Abstract
Robust Out-of-Distribution (OOD) detection is critical for the safe deployment of machine learning models. While Outlier Exposure (OE) has emerged as a popular solution, it often suffers from a generalization gap where models overfit to specific training outliers. This paper demonstrates that the key to robust OOD detection lies in the diversity rather than the quantity of auxiliary outliers. We introduce diverseMix, a provably effective method that employs an adaptive Mixup strategy to synthetically broaden the variety of training outliers. Our approach significantly reduces false positive rates across major benchmarks, including CIFAR and ImageNet.
The Challenge of OOD Generalization
Machine learning models frequently assign high-confidence "In-Distribution" (ID) labels to data that significantly differs from their training set. This overconfidence poses severe risks in safety-critical applications like autonomous driving and medical diagnosis. While training on auxiliary outliers (Outlier Exposure) helps, models often fail to generalize to novel OOD types encountered at test time. This "generalization gap" stems from the distribution shift between available training outliers and unknown real-world OOD data.
Methodology: diverseMix
The authors provide a theoretical generalization bound showing that OOD detection error is directly linked to the diversity of the auxiliary outlier set. To bridge the generalization gap without the high cost of manual data collection, we propose diverseMix.
Diversity-induced Mixup
The core of diverseMix is a three-pronged strategy:
- Semantic Interpolation: The method creates "mixed" outliers by interpolating between different outlier samples, generating synthetic data points that lie outside original data manifolds.
- Score-Adaptive Weighting: Instead of random interpolation, diverseMix uses the model's current OOD score to prioritize generating outliers that the model finds "unfamiliar."
- Regularization: These synthetic outliers are used in a regularization term to push the model toward low confidence scores in OOD regions.
By using the model's own scores to guide Mixup, diverseMix effectively "mines" the boundaries of the model's knowledge, forcing it to learn about OOD regions it would otherwise ignore.
Experimental Evaluation
We evaluated diverseMix on standard benchmarks including CIFAR-10, CIFAR-100, and ImageNet-200, using a wide battery of OOD test sets such as SVHN, Textures, and Places365.
Key Results
The results demonstrate a striking reduction in the FPR95 (False Positive Rate at 95% True Positive Rate):
- CIFAR-10: diverseMix achieved an FPR95 of 1.92%, a relative error reduction of 24.4% over previous state-of-the-art methods.
- CIFAR-100: FPR95 dropped to 8.51%, representing a 43.8% relative improvement.
On the large-scale ImageNet benchmark, the method proved robust across both "near-OOD" and "far-OOD" scenarios, outperforming specialized methods like MixOE.
Conclusion
The findings of this paper suggest that Diversity > Quantity in OOD detection. By increasing the semantic range of outliers through provable synthetic augmentation, diverseMix lowers the theoretical upper bound on detection error. While future work may explore initial outlier quality and computational efficiency, diverseMix sets a new state-of-the-art for robust OOD detection in computer vision.
Summary updated