MM++: Unsupervised Scale-Invariant Multilayer OOD Detection via Top-K Gated Feature Fusion
Researchers have developed MM++ (Multilayer Mahalanobis++), a novel unsupervised framework designed for Out-of-Distribution (OOD) detection. This method constructs a joint feature space by identifying and fusing discriminative intermediate layers with the terminal representation, capturing cross-layer correlations while filtering out noise. MM++ utilizes a regularized tied covariance matrix for stable distance estimation and requires no additional OOD data, classifier fine-tuning, or architectural changes, demonstrating robust performance across various architectures for both near and far OOD detection. AI