TL;DR for operators

A vision classifier can retain the same application interface after a model upgrade while its unfamiliar-image detector becomes substantially less reliable. The problem may lie in how the detector processes internal features, not in the classifier’s output labels.

Anti-coherence Regularized Absolute Scaling (Anti-coh) addresses two issues: cancellation between positive and negative activations, and excessive amplification of features that oppose the predicted class. Across ten ImageNet backbones, it raises average out-of-distribution (OOD) detection AUROC from 71.28% to 85.81% relative to its predecessor, SCALE.

For teams maintaining image-classification services, the finding supports testing the detection layer whenever the backbone changes. Anti-coh is a candidate when original training data are unavailable but model features and classifier weights remain accessible. It still requires representative validation examples, and stronger alternatives exist for some architectures and datasets. Benchmark improvement alone does not establish that a production exception-routing policy is reliable.

The classifier changed, but the unfamiliar-image filter did not

Consider a company upgrading the pretrained backbone behind an image-classification service. The new model returns the same categories through the same interface. Its existing detector, intended to flag unfamiliar images before their classifications are accepted, remains attached.

That continuity can be misleading. The replacement model may encode images using quite different internal feature values. A detector that depends on those values can change behavior even when its own implementation is unchanged.

Donghwan Kim, Youngwoong Jin, and Hyunsoo Yoon examine this problem in Anti-coherence Regularized Absolute Scaling for Out-of-Distribution Detection.1

One result illustrates the magnitude. On ImageNet with a SwinV2-B backbone, the existing SCALE method records 51.24% AUROC and 89.40% FPR@95. Replacing its signed-activation calculation with an absolute-magnitude calculation raises AUROC to 81.05% and lowers FPR@95 to 61.72% (Section 5.3, Table 4).

AUROC measures how effectively a detector ranks familiar and unfamiliar images. FPR@95 measures the share of unfamiliar images incorrectly accepted as familiar when 95% of familiar images are accepted. Higher AUROC and lower FPR@95 are desirable.

The size of this change raises a specific engineering question: what assumption inside the original detector became unreliable when the classifier architecture changed?

Signed feature sums can conceal the signal being measured

An image classifier converts an input into an internal feature vector before its final classification layer produces class scores. This vector, called the penultimate activation, may contain only non-negative values on some architectures and substantial negative values on others.

SCALE computes a percentile-based ratio from these activations. It compares their total sum with the sum of values above a selected percentile, then uses the resulting ratio to determine how strongly to amplify the feature vector.

The problem is that the original sums retain their signs. Positive and negative components can cancel, leaving a ratio that no longer reflects the concentration of activation magnitude.

The authors’ Absolute Fix replaces signed values with their absolute magnitudes when calculating that ratio:

$$ r_{\mathrm{abs}}= \frac{\sum_i |z_i|} {\sum_{i:|z_i|\gt P_{p,\mathrm{abs}}(\mathbf z)}|z_i|} $$

Here, $z_i$ denotes an internal feature and $P_{p,\mathrm{abs}}$ is the chosen percentile threshold over absolute feature magnitudes.

Crucially, the method does not replace the classifier’s features with their absolute values. It uses magnitudes to calculate a more stable scaling factor, then scales the original signed features. Negative activation information remains available to the classification layer.

The component ablation in Section 5.3, Table 4 tests this correction separately from the complete method. Across seven negative-activation ImageNet backbones, absolute scaling improves average AUROC and FPR@95 relative to a ReLU-based correction that removes negative contributions from the ratio calculation.

On EfficientNet-B1, for example, original SCALE achieves 50.79% AUROC, while the Absolute Fix reaches 80.89%. The correction is not universal: on DeiT3-B, absolute scaling ties original SCALE in AUROC.

The evidence supports a narrower conclusion than architecture independence. Measuring activation magnitudes can correct a failure caused by signed aggregation, but the resulting detector still needs to be evaluated on each backbone.

