DOS: Diverse Outlier Sampling for Out-of-Distribution Detection

Authors: Wenyu Jiang, Hao Cheng, MingCai Chen, Chongjun Wang, Hongxin Wei

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the superiority of DOS, reducing the average FPR95 by up to 25.79% on CIFAR-100 with TI-300K.
Researcher Affiliation Academia 1Department of Statistics and Data Science, Southern University of Science and Technology 2State Key Laboratory for Novel Software Technology, Nanjing University
Pseudocode Yes Algorithm 1: DOS: Diverse Outlier Sampling
Open Source Code Yes Code is available at: https://github.com/lygjwy/DOS.
Open Datasets Yes Datasets. We conduct experiments on CIFAR100 (Krizhevsky & Hinton, 2009) as common benchmark and Image Net-10 (Ming et al., 2022a) as large-scale benchmark. For CIFAR100, a down-sampled version of Image Net (Image Net-RC) (Chen et al., 2021) is utilized as an auxiliary OOD training dataset. Additionally, we use the 300K random Tiny Images subset (TI-300K)1 as an alternative OOD training dataset, due to the unavailability of the original 80 Million Tiny Images2 in previous work (Hendrycks et al., 2019b). 1https://github.com/hendrycks/outlier-exposure
Dataset Splits No The paper mentions training and test datasets but does not explicitly provide details about a separate validation dataset split (e.g., specific percentages or sample counts) used for hyperparameter tuning or early stopping. While FPR95 at 95% TPR on ID data implies a threshold selection, it does not constitute a distinct validation *split*.
Hardware Specification Yes All the experiments are conducted on NVIDIA V100 and all methods are implemented with default parameters using Py Torch.
Software Dependencies No The paper states that methods are 'implemented with default parameters using Py Torch,' but it does not provide a specific version number for PyTorch or any other software dependencies.
Experiment Setup Yes The model is trained for 100 epochs using SGD with a momentum of 0.9, a weight decay of 0.0001, and a batch size of 64, for both ID and OOD training data. The initial learning rate is set as 0.1 and decays by a factor of 10 at 75 and 90 epochs. Without tuning, we keep the number of the clustering center the same as the batch size. The model is fine-tuned for 10 epochs using SGD with a momentum of 0.9, a learning rate of 0.001, and a weight decay of 0.00001.