Learning by Erasing: Conditional Entropy based Transferable Out-Of-Distribution Detection

TL;DR

This paper proposes a Deep Generative Models based transferable OOD detection that does not require retraining on the new ID dataset, and presents an innovative image-erasing strategy, which is designed to create distinct conditional entropy distributions for each individual ID dataset.

Abstract

Out-of-distribution (OOD) detection is essential to handle the distribution shifts between training and test scenarios. For a new in-distribution (ID) dataset, existing methods require retraining to capture the dataset-specific feature representation or data distribution. In this paper, we propose a deep generative models (DGM) based transferable OOD detection method, which is unnecessary to retrain on a new ID dataset. We design an image erasing strategy to equip exclusive conditional entropy distribution for each ID dataset, which determines the discrepancy of DGM's posteriori ucertainty distribution on different ID datasets. Owing to the powerful representation capacity of convolutional neural networks, the proposed model trained on complex dataset can capture the above discrepancy between ID datasets without retraining and thus achieve transferable OOD detection. We validate the proposed method on five datasets and verity that ours achieves comparable performance to the state-of-the-art group based OOD detection methods that need to be retrained to deploy on new ID datasets. Our code is available at https://github.com/oOHCIOo/CETOOD.

Figures (10)

• Figure 1: Original images (top row) and their reconstruction results (bottom row) by a pre-trained DGM. The model pre-trained on CI (CIFAR10) / Fa (FashionMNIST) can reconstruct SV (SVHN) / MN (MNIST) samples well, but not vice versa. Reconstruction performances on the test set of the target datasets are evaluated with (SSIM$\uparrow$ / PSNR$\uparrow$).
• Figure 2: The (a) and (b) are the DGM's negative log-likelihood distribution on different datasets. The model of (a) is trained by reconstructing the input, while the model of (b) is trained by generating the erased patch based on its surrounding information. The real negative conditional entropy distribution between the erased patch and its surrounding is given in (c). The DGM is an auto-encoder proposed in this paper and is trained with ImageNet. The image size is $32\times32$, and the erased image patch in (b) and (c) is at the center of the image with the size of $16\times16$. Kernel Density Estimation (KDE) is used to estimate the probability distribution.
• Figure 3: The pipeline of the proposed CETOOD.
• Figure 4: The center (a), corner (b) and side (c) of the image is erased, which is indicated with white color.
• Figure 5: The time ($t$, hours) and space ($m$, FLOPs) complexity comparison between our method and the baseline approaches.