🤖 machine learning

GEODE: Angle-Adaptive OOD Detection with Universal Scorer Compatibility

GEODE is a geometry-preserving OOD detection method that uses an angle-adaptive norm loss to synthetically replicate the boundary calibration benefits of Outlier Exposure, achieving robust performance across all seven standard scorers without requiring curated auxiliary data.

Original authors: Bruno Abrahao

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bruno Abrahao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart security guard (a computer program) trained to recognize specific people, like your friends and family (these are the "In-Distribution" or ID data). The problem arises when strangers walk up to the gate. Some are clearly strangers from a different country (Far-OOD), but others look suspiciously like your friends—they might be wearing similar clothes or have similar hair (Near-OOD).

The goal of OOD (Out-of-Distribution) detection is to make sure the guard can spot anyone who isn't on the approved list, no matter how they look.

The Problem: The "One-Size-Fits-All" Failure

For a long time, researchers tried to train guards to spot strangers using two main tricks:

The "Logit" Guard: Looks at the guard's internal confidence score. "How sure am I this is my friend?"
The "Feature" Guard: Looks at the physical distance in a map of features. "How far is this person from the group of my friends?"

The paper points out a frustrating trade-off: If you train the guard to be great at checking confidence scores, they often become terrible at checking physical distances, and vice versa. It's like training a guard to be an expert at reading ID cards but making them blind to facial features.

The Old Solution: "Outlier Exposure" (OE)

The previous best method was called Outlier Exposure (OE). Imagine you hire a trainer to show the guard photos of random strangers (like people from a different city) so they learn what "not my friend" looks like.

The Catch: You need a huge, carefully curated album of these "stranger" photos. If the strangers you show them don't match the strangers who actually show up at the gate, the training fails.
The Secret: The authors discovered why OE works. It's not because the guard memorized every stranger. It's because the training accidentally pushed the "stranger" photos to sit right on the borderline between "friend" and "stranger." It's like painting a "Do Not Cross" line right at the edge of the friend zone.

The New Solution: GEODE

The authors created GEODE (GEOmetry-preserving DEtection). Instead of needing a real album of strangers, GEODE creates fake strangers on the fly and uses a clever, adaptive rule to train the guard.

Here is the analogy for how GEODE works:

1. The "Shape-Shifting" Stranger

Instead of just showing the guard random photos, GEODE takes the features of your friends and shuffles them around (like mixing up the pixels of a face). This creates a "fake stranger" that looks weird but isn't a real person.

2. The "Smart Ruler" (Angle-Adaptive Norm)

This is the core innovation. Imagine the guard has a ruler to measure how "far out" a person is from the center of the friend group.

Old Method: "Push every stranger all the way to the edge of the universe!" (This makes them easy to spot by confidence, but destroys the map for distance-based checks).
GEODE Method: "Use a Smart Ruler."
- If a fake stranger is standing close to a specific friend's group (high similarity), the ruler says: "Stay a little bit away, but don't go too far." This preserves the local shape of the map so distance-based detectors (like KNN) still work.
- If a fake stranger is standing far from any friend group (low similarity), the ruler says: "Run all the way to the center!" This makes them very easy to spot by confidence-based detectors.

This "Smart Ruler" adjusts the training target for every single fake stranger based on where they are standing. It's like a dance instructor who tells dancers near the center to stay close to the group, but tells dancers in the back to move to the front, ensuring everyone stays in a formation that is easy to monitor from all angles.

Why This Matters

Universal Compatibility: Most methods break one type of detector to fix another. GEODE is the first to say, "I can make the guard good at both reading confidence scores and checking physical distances." It works with all seven standard ways of checking for strangers without crashing any of them.
No Extra Data Needed: You don't need to go find a dataset of "strangers" to train on. GEODE creates the training signal itself by shuffling the data it already has.
The "Null Space" Trap: The paper warns against other methods (like PFS) that try to push strangers into a "blind spot" (null space) where the guard can't see them. While this makes the guard great at confidence scores, it makes them completely blind to distance-based checks (like a guard who can read ID cards but is blindfolded). GEODE avoids this trap by keeping the "strangers" visible in the right places.

