Portion-Level OOD Detection Using Task-Agnostic Etalons
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing out-of-distribution (OOD) detection methods focus on improving robustness on in-distribution data, assuming a specific form of anomalies, which is against the definition of anomalies, and lack applicability to a wide spectrum of problems without task-specific biases.
Innovation Solution
A system and method for detecting whether a portion of an input belongs to at least one class by encoding fragments, obtaining in-distribution and out-of-distribution probability distributions, and comparing likelihoods using a classifier model, with optional etalon representation and iterative etalon determination to enhance discrimination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing OOD detection methods focus on improving robustness on in-distribution data with assumed anomaly forms, then the detection accuracy for known classes is improved, but the applicability to wide spectrum of problems without task-specific biases deteriorates
Solution Approach 1:
The patent employs a universal etalon representation approach that does not depend on task-specific assumptions about anomaly forms. The etalons are learned in a task-agnostic manner from in-distribution data only, enabling the system to detect OOD portions across diverse applications (image classification, segmentation, detection) without requiring task-specific training or auxiliary OOD data, thus achieving both reliability and versatility
2Measurement precision
If input-level classification is used for OOD detection, then the classification accuracy is improved, but the ability to detect OOD elements in portions of inputs deteriorates
Solution Approach 1:
The patent segments the input into multiple portions and processes each portion independently through the classifier model. The etalon-based OOD detection is applied at the portion level rather than treating the entire input as a single unit, enabling precise localization of OOD elements within specific regions while maintaining the benefits of accurate classification
Data Source
Figure 1
Figure 2(a)~3
AI summary
A system (30) for detecting whether a portion of an input (10) belongs to at least one class, the system being configured to: - encode at least one fragment (p) of the input, including the portion; - obtain an in-distribution probability distribution of the class, i.e. a distribution of a probability to belong to the class as a function of a result of at least one classifier model (14) configured to classify the fragment; - obtain an out-of-distribution probability distribution, i.e. a distribution of a probability not to belong to any one of the classes as a function of the result of the classifier model (14); - determine whether the portion belongs to the classes based on a comparison between a likelihood that the fragment (p) belongs to one of the classes, determined based on the in-distribution probability distribution, and a likelihood that the fragment (p) does not belong to any class, determined based on the out-of-distribution probability distribution.