Hypercone Assisted Contour Generation for Out-of-Distribution Detection

2025-01-17

AI Summary

This paper addresses the out-of-distribution (OOD) detection problem for high-dimensional data (e.g., CIFAR-100) under *no distributional assumptions*. To this end, it proposes a geometric modeling framework based on *adaptive hyperconic clusters*: for the first time, hyperconic structures are introduced to quantify the *local angular boundaries* of in-distribution (ID) sample support in feature space—without requiring OOD supervision or prior distributional knowledge—and dynamically adapt to the underlying data manifold geometry. The method integrates *k-nearest-neighbor angular distance maximization*, *hyperconic space partitioning*, and *silhouette-based boundary estimation* to achieve a compact geometric characterization of ID regions. On the CIFAR-100 benchmark, it achieves state-of-the-art performance for both Near-OOD and Far-OOD detection—attaining the lowest FPR@95 and highest AUROC—while being trained *exclusively on ID data*, with *zero OOD samples used during training*.

Technology Category

• Data Mining & Knowledge Management: Anomaly/Outlier Detection
• Machine Learning: Learning with Manifolds
• Computer Vision: Other Foundations of Computer Vision

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

Recent advances in the field of out-of-distribution (OOD) detection have placed great emphasis on learning better representations suited to this task. While there are distance-based approaches, distributional awareness has seldom been exploited for better performance. We present HAC$_k$-OOD, a novel OOD detection method that makes no distributional assumption about the data, but automatically adapts to its distribution. Specifically, HAC$_k$-OOD constructs a set of hypercones by maximizing the angular distance to neighbors in a given data-point's vicinity to approximate the contour within which in-distribution (ID) data-points lie. Experimental results show state-of-the-art FPR@95 and AUROC performance on Near-OOD detection and on Far-OOD detection on the challenging CIFAR-100 benchmark without explicitly training for OOD performance.

Problem

• Out-of-Distribution Detection
• Anomaly Detection
• High-Dimensional Data

Innovation

• HAC_k-OOD
• Anomaly Detection
• Hypercone Construction