The Bottom Line

GEODE is a new training technique that teaches a computer model to spot strangers by creating smart, adaptive fake strangers on the fly. Instead of forcing all strangers into a single "bad" zone, it places them in a geometric pattern that keeps the "friend" and "stranger" boundaries clear for every type of detection method.

It achieves this without needing extra data, and it avoids the common pitfall of fixing one problem while breaking another. In short, it makes the security guard robust, versatile, and ready for any type of stranger that might show up.

Technical Summary: GEODE - Angle-Adaptive OOD Detection with Universal Scorer Compatibility

1. Problem Statement

Out-of-Distribution (OOD) detection aims to flag inputs that a deployed classifier was not trained on. Current approaches face a fundamental tension between post-hoc scoring (applied to frozen models) and training-based methods (which reshape features or logits).

The Trade-off: Training methods that improve logit-based scorers (e.g., MSP, Energy) often degrade feature-space scorers (e.g., KNN), and vice versa. No existing method achieves strong detection across all standard scorers and across both near-OOD and far-OOD regimes simultaneously.
Limitations of Outlier Exposure (OE): While OE is a dominant training-based approach using real auxiliary OOD data, it suffers from two issues:
1. Scorer Dependency: It exhibits trade-offs where it is strong on some scorers but weak on others.
2. Data Dependency: It requires curated auxiliary data that must match the true test-time OOD distribution, a requirement that is labor-intensive and fragile.
The Null-Space Failure: Recent methods like PFS attempt to push OOD features into the classifier's null space to boost logit scores. However, this catastrophically destroys feature-space geometry, causing KNN-based detection to fail (performing worse than random).

The core question addressed is: What geometric property makes OE effective, and can this be replicated synthetically without real OOD data to achieve universal scorer compatibility?

2. Methodology: GEODE

The authors propose GEODE (GEOmetry-preserving DEtection), a training-based method that synthesizes OOD signals to calibrate the decision boundary while preserving feature geometry.

2.1 Geometric Analysis of OE

Through geometric analysis in the penultimate layer of a trained backbone (ResNet-18), the authors found that:

OE is Boundary Calibration, not Coverage: OE features are geometrically indistinguishable from real near-OOD data. Specifically, they match in norm ratio, maximum cosine similarity to class means, and subspace projection.
The Quartile Effect: The "boundary-adjacent quartile" of OE data (samples with high cosine similarity to ID class means, $>0.85$ ) drives nearly all of OE's performance gain. Examples far from the boundary contribute little.
Conclusion: The effective mechanism is calibrating the boundary region, not covering the entire OOD space.

2.2 The GEODE Mechanism

GEODE replicates this calibration synthetically using feature shuffling (permuting dimensions across a batch) and an angle-adaptive norm loss.

Synthetic OOD Generation: Instead of using auxiliary data, GEODE creates synthetic OOD features ( $h_{ood}$ ) by independently permuting feature dimensions across a batch of ID features. This breaks inter-dimensional correlations while preserving marginal statistics.
Angle-Adaptive Target: The core innovation is a target norm $t$ $t$ that scales per-sample based on the cosine similarity ( $s$ $s$ ) of the synthetic feature to the nearest class mean ( $\mu_c$ $μ_{c}$ ):
$s = \max_c \cos(h_{ood}, \mu_c)$
$t = s^2 \cdot \alpha \cdot r_{id}$
Where $r_{id}$ $r_{i d}$ is the average ID feature norm and $\alpha$ $α$ is a relaxation parameter.
- Near Class Means ( $s \approx 1$ ): The target norm is large ( $t \approx \alpha \cdot r_{id}$ ). This preserves the local feature geometry required by distance-based scorers like KNN.
- Far from Classes ( $s \approx 0$ ): The target norm is small ( $t \approx 0$ ). This pushes features toward the origin, sharpening separation for logit-based scorers like MSP and Energy.
Contrastive ID Margin: To widen the margin from both sides, a contrastive term encourages ID features to have norms slightly larger than $r_{id}$ (penalizing ID features only when $\|h\| < r_{id}$ ).
Loss Function: The total loss combines Cross-Entropy, the adaptive OOD norm penalty, and the contrastive ID margin.

