Single Layer Predictive Normalized Maximum Likelihood for Out-of-Distribution Detection

Authors: Koby Bibas, Meir Feder, Tal Hassner

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We extensively evaluate our approach on 74 OOD detection benchmarks using Dense Net-100, Res Net-34, and Wide Res Net40 models trained with CIFAR-100, CIFAR-10, SVHN, and Image Net-30 showing a significant improvement of up to 15.6% over recent leading methods.
Researcher Affiliation Collaboration Koby Bibas School of Electrical Engineering Tel Aviv University kobybibas@gmail.com Meir Feder School of Electrical Engineering Tel Aviv University meir@eng.tau.ac.il Tal Hassner Facebook AI talhassner@gmail.com
Pseudocode No The paper describes methods using mathematical equations and prose, but it does not include a dedicated pseudocode or algorithm block.
Open Source Code Yes Code is available in https://github.com/kobybibas/pnml_ood_detection
Open Datasets Yes For datasets that represent known classes, we use CIFAR-100, CIFAR-10 (Krizhevsky et al., 2014) and SVHN (Netzer et al., 2011). [...] In addition, to evaluate higher resolution images, we use Image Net-30 set (Hendrycks et al., 2019b).
Dataset Splits No The paper mentions using standard datasets like CIFAR-100, CIFAR-10, SVHN, Image Net-30 for IND sets and Tiny Image Net, LSUN, i SUN, Uniform noise, and Gaussian noise for OOD sets, but it does not specify the exact training, validation, and test splits used for these datasets within the paper.
Hardware Specification Yes We ran all experiments on NVIDIA K80 GPU.
Software Dependencies No The paper does not provide specific version numbers for ancillary software components, only mentioning the use of various models and datasets without detailing the software environment (e.g., specific library versions).
Experiment Setup No The paper mentions using pretrained models (Res Net-34, Dense Net-BC-100, Wide Res Net-40) and the datasets they were trained with (CIFAR-100, CIFAR-10, SVHN), but it does not provide specific hyperparameters such as learning rates, batch sizes, number of epochs, or optimizer settings for its own experimental setup.