How to Exploit Hyperspherical Embeddings for Out-of-Distribution Detection?

Authors: Yifei Ming, Yiyou Sun, Ousmane Dia, Yixuan Li

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTS
Researcher Affiliation Collaboration Department of Computer Sciences, University of Wisconsin-Madison1 Meta2
Pseudocode Yes Algorithm 1: Pseudo-code of CIDER.
Open Source Code Yes Code is available at https://github.com/deeplearning-wisc/cider.
Open Datasets Yes we consider CIFAR-10 and CIFAR-100 (Krizhevsky et al., 2009) as in-distribution datasets. For OOD test datasets, we use a suite of natural image datasets including SVHN (Netzer et al., 2011), Places365 (Zhou et al., 2017), Textures (Cimpoi et al., 2014b), LSUN (Yu et al., 2015), and i SUN (Xu et al., 2015).
Dataset Splits No The paper mentions using CIFAR-10 and CIFAR-100 datasets which have standard train/test splits, and implies validation is used for hyperparameter tuning ("K can be tuned using a validation method"), but does not explicitly state the training/validation/test dataset splits in terms of percentages, counts, or a specific method for creating them from the full datasets.
Hardware Specification Yes We run all the experiments on NVIDIA Ge Force RTX-2080Ti GPU for small to medium batch size and on NVIDIA A100 GPU for large batch size and larger network encoder.
Software Dependencies Yes All methods are implemented in Pytorch 1.10.
Experiment Setup Yes We train the model using stochastic gradient descent with momentum 0.9, and weight decay 10 4. ... the initial learning rate is 0.5 with cosine scheduling, the batch size is 512, and the training time is 500 epochs. We choose the default weight λc = 2... The temperature τ is 0.1.