Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection

1. 1Revisit the assumption that penultimate-layer activations are the most informative for OOD detection
2. 2Explore the use of intermediate layers for encoding rich and discriminative in-distribution representations
3. 3Apply Prototype Fusion to combine representations from multiple layers for improved OOD detection

Machine learning researchers and engineers working on safety-critical applications can benefit from this approach to improve the robustness of their models, and data scientists can apply this method to detect anomalies in their datasets

💡 Intermediate layers can encode equally rich and discriminative representations as penultimate layers for OOD detection

🚀 Improve OOD detection with Prototype Fusion, a training-free multi-layer approach!