Learning to Re-weight Examples with Optimal Transport for Imbalanced Classification

Authors: Dandan Guo, Zhuo Li, meixi zheng, He Zhao, Mingyuan Zhou, Hongyuan Zha

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on image, text and point cloud datasets demonstrate that our proposed re-weighting method has excellent performance, achieving state-of-the-art results in many cases and providing a promising tool for addressing the imbalanced classification issue.
Researcher Affiliation Academia Dandan Guo 1,2, Zhuo Li3,4, Meixi Zheng5, He Zhao 6, Mingyuan Zhou7, Hongyuan Zha1,8 1School of Data Science, The Chinese University of Hong Kong, Shenzhen 2 Institute of Robotics and Intelligent Manufacturing 3School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen 4 Shenzhen Research Institute of Big Data 5Xidian University 6CSIRO s Data61 7The University of Texas at Austin 8Shenzhen Institute of Artificial Intelligence and Robotics for Society
Pseudocode Yes Algorithm 1 Workflow about our re-weighting method for optimizing θ and w.
Open Source Code Yes The code has been made available at https://github.com/Dandan Guo1993/reweight-imbalance-classification-with-OT.
Open Datasets Yes We evaluate our method on CIFAR-LT-10, CIFAR-LT-100, Image Net-LT and Places-LT. We create CIFAR-LT-10 (CIFAR-LT-100) from CIFAR-10 (CIFAR-100)[38] by downsampling samples per class with IF {200, 100, 50, 20} [5, 13]. Image Net-LT is built from the classic Image Net with 1000 classes[39] and IF=1280/5 [5, 24]. Places-LT is created from Places-2 [40] with 365 classes and IF=4980/5 [4, 24].
Dataset Splits Yes Besides, consider a small balanced meta set Dmeta = {(xj, yj)}M j=1, where M is the amount of total samples and M N. ... We randomly select 10 training images per class as meta set [5]; see more details in Appendix B.
Hardware Specification Yes Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix B.
Software Dependencies No The paper mentions software like 'ResNet-32' and 'BERT (base, uncased) model' but does not provide specific version numbers for any libraries or frameworks used in the implementation (e.g., PyTorch version, TensorFlow version).
Experiment Setup Yes For a fair comparison, we use Res Net-32 [41] as the backbone on CIFAR-LT-10 and CIFAR-LT-100. Following Li et al. [5], at stage 1, we use 200 epochs, set the learning rate α of θ as 0.1, which is decayed by 1e-2 at the 160th and 180th epochs. At stage 2, we use 40 epochs, set α as 2e-5 and learning rate β of weights as 1e-3. We use the SGD optimizer with momentum 0.9, weight decay 5e-4 and set the batch size as 16.