Self-Supervised Learning for Generalizable Out-of-Distribution Detection

Authors: Sina Mohseni, Mandar Pitale, JBS Yadawa, Zhangyang Wang5216-5223

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

Reproducibility Variable Result LLM Response
Research Type Experimental We perform multiple image classification experiments and observe our technique to perform favorably against stateof-the-art OOD detection methods.
Researcher Affiliation Collaboration Sina Mohseni,1,2 Mandar Pitale,2 JBS Yadawa,2 Zhangyang Wang1 1Texas A&M University, College Station, TX 2NVIDIA, Santa Clara, CA
Pseudocode Yes Algorithm 1 Two-Step Training for Inand Out-of-Distribution Training Sets
Open Source Code No The paper does not provide any statements about open-sourcing the code or a link to a code repository.
Open Datasets Yes For example, in the MNIST (Le Cun et al. 1998) experiment, while the normal Din train is handwritten digits, we used English letters from E-MNIST (Cohen et al. 2017) as the source of Dout train set. We then evaluate the OOD detection performance with unseen Dout test including Kuzushiji-MNIST (Clanuwat et al. 2018), not-MNIST (Bulatov 2011), and Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017) datasets to measure how well can the model generalize on unseen distributions. Other experiments include training multi-class classifiers on CIFAR-10, CIFAR-100 (Krizhevsky, Hinton, and others 2009), and SVHN (Netzer et al. 2011) datasets. In all experiments (except the MNIST) we used 80 Million Tiny Images dataset (Torralba, Fergus, and Freeman 2008) as the source of unlabeled Dout train.
Dataset Splits No The paper mentions using training and test sets and discusses 'Din train' and 'Dout train', but does not explicitly provide details about standard train/validation/test splits (e.g., percentages or counts for each split) or refer to pre-defined splits for validation.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory, or specific cloud instance types used for the experiments.
Software Dependencies No The paper does not mention any specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We trained the model for 100 epochs in CIFAR-10 and CIFAR-100 experiments, 20 epochs for the SVHN training set, and 10 epochs for the MNIST experiment. We used batch size of 128, learning rate of 0.1 (decayed on a cosine learning rate schedule), and dropout rate of 0.3 for the CIFAR-10, CIFAR-100, and SVHN experiments. For the MNIST experiment, we used batch size of 64, learning rate of 0.01 (decayed on a cosine learning rate schedule), and dropout rate of 0.1.