Online Feature Updates Improve Online (Generalized) Label Shift Adaptation

Authors: Ruihan Wu, Siddhartha Datta, Yi Su, Dheeraj Baby, Yu-Xiang Wang, Kilian Q. Weinberger

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiment In this section, we initiate OLS-OFU with three popular SSL techniques and empirically evaluate how OLS-OFU improves the original OLS methods on both online label shift and online generalized label shift on various datasets and shift patterns.
Researcher Affiliation Collaboration Ruihan Wu UC San Diego ruw076@ucsd.edu Siddhartha Datta University of Oxford siddhartha.datta@cs.ox.ac.uk Yi Su Google Deep Mind yisumtv@google.com Dheeraj Baby UC Santa Barbara dheeraj@ucsb.edu Yu-Xiang Wang UC San Diego yuxiangw@ucsd.edu Kilian Q. Weinberger Cornell University kilian@cornell.edu
Pseudocode Yes Algorithm 1 Online label shift adaptation with online feature updates (OLS-OFU).
Open Source Code Yes Code is released at https://github.com/dattasiddhartha/online-feature-updates-olsofu
Open Datasets Yes For online label shift, we evaluate the efficacy of our algorithm on CIFAR-10 [29], STL10 [12], CINIC [13], and Euro SAT [25]. For online generalized label shift, the offline train and validation sets are the CIFAR-10 images. The test unlabeled batches are drawn from CIFAR-10C [26], a benchmark with the same objects as CIFAR-10 but with various types of corruption.
Dataset Splits Yes For each dataset, we split the original train set into the offline train (i.e., D0) and validation sets (i.e., D 0) following a ratio of 4 : 1.
Hardware Specification No The paper mentions evaluating methods with ResNet18 and discusses time costs, but does not specify the hardware (e.g., GPU model, CPU, memory) used for these experiments.
Software Dependencies No The paper refers to using PyTorch (implicitly for neural networks) and ResNet18, but does not specify version numbers for any software dependencies like Python, PyTorch, or other libraries.
Experiment Setup Yes We experiment with T = 1000 and batch size B = 10 at each time step, following Baby et al. [6]. The frequency parameter τ is fixed as 100 for most experiments unless we particularly mention it. The seed used to train our model is 4242, and we train an additional 4 models on seeds 4343, 4545, 4646, 4747.