Mining In-distribution Attributes in Outliers for Out-of-distribution Detection
Authors: Yutian Lei, Luping Ji, Pei Liu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the superiority of our framework to others. MVOL effectively utilizes both auxiliary OOD datasets and even wild datasets with noisy ID data. |
| Researcher Affiliation | Academia | School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China |
| Pseudocode | No | The paper includes definitions, propositions, assumptions, and theorems, along with mathematical equations, but it does not contain any clearly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | Code https://github.com/UESTC-nnLab/MVOL |
| Open Datasets | Yes | Datasets. (1) Following the common benchmarks in literature, we use CIFAR-10 and CIFAR-100 as ID datasets, and six diverse OOD test sets. The Tiny Images dataset used in (Hendrycks, Mazeika, and Dietterich 2019; Liu et al. 2020) for auxiliary outliers has been withdrawn due to its ethical wrong. Instead, we use 300K Random Images, a cleaned subset of the Tiny Images dataset by (Hendrycks, Mazeika, and Dietterich 2019), as done in (Katz-Samuels et al. 2022). |
| Dataset Splits | Yes | (2) For experiments with wild datasets, we partition CIFAR10 into 2 halves: one provides ID training data, and the other provides auxiliary noisy ID data. CIFAR-100 provides auxiliary outliers. The auxiliary ID and OOD data are mixed with varied noise levels α, i.e., 0, 0.05, 0.1, 0.3, 0.5, and the total images in auxiliary datasets are maintained at 50000. |
| Hardware Specification | No | The paper does not specify any particular hardware used for running the experiments, such as GPU models, CPU models, or memory. |
| Software Dependencies | No | The paper mentions using 'Wide Res Net' and 'cross entropy loss' but does not specify any software libraries or their version numbers. |
| Experiment Setup | Yes | Training Details. (1) We use the Wide Res Net with 40 layers and a widen factor of 2 for all experiments. (2) Involved two experimental settings: (i) Single model: single neural network is trained with one-hot labels and cross entropy loss. (ii) Ensemble distillation model: a single network is trained with the standard ensemble distillation algorithm in (Hinton, Vinyals, and Dean 2015), where soft labels are generated by ensemble models. |