Igeood: An Information Geometry Approach to Out-of-Distribution Detection

Authors: Eduardo Dadalto Camara Gomes, Florence Alberge, Pierre Duhamel, Pablo Piantanida

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we show that IGEOOD outperforms competing state-of-the-art methods on a variety of network architectures and datasets. (Abstract) / 4 EXPERIMENTAL RESULTS We show the effectiveness of IGEOOD comparing to state-of-the-art methods. Details about the experimental setup and additional results are given in appendices (see Sections C, D, and E).
Researcher Affiliation Academia Eduardo D. C. Gomes, Florence Alberge & Pierre Duhamel Laboratoire des signaux et syst emes (L2S) Universit e Paris-Saclay CNRS Centrale Sup elec 91190, Gif-sur-Yvette, France. / Pablo Piantanida International Laboratory on Learning Systems (ILLS) Mc Gill ETS MILA CNRS Universit e Paris-Saclay Centrale Sup elec H3C 1K3 Quebec, Canada
Pseudocode Yes B IGEOOD ALGORITHMS AND COMPUTATION DETAILS / Algorithm 1: Evaluating IGEOOD score based on the logits. / Algorithm 2: Evaluating feature-wise IGEOOD score. / Algorithm 3: Evaluating feature-wise IGEOOD+ score.
Open Source Code Yes Our code is publicly available at https://github.com/edadaltocg/Igeood.
Open Datasets Yes We take as in-distribution data images from CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 and SVHN (Netzer et al., 2011) datasets. (Section 4) / C.3 DATASETS: Provides detailed descriptions and original citations for CIFAR-10, CIFAR-100, SVHN, Tiny-Image Net, LSUN, i SUN, Textures (DTD), Chars74K, Places365, Gaussian.
Dataset Splits Yes For the BLACK-BOX and GREY-BOX experimental settings, we tune hyperparameters for all of the OOD detectors only based on the DNN classifier architecture, the in-distribution dataset, and a validation dataset. The i SUN (Xu et al., 2015) dataset is chosen as a source of OOD validation data, independently from OOD test data. (Section 4.1)
Hardware Specification No Not found. The paper mentions 'one GPU' and 'one high-end GPU' but does not specify the exact model (e.g., NVIDIA A100, V100, RTX 3090) or other hardware components like CPU or memory for reproducibility.
Software Dependencies No Not found. While the paper mentions using SGD, cross-entropy loss, and implicitly deep learning frameworks, it does not specify any software names with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 11.x).
Experiment Setup Yes We train each model by minimizing the cross-entropy loss using SGD with Nesterov momentum equal to 0.9, weight decay equal to 0.0001, and a multi-step learning rate schedule starting at 0.1 for 300 epochs. (C.1) / We used a constant learning rate of 0.01 and a batch size of 128 for 100 epochs. (B.1) / For temperature T, we ran a Bayesian optimization for 500 epochs in the interval of temperature values between 1 and 1000... For the input pre-processing noise magnitude ε tuning, we ran a grid search optimization with 21 equally spaced values in the interval [0, 0.002]. (E.2) / Table 7: Best temperatures T for the BLACK-BOX setup, best temperature and noise magnitude (T, ε) for the GREY-BOX setup, and best ε for the Mahalanobis score and (T, ε) for IGEOOD and IGEOOD+ in the WHITE-BOX setup with adversarial tuning.