Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection

Out-of-distribution (OOD) detection is essential for reliably deploying machine learning models in the wild. Yet, most methods treat large pre-trained models as monolithic encoders and rely solely on their final-layer representations for detection. We challenge this wisdom. We reveal the $\mathit{\text{intermediate}}\mathit{\text{layers}}$ of pre-trained models, shaped by residual connections that subtly transform input projections, $\mathit{\text{can}}$ encode $\mathit{\text{surprisingly}}\mathit{\text{rich}}\mathit{\text{and}}\mathit{\text{diverse}}\mathit{\text{signals}}$ for detecting distributional shifts. Importantly, to exploit latent representation diversity across layers, we introduce an entropy-based criterion to $\mathit{\text{automatically}}$ identify layers offering the most complementary information in a training-free setting – $\mathit{\text{without}}\mathit{\text{access}}\mathit{\text{to}}\mathit{\text{OOD}}\mathit{\text{data}}$. We show that selectively incorporating these intermediate representations can increase the accuracy of OOD detection by up to $10\mathrm{\%}$ in far-OOD and over $7\mathrm{\%}$ in near-OOD benchmarks compared to state-of-the-art training-free methods across various model architectures and training objectives. Our findings reveal a new avenue for OOD detection research and uncover the impact of various training objectives and model architectures on confidence-based OOD detection methods.