Enhancing Class-Imbalanced Learning with Pre-Trained Guidance through Class-Conditional Knowledge Distillation

Authors: Lan Li, Xin-Chun Li, Han-Jia Ye, De-Chuan Zhan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on various imbalanced datasets demonstrate an average accuracy improvement of 7.4% using our method.
Researcher Affiliation Academia 1School of Artificial Intelligence, Nanjing University, China 2National Key Laboratory for Novel Software Technology, Nanjing University, China.
Pseudocode Yes Appendix B. Pseudo Code of ACCKD. Algorithm 1 Augmented Class-Conditional Knowledge Distillation (ACCKD)
Open Source Code Yes Our code is available at https://github.com/Lain810/CCKD.
Open Datasets Yes Dataset. We performed extensive experiments on three widely used imbalanced datasets: CIFAR-100-LT (Cui et al., 2019), Image Net-LT (Russakovsky et al., 2015), and i Naturalist 2018 (Horn et al., 2018).
Dataset Splits Yes It should be estimated using a validation set Dval that shares the same distribution as the test data... For CIFAR-100, following Cui et al. (2019), we created class imbalance by the imbalance ratio of Nmax/Nmin = 100. Image Net-LT consists of 115.8K images across 1000 classes, with a maximum of 1280 images and a minimum of 5 images per class.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or specific computing environments) used for running the experiments.
Software Dependencies No The paper mentions using Res Net-32 and Res Net-18 as backbone networks but does not specify any software libraries or their version numbers (e.g., PyTorch, TensorFlow, CUDA versions) required to reproduce the experiments.
Experiment Setup Yes We set the batch size to 128 and train for 200 epochs, using SGD as the optimizer. The initial learning rate of the optimizer is set to 0.1, with a momentum of 0.9 and a weight decay of 5 10 4. In terms of hyperparameters, we set the balancing parameters for the loss terms, α1 and α2, to 1 and 2, respectively. The temperature parameter for distillation, τ, is set to 2.