Key Feature Replacement of In-Distribution Samples for Out-of-Distribution Detection
Authors: Jaeyoung Kim, Seo Taek Kong, Dongbin Na, Kyu-Hwan Jung
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | an exhaustive comparison with state-of-the-art algorithms demonstrates KIRBY s competitiveness on various benchmarks. Numerical experiments confirm that our procedure indeed generates surrogate OOD data close to ID examples. Accordingly, a rejection network trained on this construction outperforms state-of-the-art OOD detection algorithms on most benchmarks. |
| Researcher Affiliation | Collaboration | Jaeyoung Kim*1, Seo Taek Kong* 2, Dongbin Na1, Kyu-Hwan Jung 3 1 VUNO, Inc. 2 University of Illinois, Urbana-Champaign 3 Samsung Advanced Institute for Health Sciences and Technology, Sungkyunkwan University |
| Pseudocode | No | The paper includes a 'Diagram of KIRBY' in Figure 4, but this is a visual diagram, not structured pseudocode or an algorithm block. |
| Open Source Code | Yes | Code is available at https://github.com/vuno/KIRBY |
| Open Datasets | Yes | Datasets Following (Liang, Li, and Srikant 2018; Liu et al. 2020b), we test OOD detection performance using CIFAR-10 and CIFAR-100 as ID sets and the following OOD sets: SVHN (Netzer et al. 2011). ... Textures (Cimpoi et al. 2014) ... LSUN-crop & LSUN-resize ... Place-365 (Zhou et al. 2017) ... i SUN (Xu et al. 2015) ... STL-10 (Coates, Ng, and Lee 2011) |
| Dataset Splits | No | The paper mentions using CIFAR-10 and CIFAR-100 as ID sets and various OOD datasets for testing, but it does not specify the train/validation/test splits (e.g., percentages or exact counts) for the ID datasets used for training and evaluating the models. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory specifications) used for running the experiments. It mentions an 'Intel(R) Xeon(R) Gold 5220 CPU' only in the context of latency comparison for inpainting methods, not for the overall experimental setup. |
| Software Dependencies | No | The paper mentions network architectures like Wide Res Net-40-2, Dense Net-BC, and Res Net-34, but it does not specify the version numbers of any software dependencies such as deep learning frameworks (e.g., PyTorch, TensorFlow), programming languages (e.g., Python), or other libraries. |
| Experiment Setup | Yes | Training converges in 5 epochs using SGD-momentum with the initial learning rate of 0.01 and weight decay 5 10 4. The rejection network is optimized with cross-entropy loss. The threshold parameter λ is set as 0.3, and the augmentation that replaces an image with a patch is applied to each sample with probability 0.5. |