Deep Anomaly Detection with Outlier Exposure

Authors: Dan Hendrycks, Mantas Mazeika, Thomas Dietterich

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In extensive experiments on natural language processing and smalland large-scale vision tasks, we find that Outlier Exposure significantly improves detection performance.
Researcher Affiliation Academia Dan Hendrycks University of California, Berkeley hendrycks@berkeley.edu Mantas Mazeika University of Chicago mantas@ttic.edu Thomas Dietterich Oregon State University tgd@oregonstate.edu
Pseudocode No The paper presents a formal objective function for Outlier Exposure but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Our code is made publicly available at https://github.com/hendrycks/outlier-exposure.
Open Datasets Yes SVHN. The SVHN dataset (Netzer et al., 2011) contains 32 32 color images of house numbers. CIFAR. The two CIFAR (Krizhevsky & Hinton, 2009) datasets contain 32 32 natural color images. Tiny Image Net. The Tiny Image Net dataset (Johnson et al.) is a 200-class subset of the Image Net (Russakovsky et al., 2015) dataset.
Dataset Splits Yes Instead, the λ coefficients were determined early in experimentation with validation Dval out distributions described in Appendix A. In particular, we use λ = 0.5 for vision experiments and λ = 1.0 for NLP experiments. We train SVHN, CIFAR-10, CIFAR-100, and Tiny Image Net classifiers with 5000, 5000, 5000, and 10000 training examples held out, respectively.
Hardware Specification No The paper states 'We thank NVIDIA for donating GPUs used in this research,' but it does not specify the exact models or detailed specifications of the GPUs or any other hardware components used for the experiments.
Software Dependencies No The paper describes the models and datasets used but does not list specific software dependencies with version numbers, such as 'Python 3.x' or 'PyTorch 1.x'.
Experiment Setup Yes For CIFAR-10, CIFAR-100, and Tiny Image Net classification experiments, we use a 40-2 Wide Residual Network (Zagoruyko & Komodakis, 2016). The network trains for 100 epochs with a dropout rate of 0.3. The initial learning rate of 0.1 decays following a cosine learning rate schedule (Loshchilov & Hutter, 2017). During fine-tuning of the entire network, we again use a cosine learning rate schedule but with an initial learning rate of 0.001. We use standard flipping and data cropping augmentation, Nesterov momentum, and ℓ2 weight decay with a coefficient of 5 10 4.