IPMix: Label-Preserving Data Augmentation Method for Training Robust Classifiers

Authors: Zhenglin Huang, Xiaoan Bao, Na Zhang, Qingqi Zhang, Xiao Tu, Biao Wu, Xi Yang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that IPMix outperforms state-of-the-art corruption robustness on CIFAR-C and Image Net-C. In addition, we show that IPMix also significantly improves the other safety measures, including robustness to adversarial perturbations, calibration, prediction consistency, and anomaly detection, achieving state-of-the-art or comparable results on several benchmarks.
Researcher Affiliation Academia Zhenglin Huang1, Xianan Bao1, Na Zhang1 , Qingqi Zhang2, Xiaomei Tu3, Biao Wu1, and Xi Yang4 1School of Artificial Intellienge, Zhejiang Sci-Tech University 2Yamaguchi University 3ZGUC 4University of Science and Technology of China, Hefei, China
Pseudocode Yes The algorithm of IPMix is summarized in Appendix D.
Open Source Code Yes Code is available at https://github.com/hzlsaber/IPMix.
Open Datasets Yes We present the evaluation results of IPMix for image classification on three datasets CIFAR-10, CIFAR-100 [65], and Image Net [66] across various models.
Dataset Splits No The paper states 'Please refer to Appendix A for more details about the training configurations.' and mentions 'validation' in the context of calibrated uncertainty estimates, but does not explicitly provide specific dataset split percentages or sample counts for training, validation, and test sets in the main text.
Hardware Specification No The paper does not explicitly describe the specific hardware used to run its experiments, such as exact GPU or CPU models.
Software Dependencies No The paper mentions using an 'SGD optimizer' but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) needed to replicate the experiment.
Experiment Setup Yes We utilize SGD optimizer with an initial learning rate of 0.01 to train Res Net-50 for 180 epochs following a cosine decay schedule. We experiment with different backbone architectures on CIFAR-10 and CIFAR-100, including 40-4 Wide Res Net [67], 28-10 Wide Res Net, Res Ne Xt-29 [68], and Resnet-18 [69].