Exploiting Inter-Sample Information for Long-tailed Out-of-Distribution Detection

TL;DR

This work tackles OOD detection under long-tailed ID distributions by building a graph that encodes inter-sample relationships using initial embeddings from a $k$-NN graph. It enhances these representations through Gaussianization of activations from a pre-trained backbone and refinement via a graph convolutional network, producing a discriminative feature space for LT-OOD detection. Evaluations on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT show state-of-the-art improvements in OOD detection metrics (AUROC, AUPR, FPR95) and notably better tail-class ID accuracy, demonstrating robustness to distribution shifts. The approach offers practical impact for safe deployment in real-world LT settings and illustrates the value of combining pre-training, normalization alignment, and relational graph learning for OOD tasks.

Abstract

Detecting out-of-distribution (OOD) data is essential for safe deployment of deep neural networks (DNNs). This problem becomes particularly challenging in the presence of long-tailed in-distribution (ID) datasets, often leading to high false positive rates (FPR) and low tail-class ID classification accuracy. In this paper, we demonstrate that exploiting inter-sample relationships using a graph-based representation can significantly improve OOD detection in long-tailed recognition of vision datasets. To this end, we use the feature space of a pre-trained model to initialize our graph structure. We account for the differences between the activation layer distribution of the pre-training vs. training data, and actively introduce Gaussianization to alleviate any deviations from a standard normal distribution in the activation layers of the pre-trained model. We then refine this initial graph representation using graph convolutional networks (GCNs) to arrive at a feature space suitable for long-tailed OOD detection. This leads us to address the inferior performance observed in ID tail-classes within existing OOD detection methods. Experiments over three benchmarks CIFAR10-LT, CIFAR100-LT, and ImageNet-LT demonstrate that our method outperforms the state-of-the-art approaches by a large margin in terms of FPR and tail-class ID classification accuracy.

Figures (5)

• Figure 1: Feature space representations of CIFAR10-LT as ID test set and CIFAR-100 as OOD test set using t-SNE. (a) Feature distribution of ResNet18 trained from scratch using OE hendrycks_2018OE, (b) ResNet18 pre-trained on Downsampled ImageNet, (c) same model in (b) after applying Gaussianization, and (d) model in (c) after message passing using GCN. Tail ID classes (e.g., Horse, Ship, Truck) heavily overlap with OOD samples (represented as red dots) in (a). Separability enhanced from left to right, wherein (d) there are clear gaps between the decision boundaries.
• Figure 2: Overview of proposed methodology. Given $\mathcal{D}_{\mathrm{in}}$ with $K$ classes and and auxiliary OOD training set $\mathcal{D}_{\mathrm{out}}^{\mathrm{OE}}$, we extract the initial feature embeddings from a pre-trained backbone model $f$ after applying Gaussianization using the same training data. We then create a $k$-NN graph to refine these initial features using message passing in the form of GCN $g$. Then, we incorporate a fully-connected classifier to arrive at the objective loss for training $g$.
• Figure 3: Distribution of number of ID misclassifications over each class for CIFAR100-LT, with (a) one of the baselines "Pre-train" and (b) our method. (c) Head-class ID, tail-class ID and OOD neighbor node distribution over each ID class of CIFAR100-LT. Average neighbor values per each class are plotted due to the long-tailed nature. ID class labels higher than 50 belong to tail-classes.
• Figure 4: Variation of CIFAR10-LT ID classification accuracy over selection of $k$ in the $k$-NN graph.
• Figure 5: Feature space representations obtained from the (a) MLP and (b) GCN models for CIFAR10-LT as ID test set and CIFAR-100 as OOD test set using t-SNE. Equal number of test samples from each ID class are visualized. The GCN demonstrates distinct gaps between the decision boundaries of ID and OOD samples compared to the MLP.