Interpolation for Robust Learning: Data Augmentation on Wasserstein Geodesics
Authors: Jiacheng Zhu, Jielin Qiu, Aritra Guha, Zhuolin Yang, Xuanlong Nguyen, Bo Li, Ding Zhao
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental validations of the proposed strategy on four datasets, including CIFAR-100 and Image Net, establish the efficacy of our method, e.g., our method improves the baselines certifiable robustness on CIFAR10 up to 7.7%, with 16.8% on empirical robustness on CIFAR-100.5. Experiments and Discussion We evaluate our proposed method in terms of both empirical robustness and certified robustness (Cohen et al., 2019a) on the MNIST (Le Cun et al., 1998), CIFAR-10 and CIFAR-100 (Krizhevsky, 2009) dataset, and samples from Image Net(64 64) dataset (Deng et al., 2009; Le & Yang, 2015). |
| Researcher Affiliation | Collaboration | 1Carnegie Mellon University 2Data Science & AI Research, AT&T Chief Data Office 3University of Illinois at Urbana Champaign 4University of Michigan Ann Arbor. |
| Pseudocode | Yes | A pseudocode Algo.(2) is attached in the Appendix.Algorithm 1 The data augmentation algorithmAlgorithm 2 Sinkhorn Barycenter |
| Open Source Code | No | The paper does not include any explicit statements about releasing source code or links to a code repository. |
| Open Datasets | Yes | 5. Experiments and Discussion We evaluate our proposed method in terms of both empirical robustness and certified robustness (Cohen et al., 2019a) on the MNIST (Le Cun et al., 1998), CIFAR-10 and CIFAR-100 (Krizhevsky, 2009) dataset, and samples from Image Net(64 64) dataset (Deng et al., 2009; Le & Yang, 2015). |
| Dataset Splits | No | The paper does not explicitly state specific training, validation, and test splits with percentages or sample counts. It mentions using data augmentation to double the training set size but does not detail how the overall dataset was partitioned for training, validation, and testing. |
| Hardware Specification | Yes | The experiments are carried out on several NVIDIA RTX A6000 GPUs and two NVIDIA Ge Force RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions software components like 'VAE' and 'Res Nets' but does not provide specific version numbers for any software libraries, frameworks, or programming languages used (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | Typically, we use data augmentation to double the size of the training set ma = 1 at first and use the regularization with a fixed weight αreg = 5.0 during the training, as implicit data augmentation. For the MNIST samples, we train a Le Net (Le Cun et al., 1998) classifier with a learning rate lr = 0.01 for 90 epochs. For the CIFAR dataset, we use the Res Net110 (He et al., 2016) for the certifiable task on CIFAR10 and Preact Res Net18 on CIFAR-100. The Sinkhorn entropic coefficient is chosen to be ϵ = 0.01. |