Multi-model Ensemble Conformal Prediction in Dynamic Environments

Authors: Erfan Hajihashemi, Yanning Shen

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, the performance of the proposed method, SAMOCP, is assessed within the context of classification tasks. We conduct a comprehensive comparison with recently proposed methods in online conformal prediction for dynamic environments within classification tasks.
Researcher Affiliation Academia Erfan Hajihashemi Department of Electrical Engineering & Computer Science University of California, Irvine ehajihas@uci.edu Yanning Shen Department of Electrical Engineering & Computer Science University of California, Irvine yannings@uci.edu
Pseudocode Yes Algorithm 1 Multi-model Ensemble Online Conformal Prediction (MOCP)
Open Source Code Yes Codes are available at hyperrefhttps://github.com/erfanhajihashemi/Multi-model-Ensemble-Conformal-Predictionin-Dynamic-Environments.
Open Datasets Yes Dataset: We utilize corrupted versions of CIFAR-10 and CIFAR-100 [Krizhevsky, 2009], known as CIFAR-10C and CIFAR-100C [Hendrycks and Dietterich, 2019]. ... All real datasets are downloaded from the Zenodo repository.
Dataset Splits No The paper mentions data being split into 'batches of 500 data samples each' and uses a 'calibration dataset' which is an 'evolving calibration dataset' in the online setting, but does not provide explicit train/validation/test dataset splits with percentages or sample counts for reproducibility in a typical static ML setup.
Hardware Specification Yes All experiments were performed on a workstation with NVIDIA RTX A4000 GPU.
Software Dependencies No The paper mentions various learning models (e.g., Res Net-50, Goog Le Net) and notes that codes are available on GitHub, but it does not specify software dependencies with version numbers (e.g., Python version, PyTorch/TensorFlow version, specific library versions).
Experiment Setup Yes For every experiment conducted on the synthetic dataset, CIFAR-10C, CIFAR-100C, parameters ϵ, σ, and η were selected through grid search, with values of 0.9, 140, and 0.05, respectively. The hyperparameters ξ and kreg are set to 0.02 and 5 for CIFAR-100C, and 0.1 and 1 for Cifar-10C, respectively.