CORE: Robust OOD Detection via Orthogonal Scoring

🔑 Enhanced Key Takeaways

•CORE identifies a 'directional signature' in the residual subspace (orthogonal to the predicted class weight) that remains stable for in-distribution data but deviates for OOD samples, even when those samples trigger high-confidence logits.
•The method leverages the mathematical independence of orthogonal subspaces to ensure that failure modes of the confidence signal and the membership signal do not overlap, providing a safety net when one signal is compromised.
•Empirical testing across ResNet, ViT, SwinV2, and DeiT architectures demonstrates that CORE maintains a high grand average AUROC (84.9%) without the architecture-specific performance degradation common in activation-shaping methods like ASH.
•CORE introduces a z-score normalized summation technique that allows the combination of disparate signal scales (logits vs. feature norms) without requiring manual hyperparameter tuning for each new dataset.

📊 Competitor Analysis▸ Show

Method	Type	Computational Overhead	Key Mechanism	Benchmark Rank (Avg AUROC)
CORE	Post-hoc / Orthogonal	Negligible ($O(d)$)	Orthogonal Subspace Decomposition	1st (84.9%)
Energy	Post-hoc / Logit	Minimal	Log-sum-exp of logits	Competitive (Baseline)
ReAct	Post-hoc / Activation	Minimal	Activation Truncation (Rectification)	High (Architecture Sensitive)
ASH	Post-hoc / Activation	Minimal	Activation Pruning/Shaping	High (Penultimate Layer only)
KNN	Post-hoc / Feature	High ($O(N)$)	Nearest Neighbor Distance	High (Memory Intensive)

🛠️ Technical Deep Dive

Detailed implementation and architectural details of the CORE framework:

Feature Decomposition: The penultimate feature vector $z$ is decomposed into $z = z_{\parallel} + z_{\perp}$, where $z_{\parallel}$ is the projection onto the weight vector $w_c$ of the predicted class $c$, and $z_{\perp}$ is the orthogonal residual.
Subspace Scoring: The confidence score is derived from the magnitude of $z_{\parallel}$ (which directly influences the logit), while the membership score is derived from the directional consistency of $z_{\perp}$.
Normalization Strategy: To combine these signals, CORE applies z-score normalization: $S_{combined} = \frac{S_{\parallel} - \mu_{\parallel}}{\sigma_{\parallel}} + \frac{S_{\perp} - \mu_{\perp}}{\sigma_{\perp}}$, where $\mu$ and $\sigma$ are estimated from a small held-out in-distribution validation set.
Complexity: The operation requires only a single vector projection and two norm calculations per inference, maintaining $O(d)$ time complexity where $d$ is the feature dimension.

🔮 Future ImplicationsAI analysis grounded in cited sources

Standardization of Subspace-Aware OOD

The success of CORE suggests that future OOD detection will move away from simple scalar logit analysis toward multi-dimensional vector-space decomposition.

Integration into Real-time Edge AI

With its $O(d)$ complexity and zero-retraining requirement, CORE is uniquely positioned for low-latency deployment in safety-critical edge devices like autonomous drones or medical sensors.

⏳ Timeline

2020-12

Energy-based OOD detection established as a logit-space baseline.

2021-11

ReAct introduces activation rectification to mitigate OOD overconfidence.

2023-02

ASH (Activation Shaping) improves OOD detection via penultimate layer pruning.

2024-08

Critical analysis of OOD benchmarks highlights architecture-sensitive failure modes.

2026-03

CORE (arXiv:2603.18290) published, introducing orthogonal residual scoring.

CORE: Robust OOD Detection via Orthogonal Scoring

⚡ 30-Second TL;DR

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (9)

👉Related Updates

Bias Mitigation Evaluated in LLM Judges