2.3 Theoretical Justification

The design is grounded in Neural Collapse (NC) theory, where class means converge to a Simplex Equiangular Tight Frame (ETF).

Theorem 1: Shuffled features concentrate at the global mean, justifying the need for adaptive targeting based on individual sample angles.
Theorem 2 & 3: The adaptive target preserves local pairwise distances near class boundaries (crucial for KNN) better than a uniform norm penalty.
Theorem 4: Pushing OOD features into the null space (as PFS does) eliminates class-conditional distance variation, causing KNN to fail. GEODE avoids this by keeping OOD features within the classifier subspace.

3. Key Contributions

Geometric Framing: The paper provides the first geometric characterization of why OE works (boundary calibration) and diagnoses the catastrophic failure of null-space methods (PFS) on KNN.
Method Design: GEODE introduces an angle-adaptive norm loss that synthesizes OOD signals without auxiliary data, preserving the geometry required for both logit-based and feature-space scorers.
Empirical Universality: GEODE achieves competitive AUROC across all seven standard scorers (MSP, Energy, ODIN, ReAct, Scale, ASH, KNN) on CIFAR-10 and CIFAR-100 without catastrophic failure on any single scorer.
Composability: GEODE is orthogonal to OE. Combining them yields state-of-the-art results (e.g., 95.0 MSP / 94.8 KNN on CIFAR-10), beating OE alone on every scorer.

4. Experimental Results

Evaluated on OpenOOD v1.5 benchmarks (CIFAR-10, CIFAR-100, ImageNet-200) using ResNet-18 and WRN-28-10:

CIFAR-10 (Near-OOD): GEODE achieves 92.30 AUROC (EBO) and ranges from 89.0 to 92.3 across all seven scorers. It avoids the "catastrophic failure" seen in other methods (e.g., PFS drops KNN to 14.38; GEODE maintains 89.04, a deliberate 1.6-point trade-off for massive logit gains).
CIFAR-100: GEODE outperforms vanilla CE at matched epoch counts and beats OE on all three scorers (MSP, Energy, KNN) when combined.
ImageNet-200: GEODE shows no catastrophic failures across scorers (range 77.93–83.48), whereas LogitNorm fails catastrophically on Energy (29.67) and ReAct (26.53).
Training Dynamics: GEODE outperforms vanilla CE as training progresses (up to 400 epochs), benefiting from Neural Collapse convergence. In contrast, LogitNorm degrades with longer training because it interferes with NC convergence.

5. Significance and Claims

The paper claims that GEODE resolves the long-standing trade-off between logit-based and feature-space OOD detection. By recognizing that effective OOD detection relies on boundary calibration rather than full OOD coverage, GEODE synthesizes a training signal that respects the geometric locus of the decision boundary.

Universal Compatibility: Unlike previous methods that optimize for a specific scorer family at the expense of others, GEODE works across all standard scorers without catastrophic failure.
Data Efficiency: It achieves these results without requiring curated auxiliary OOD datasets, relying instead on synthetic feature shuffling.
Theoretical Alignment: The method is designed to be compatible with the natural convergence of Neural Collapse, unlike logit-modifying methods that disrupt this convergence.

The authors modestly note limitations: the contrastive margin component does not generalize well to CIFAR-100 (due to fragile 100-class ETF geometry), and GEODE does not improve upon vanilla ViT-B/16 on small datasets (likely due to LayerNorm erasing norm variations and the model not reaching the NC regime). Future work is suggested for Transformer-friendly variants and larger-scale benchmarks.