Tuesday 08 April 2025
A new approach to detecting out-of-distribution data, which is crucial for ensuring the reliability and safety of machine learning models in real-world applications, has been proposed by a team of researchers. The method, dubbed OT-DETECTOR, leverages optimal transport theory to quantify both semantic and distributional discrepancies between test samples and in-distribution labels.
The problem of out-of-distribution (OOD) detection arises when machine learning models are faced with data from unknown classes or distributions. This can lead to catastrophic outcomes in high-stakes applications such as autonomous driving and medical diagnostics, where accurate detection of OOD data is essential. Existing methods for OOD detection have primarily focused on semantic matching, but this approach has limitations. OT-DETECTOR addresses these limitations by incorporating both semantic and distributional discrepancies into its evaluation framework.
The researchers behind OT-DETECTOR employ a novel technique called cross-modal transport mass to quantify the semantic discrepancy between test samples and in-distribution labels. This involves computing the optimal transport plan between the feature distributions of the test samples and in-distribution labels, which provides a measure of their similarity. The team also introduces a distributional discrepancy metric based on the optimal transport cost, which captures the difference between the distributions of the test samples and in-distribution labels.
To further enhance its performance, OT-DETECTOR incorporates a semantic-aware content refinement (SaCR) module. This module uses semantic cues from the in-distribution labels to amplify the distributional discrepancy between in-distribution and hard OOD samples. The SaCR module is trained on a subset of the training data and can be adapted to different datasets.
The researchers evaluated OT-DETECTOR on several benchmark datasets, including ImageNet, CUB-200, Stanford Cars, Food-101, Oxford Pets, iNaturalist, SUN, Places, and Textures. The results show that OT-DETECTOR achieves state-of-the-art performance across various OOD detection tasks, particularly in challenging hard-OOD scenarios.
One of the key advantages of OT-DETECTOR is its robustness to variations in prompt templates. Unlike other methods, which are sensitive to changes in the prompt template, OT-DETECTOR consistently delivers stable performance regardless of the template used. This makes it a more reliable and practical solution for real-world applications.
The team also evaluated the impact of different batch sizes on OT-DETECTOR’s performance.
Cite this article: “Unveiling the Power of Optimal Transport: A Novel Approach to Zero-Shot Out-of-Distribution Detection”, The Science Archive, 2025.
Machine Learning, Out-Of-Distribution, Detection, Optimal Transport Theory, Semantic Matching, Distributional Discrepancy, Feature Distributions, Optimal Transport Cost, Semantic-Aware Content Refinement, Benchmark Datasets.
