MM++ Offers Unsupervised, Scale-Invariant OOD Detection.

Detecting Out-of-Distribution (OOD) data is a critical challenge in machine learning, often requiring a trade-off between scale invariance and the ability to capture hierarchical features. MM++ (Multilayer Mahalanobis++) introduces a novel, fully unsupervised and strictly post-hoc framework designed to address this, providing scale-invariant OOD detection without needing auxiliary OOD data, classifier fine-tuning, or architectural modifications. The core of MM++ lies in its principled construction of a joint feature space. It intelligently identifies discriminative intermediate layers within a neural network by observing entropy density drops, which signal boundaries of significant semantic compression. These selected layers are then fused with the network's terminal representation, allowing the framework to capture latent cross-layer correlations while simultaneously mitigating noise from early layers. To ensure reliable distance estimation within this unified space, MM++ employs a Ledoit-Wolf regularized tied covariance matrix. This stabilization technique contributes to the framework's robust performance across various architectures for both near-OOD and far-OOD detection scenarios. Its ability to operate without any prior knowledge of OOD data makes it highly practical for real-world applications.