iDECODe: In-Distribution Equivariance for Conformal Out-of-Distribution Detection

Authors: Ramneet Kaur, Susmit Jha, Anirban Roy, Sangdon Park, Edgar Dobriban, Oleg Sokolsky, Insup Lee7104-7114

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of i DECODe by experiments on image and audio datasets, obtaining state-of-the-art results. We also show that i DECODe can detect adversarial examples.
Researcher Affiliation Collaboration Ramneet Kaur1, Susmit Jha2, Anirban Roy2, Sangdon Park13, Edgar Dobriban4, Oleg Sokolsky1, Insup Lee1 1 Department of Computer and Information Science, University of Pennsylvania, Philadelphia, USA 2 Computer Science Laboratory, SRI International, Menlo Park, USA 3 School of Computer Science, Georgia Institute of Technology 4 Statistics & Computer Science, University of Pennsylvania
Pseudocode Yes Algorithm 1: i D Equivariance for Conformal OOD Detection
Open Source Code Yes Code, pre-trained models, and data are available at https://github.com/ramneetk/i DECODe.
Open Datasets Yes We demonstrate the efficacy of i DECODe by experiments on image and audio datasets... ablations studies on CIFAR-10 as the i D dataset and SVHN, LSUN, Image Net, CIFAR100, and Places365 datasets as OOD... FSDNoisy18k (FSD) (Fonseca et al. 2019) is an audio dataset
Dataset Splits Yes Here, the training set is split into a proper training set Xtr = {xi : i = 1, . . . , m} and a calibration set Xcal = {xj : j = m+1, . . . , l}... Calibration set of size 1000 images is randomly sampled with replacement from a held-out (not used in training) set of 5000 images... train the AVT model with WRN architecture on the proper training set (90% of the total training set) of one class of CIFAR-10... train the AVT model with Res Net-18 architecture on the proper training set (90% of the total training set) of CIFAR-100.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions using frameworks like ResNet and WRN architectures, and implies PyTorch due to the code base URL, but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes The details about the AVT model M with Res Net architecture trained to be G-equivariant on the proper training set (90% of the total training data) of CIFAR-10 (with the same set of hyperparameters from Qi et al. (2019) s AVT model on CIFAR-10) are given in Appendix C.1.1.