PDE+: Enhancing Generalization via PDE with Adaptive Distributional Diffusion
Authors: Yige Yuan, Bingbing Xu, Bo Lin, Liang Hou, Fei Sun, Huawei Shen, Xueqi Cheng
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of PDE+ is validated through extensive experimental settings, demonstrating its superior performance compared to state-of-the-art methods. Our code is available at https://github.com/yuanyige/pde-add. In this section, we empirically evaluates PDE+ through the following questions. Due to the space limitations, more comprehensive experiments including full results on corruptions and diffusion scale analysis are provided in Appendix E. (Q1) Does PDE+ improve generalization compared to SOTA methods on various benchmarks? (Q2) Does PDE+ learns appropriate diffusion distribution coverage? (Q3) Does PDE+ improve generalization beyond observed (training) distributions? Table 1 illustrates the results of PDE+ on CIFAR10(C), CIFAR100(C) and Tiny Image Net(C) compared to baselines. Table 2 illustrates the results of PDE+ on PACS datasets. |
| Researcher Affiliation | Academia | Yige Yuan1,2, Bingbing Xu1*, Bo Lin3, Liang Hou1, Fei Sun1, Huawei Shen1,2*, Xueqi Cheng1,2* 1CAS Key Laboratory of AI Safety & Security, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2University of Chinese Academy of Sciences, Beijing, China 3Department of Mathematics, National University of Singapore, Singapore |
| Pseudocode | Yes | The algorithmic pseudocode for both training and testing phase can be found in Algorithms 1 and 2 in Appendix D. |
| Open Source Code | Yes | Our code is available at https://github.com/yuanyige/pde-add. |
| Open Datasets | Yes | Our experiments primarily focus on two types of datasets: (i) The original and 15 shift corruption distributions provided by CIFAR-10(C), CIFAR-100(C) and Tiny-Image Net(C) (Krizhevsky, Hinton et al. 2009; Le and Yang 2015; Hendrycks and Dietterich 2019). (ii) The PACS dataset (Li et al. 2017) encompasses four different domains: photo, art, cartoon, and sketch. |
| Dataset Splits | No | The paper mentions training and test distributions, but does not explicitly provide details about training, validation, or test dataset splits (e.g., percentages, counts, or specific split files/methods). |
| Hardware Specification | No | The paper does not specify any hardware details such as GPU models, CPU types, or memory used for running the experiments. |
| Software Dependencies | No | The paper mentions software dependencies implicitly through the use of terms like 'Res Net' and 'Aug Mix', but does not provide specific version numbers for any libraries, frameworks (e.g., PyTorch, TensorFlow), or other software components. |
| Experiment Setup | No | The paper has a section '5 Experiments' and refers to 'Appendix F' for details on experiment settings, but the main body of the paper does not explicitly list specific experimental setup details such as hyperparameters (e.g., learning rate, batch size, epochs), optimizer settings, or model initialization details. |