CSI: Novelty Detection via Contrastive Learning on Distributionally Shifted Instances

Authors: Jihoon Tack, Sangwoo Mo, Jongheon Jeong, Jinwoo Shin

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We verify the effectiveness of CSI under various environments of detecting OOD, including unlabeled one-class, unlabeled multi-class, and labeled multi-class settings. To our best knowledge, we are the first to demonstrate all three settings under a single framework. Overall, CSI outperforms the baseline methods for all tested datasets. ... 3 Experiments
Researcher Affiliation Academia Jihoon Tack , Sangwoo Mo , Jongheon Jeong , Jinwoo Shin Graduate School of AI, KAIST School of Electrical Engineering, KAIST {jihoontack,swmo,jongheonj,jinwoos}@kaist.ac.kr
Pseudocode No The paper provides mathematical formulations for its methods but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Code and pre-trained models are available at https://github.com/alinlab/CSI.
Open Datasets Yes We run our experiments on three datasets, following the prior work [15, 25, 2]: CIFAR-10 [33], CIFAR-100 labeled into 20 super-classes [33], and Image Net-30 [25] datasets. ... We compare CSI on two in-distribution datasets: CIFAR-10 [33] and Image Net-30 [25]. We consider the following datasets as out-of-distribution: SVHN [48], resized LSUN and Image Net [39], CIFAR100 [33], and linearly-interpolated samples of CIFAR-10 (Interp.) [11] for CIFAR-10 experiments, and CUB-200 [67], Dogs [29], Pets [51], Flowers [49], Food-101 [3], Places-365 [75], Caltech256 [18], and DTD [8] for Image Net-30.
Dataset Splits No The paper describes the datasets used and the evaluation metrics, and references prior work for setups, but it does not explicitly provide the specific training, validation, and test dataset splits (e.g., percentages or sample counts) within the main text or with direct citations to papers defining those splits for all datasets.
Hardware Specification No The paper mentions using a Res Net-18 architecture but does not specify any hardware details such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes Setup. We use Res Net-18 [20] architecture for all the experiments. For data augmentations T , we adopt those used by Chen et al. [5]: namely, we use the combination of Inception crop [64], horizontal flip, color jitter, and grayscale. For shifting transformations S, we use the random rotation 0 , 90 , 180 , 270 unless specified otherwise... By default, we train our models from scratch with the training objective in (5)... We simply set λ = 1 for all our experiments.