Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Multi-dimensional Semantic Surprise Framework Based on Low-Entropy Semantic Manifolds for Fine-Grained Out-of-Distribution Detection

Ningkang Peng, Yuzhe Mao, Yuhao Zhang, Linjin Qian, Qianfeng Yu, Yanhui Gu, Yi Chen, Li Kong

TL;DR

This work reframes OOD detection as a semantic-surprise problem by introducing a ternary ID/Near-OOD/Far-OOD framework and Low-Entropy Semantic Manifolds learned via a Hierarchical Prototypical Network. The Semantic Surprise Vector decomposes sample risk into Conformity, Novelty, and Ambiguity, evaluated through the Normalized Semantic Risk, achieving state-of-the-art performance on the challenging ternary task and strong results on binary OOD benchmarks. The approach emphasizes structured geometry and interpretable risk diagnostics, offering a principled path toward safer AI in open-world settings. Overall, the method demonstrates robust generalization and provides a cognitive-aligned, information-theoretic alternative to traditional probabilistic OOD detection.

Abstract

Out-of-Distribution (OOD) detection is a cornerstone for the safe deployment of AI systems in the open world. However, existing methods treat OOD detection as a binary classification problem, a cognitive flattening that fails to distinguish between semantically close (Near-OOD) and distant (Far-OOD) unknown risks. This limitation poses a significant safety bottleneck in applications requiring fine-grained risk stratification. To address this, we propose a paradigm shift from a conventional probabilistic view to a principled information-theoretic framework. We formalize the core task as quantifying the Semantic Surprise of a new sample and introduce a novel ternary classification challenge: In-Distribution (ID) vs. Near-OOD vs. Far-OOD. The theoretical foundation of our work is the concept of Low-Entropy Semantic Manifolds, which are explicitly structured to reflect the data's intrinsic semantic hierarchy. To construct these manifolds, we design a Hierarchical Prototypical Network. We then introduce the Semantic Surprise Vector (SSV), a universal probe that decomposes a sample's total surprise into three complementary and interpretable dimensions: conformity, novelty, and ambiguity. To evaluate performance on this new task, we propose the Normalized Semantic Risk (nSR), a cost-sensitive metric. Experiments demonstrate that our framework not only establishes a new state-of-the-art (sota) on the challenging ternary task, but its robust representations also achieve top results on conventional binary benchmarks, reducing the False Positive Rate by over 60% on datasets like LSUN.

A Multi-dimensional Semantic Surprise Framework Based on Low-Entropy Semantic Manifolds for Fine-Grained Out-of-Distribution Detection

TL;DR

This work reframes OOD detection as a semantic-surprise problem by introducing a ternary ID/Near-OOD/Far-OOD framework and Low-Entropy Semantic Manifolds learned via a Hierarchical Prototypical Network. The Semantic Surprise Vector decomposes sample risk into Conformity, Novelty, and Ambiguity, evaluated through the Normalized Semantic Risk, achieving state-of-the-art performance on the challenging ternary task and strong results on binary OOD benchmarks. The approach emphasizes structured geometry and interpretable risk diagnostics, offering a principled path toward safer AI in open-world settings. Overall, the method demonstrates robust generalization and provides a cognitive-aligned, information-theoretic alternative to traditional probabilistic OOD detection.

Abstract

Out-of-Distribution (OOD) detection is a cornerstone for the safe deployment of AI systems in the open world. However, existing methods treat OOD detection as a binary classification problem, a cognitive flattening that fails to distinguish between semantically close (Near-OOD) and distant (Far-OOD) unknown risks. This limitation poses a significant safety bottleneck in applications requiring fine-grained risk stratification. To address this, we propose a paradigm shift from a conventional probabilistic view to a principled information-theoretic framework. We formalize the core task as quantifying the Semantic Surprise of a new sample and introduce a novel ternary classification challenge: In-Distribution (ID) vs. Near-OOD vs. Far-OOD. The theoretical foundation of our work is the concept of Low-Entropy Semantic Manifolds, which are explicitly structured to reflect the data's intrinsic semantic hierarchy. To construct these manifolds, we design a Hierarchical Prototypical Network. We then introduce the Semantic Surprise Vector (SSV), a universal probe that decomposes a sample's total surprise into three complementary and interpretable dimensions: conformity, novelty, and ambiguity. To evaluate performance on this new task, we propose the Normalized Semantic Risk (nSR), a cost-sensitive metric. Experiments demonstrate that our framework not only establishes a new state-of-the-art (sota) on the challenging ternary task, but its robust representations also achieve top results on conventional binary benchmarks, reducing the False Positive Rate by over 60% on datasets like LSUN.
Paper Structure (29 sections, 15 equations, 3 figures, 7 tables)

This paper contains 29 sections, 15 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Theoretical Foundation: The Low-Entropy Semantic Manifold
  5. Principle 1: Subclass Compactness.
  6. Principle 2: Superclass Cohesion.
  7. Principle 3: Superclass Separation.
  8. Manifold Quality Metrics.
  9. Manifold Shaping: The Hierarchical Prototypical Network
  10. Sample-Level Compactness.
  11. Prototype-Level Structure.
  12. Prototype Updating.
  13. Manifold Probing: The Semantic Surprise Vector
  14. 1. Conformity Surprise ($S_{\text{conf}}$).
  15. 2. Novelty Surprise ($S_{\text{novel}}$).
  16. ...and 14 more sections

Figures (3)

  • Figure 1: Visualization of feature spaces. (a) The PALM suffers from cognitive flattening, evidenced by the severe overlap between ID (blue) and Near-OOD (orange) clusters (highlighted by black ellipses). (b) Our method creates a significantly better-separated manifold, crucial for distinguishing fine-grained OOD risks.
  • Figure 2: Overview of our proposed framework. We learn a hierarchically-structured manifold on a hypersphere and probe it using the multi-dimensional SSV to inform a final risk classifier.
  • Figure 3: Comparison of confusion matrices for our method (left) and PALM (right). Our method shows a cleaner block-diagonal structure and significantly reduces high-risk misclassifications (e.g., Near-OOD predicted as ID).