A Unified View of Label Shift Estimation

Authors: Saurabh Garg, Yifan Wu, Sivaraman Balakrishnan, Zachary Lipton

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on synthetic data, MNIST, and CIFAR10 support our findings.
Researcher Affiliation Academia Saurabh Garg, Yifan Wu, Sivaraman Balakrishnan, Zachary C. Lipton Machine Learning Department, Department of Statistics and Data Science, Carnegie Mellon University {sgarg2,yw4,sbalakri,zlipton}@andrew.cmu.edu
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper provides links to the publicly available code for BBSE and RLLS (baselines used for comparison), but not for the MLLS methodology primarily described in this paper.
Open Datasets Yes We validate our results on synthetic data, MNIST, and CIFAR-10.
Dataset Splits Yes With CIFAR10 and MNIST, we split the full training set into two subsets: train and valid, and use the provided test set as is.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory specifications) used for running its experiments.
Software Dependencies No The paper mentions using 'Res Net-18' and 'pytorch-cifar' implementation but does not specify version numbers for any software dependencies.
Experiment Setup Yes For GMM, we control the shift in the label marginal for class 1 with a fixed target sample size of 1000. For multiclass problems -MNIST and CIFAR-10, we control the Dirichlet shift parameter with a fixed sample size of 5000. For GMM, we fix the label marginal for class 1 at 0.01 whereas for multiclass problems, MNIST and CIFAR-10, we fix the Dirichlet parameter to 0.1.