A Closer Look at Accuracy vs. Robustness
Authors: Yao-Yuan Yang, Cyrus Rashtchian, Hongyang Zhang, Russ R. Salakhutdinov, Kamalika Chaudhuri
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments with robustness methods, we argue that the gap between theory and practice arises from two limitations of current methods: either they fail to impose local Lipschitzness or they are insufficiently generalized. We explore combining dropout with robust training methods and obtain better generalization. |
| Researcher Affiliation | Academia | 1University of California, San Diego 2Toyota Technological Institute at Chicago 3Carnegie Mellon University |
| Pseudocode | No | The paper describes mathematical functions and procedures in text and equations (e.g., Section 4.1, Measuring Local Lipschitzness in Section 5.1) but does not include any clearly labeled "Pseudocode" or "Algorithm" blocks or structured code-like steps. |
| Open Source Code | Yes | Code available at https://github.com/yangarbiter/robust-local-lipschitz. |
| Open Datasets | Yes | We consider four datasets: MNIST, CIFAR-10, SVHN and Restricted Image Net (Res Image Net)... [26], [24], [33]. |
| Dataset Splits | No | The paper mentions "Train-Train Separation" and "Test-Train Separation" in Table 1 and accompanying text, and discusses "generalization gaps" between training and test accuracies. However, it does not explicitly define or specify a "validation" dataset split for hyperparameter tuning or early stopping. |
| Hardware Specification | No | The paper does not specify the hardware used for running the experiments (e.g., specific GPU/CPU models, memory, or cloud resources). |
| Software Dependencies | No | The paper does not list specific version numbers for software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | For SVHN, we use a dropout rate of 0.5 and for CIFAR-10 a rate of 0.2. More experimental details are provided in the Appendix B. Learning rates for all datasets are tuned with ADAM [23] optimizer with an initial learning rate of 10−3 and decayed by 0.1 at epochs 75 and 90. |