Multi-Class Uncertainty Calibration via Mutual Information Maximization-based Binning

Authors: Kanil Patel, William H. Beluch, Bin Yang, Michael Pfeiffer, Dan Zhang

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental I-Max is evaluated according to multiple performance metrics, including accuracy, ECE, Brier and NLL, and compared against benchmark calibration methods across multiple datasets and trained classifiers.
Researcher Affiliation Collaboration 1Bosch Center for Artificial Intelligence, Renningen, Germany 2Institute of Signal Processing and System Theory, University of Stuttgart, Stuttgart, Germany
Pseudocode Yes Algo. 1 in the appendix for pseudocode
Open Source Code Yes Code available at https://github.com/boschresearch/imax-calibration
Open Datasets Yes We evaluate post-hoc calibration methods on four benchmark datasets, i.e., Image Net (Deng et al., 2009), CIFAR 10/100 (Krizhevsky, 2009) and SVHN (Netzer et al., 2011)
Dataset Splits Yes We perform class-balanced random splits of the data test set, unless stated otherwise: the calibration and evaluation set sizes are both 25k for Image Net, and 5k for CIFAR10/100.
Hardware Specification No No explicit details on specific GPU models, CPU models, or other hardware specifications used for running experiments were found.
Software Dependencies No No specific software versions (e.g., Python 3.x, PyTorch 1.x) or library version numbers (e.g., Numpy X.Y, sklearn X.Y) are provided.
Experiment Setup Yes All scaling methods use the Adam optimizer with batch size 256 for CIFAR and 4096 for Image Net. The learning rate was set to 10 3 for temperature scaling Guo et al. (2017) and Platt scaling Platt (1999), 0.0001 for vector scaling Guo et al. (2017) and 10 5 for matrix scaling Guo et al. (2017).