AdAUC: End-to-end Adversarial AUC Optimization Against Long-tail Problems
Authors: Wenzheng Hou, Qianqian Xu, Zhiyong Yang, Shilong Bao, Yuan He, Qingming Huang
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, the extensive experimental results show the performance and robustness of our algorithm in three long-tail datasets. |
| Researcher Affiliation | Collaboration | 1Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS, Beijing, China. 2School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing, China. 3State Key Laboratory of Information Security, Institute of Information Engineering, CAS, Beijing, China. 4School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China. 5Alibaba Group, Beijing, China 6Key Laboratory of Big Data Mining and Knowledge Management, Chinese Academy of Sciences, Beijing, China. 7Artificial Intelligence Research Center, Peng Cheng Laboratory, Shenzhen, China. |
| Pseudocode | Yes | Algorithm 1 Adversarial Training for AUC Optimization |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | Binary MNIST-LT Dataset. We construct a long-tail MNIST dataset from original MNIST dataset (Le Cun et al., 1998) in the same way as CIFAR-10-LT, where the ratio of positive class size to negative class size ρ ≈ 1 : 9. |
| Dataset Splits | No | The paper describes how the long-tail datasets were constructed based on CIFAR-10, CIFAR-100, and MNIST, but it does not explicitly state the specific percentages or counts for training, validation, and test splits. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions optimizers (SGD momentum, SGDA momentum) and model architecture (Wide ResNet-28), but it does not specify versions for any software dependencies like deep learning frameworks (e.g., PyTorch, TensorFlow) or Python. |
| Experiment Setup | Yes | The adversarial training is applied with the maximal permutation ϵ of 8/255 and a step size of 2/255. The max number of iterations K is set as 10. For CE, we use SGD momentum optimizer, while for ours, we use SGDA momentum optimizer. The initial learning rate ηw is set as 0.01 with decay 5 × 10−4, and the batch size is 128. And the initial learning rate ηα is set as 0.1. In the training process, we adopt a learning rate step decay schedules, which cut the learning rate by a constant factor 0.001 every 30 constant number of epochs for all methods. |