Conformal Prediction for Class-wise Coverage via Augmented Label Rank Calibration

Authors: Yuanjie Shi, Subhankar Ghosh, Taha Belkhouja, Jana Doppa, Yan Yan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments on multiple real-world datasets demonstrate that RC3P achieves class-wise coverage and 26.25% reduction in prediction set sizes on average.
Researcher Affiliation Academia Yuanjie Shi Washington State University Subhankar Ghosh Washington State University Taha Belkhouja Washington State University Janardhan Rao Doppa Washington State University Yan Yan Washington State University
Pseudocode Yes Algorithm 1 RC3P Method for Class-Conditional CP
Open Source Code Yes The code is available at https://github.com/Yuanjie Sh/RC3P.
Open Datasets Yes We consider four datasets: CIFAR-10, CIFAR-100 [34], mini-Image Net [57], and Food-101 [8] by using the standard training and validation split.
Dataset Splits Yes Randomly split data into train Dtr and calibration Dcal and train the classifier f on Dtr.we randomly split the original (balanced) validation set into calibration samples and testing samples.
Hardware Specification No The paper mentions 'Res Net-20 [27]' as the main architecture and 'LDAM proposed by [10]' as the training algorithm, but does not specify any hardware details such as GPU/CPU models or memory used for experiments.
Software Dependencies No The paper mentions 'PyTorch [44]' but does not provide a specific version number for it or any other software components.
Experiment Setup Yes To handle imbalanced data, we employ the training algorithm LDAM proposed by [10] that assigns different margins to classes... where all models are trained with 200 epochs... learning rate 0.1, momentum 0.9, and weight decay 2e-4 for 200 epochs and 50 epochs. The batch size is 128. For Food-101, the batch size is 256... We set α = 0.1 as our main experiment setting and also report other experiment results of different α values (See Appendix C.7).