Revive Re-weighting in Imbalanced Learning by Density Ratio Estimation

Authors: JIAAN LUO, Feng Hong, Jiangchao Yao, Bo Han, Ya Zhang, Yanfeng Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted on various large scale benchmark datasets validate the effectiveness of our method. Results demonstrate substantial improvements in generalization capabilities, particularly under severely imbalanced conditions.
Researcher Affiliation Academia Jiaan Luo1,3 Feng Hong1 Jiangchao Yao1,3 Bo Han4 Ya Zhang2,3 Yanfeng Wang2,3 1Cooperative Medianet Innovation Center, Shanghai Jiao Tong University 2School of Artificial Intelligence, Shanghai Jiao Tong University 3Shanghai Artificial Intelligence Laboratory 4Hong Kong Baptist University
Pseudocode Yes The framework and pseudo-code of our method are shown in Appendix C.2 and Appendix C.3. We provide the pseudo-code of RDR to demonstrate the process of implementing our method in detail, as shown in Algorithm 1. In addition, we also provide pseudo-code that combines our method with the SAM method, as shown in Algorithm 2.
Open Source Code Yes The code is available here.
Open Datasets Yes We conduct experiments on four major long-tailed datasets, CIFAR-10-LT, CIFAR-100-LT, Image Net-LT [Liu et al., 2019] and Places-LT [Liu et al., 2019]. CIFAR-10-LT and CIFAR-100-LT are two datasets sampled from the original CIFAR [Krizhevsky et al., 2009] dataset with a total of 10 and 100 classes, respectively.
Dataset Splits No The paper mentions 'training dataset' and 'test sets' but does not explicitly specify a validation dataset split or how it was used.
Hardware Specification Yes Experiments based on CIFAR-10-LT and CIFAR-100-LT are carried out on NVIDIA Ge Force RTX 3090 GPUs, while experiments based on Image Net-LT and Places-LT are carried out on NVIDIA A100 GPUs.
Software Dependencies Yes Our code is implemented with Pytorch 1.12.1.
Experiment Setup Yes We train each model with batch size of 128 (for CIFAR-10-LT and CIFAR-100-LT) / 256 (for Places-LT and Image Net-LT), SGD optimizer with momentum of 0.9, weight decay of 0.0002. The initial learning rate is set to 0.1, with cosine learning-rate scheduling along training.