al., 2020)'. While SciPy 1.0 is implied by the citation, Cat Boost does not have a specific version number stated, and the question requires specific version numbers for multiple key components.
Experiment Setup No The paper describes the data splitting (e.g., 60% train, 35% calibration, 5% test) but does not provide specific hyperparameter values like learning rate, batch size, number of epochs, or optimizer settings used for training the models (e.g., CatBoost classifiers).