Correcting the ratio leaves a second amplification problem

Once the scaling factor is repaired, another question remains: should every image receive the same degree of amplification implied by its activation magnitudes?

Uniform scaling strengthens the entire feature vector, including contributions that work against the predicted class.

The paper’s theoretical analysis establishes that, under a zero-bias linear classifier with a unique positive maximum logit, increasing the scaling factor makes that logit’s contribution increasingly dominate the energy-based OOD score (Section 4, Lemma 4.1).

This motivates examining how the predicted class score is assembled.

Each feature contributes through its value multiplied by its corresponding classifier weight. Some products raise the predicted class score; others lower it. Anti-coherence measures the opposing contributions as a fraction of total absolute contribution magnitude:

$$ A(\mathbf z)=\frac{C_n}{C_p+C_n} $$

Here, $C_p$ is the sum of positive feature-weight contributions and $C_n$ is the absolute sum of negative contributions. Anti-coherence therefore depends on both features and classifier weights. It is not simply a count of negative activations.

The authors observe that this fraction tends to be higher for OOD images in their ImageNet distribution analyses (Figure 3 and Appendix C). They use it to reduce the corrected scaling factor:

$$ r_{\mathrm{anti}}=r_{\mathrm{abs}}-\lambda A(\mathbf z) $$

A larger opposing fraction produces stronger damping, governed by the regularization parameter $\lambda$.

The incremental ablation in Section 5.3, Table 5 is especially informative. Adding Anti-coherence to absolute scaling improves average AUROC on all ten ImageNet backbones, while FPR@95 improves on nine. ViT-B/16 is the exception: its FPR@95 increases from 61.10% to 63.18%, despite improved AUROC.

This separates the two contributions. Absolute scaling repairs how the activation ratio is measured; Anti-coherence adds a classifier-weight-aware signal that moderates amplification. The second contribution improves aggregate discrimination without guaranteeing an improvement in every operating metric.

The theoretical result also has a boundary: the implemented classifier retains bias terms, whereas the lemma assumes zero bias. The proof motivates the mechanism but does not establish its behavior for every deployed classification head.

Stronger than SCALE does not mean strongest everywhere

The principal OpenOOD v1.5 benchmark compares Anti-coh with existing activation-shaping and logit-based detectors. The supplementary experiments assess broader alternatives and additional datasets.

Evaluation setting SCALE AUROC / FPR@95 Anti-coh AUROC / FPR@95
ImageNet, 10 backbones, 5 OOD datasets 71.28% / 70.39% 85.81% / 52.76%
CIFAR-10, ResNet-18, 6 OOD datasets 86.21% / 44.13% 91.62% / 40.68%
CIFAR-100, ResNet-18, 6 OOD datasets 80.64% / 76.17% 79.38% / 79.77%

Sources: Section 5.2, Tables 2–3; Appendix G, Table 11. Entries are benchmark averages, not measurements of production failure rates.

The ImageNet comparison shows a 14.53-percentage-point AUROC gain and a 17.63-point FPR@95 reduction over SCALE. This is substantial evidence for the complete method against its direct predecessor.

The other comparisons qualify that result.

On CIFAR-10, Anti-coh achieves the highest average AUROC among the principal baselines, but the plain Energy detector has a slightly better FPR@95: 40.38% versus 40.68%.

The CIFAR-100 supplementary evaluation reverses the comparison with SCALE. Several other methods also outperform Anti-coh there.

Representation-based detectors offer another alternative. In Appendix D, CADRef achieves 86.24% AUROC on ViT-B/16 compared with Anti-coh’s 82.92%. Mahalanobis reaches 87.99% on SwinV2-B compared with Anti-coh’s 85.06%. These methods can exploit training-set feature statistics that Anti-coh avoids.

The fine-grained SSB-Hard near-OOD benchmark remains challenging as well. Strong average results across unfamiliar-image datasets should not be assumed to cover visually similar but unsupported categories.

These experiments serve different evidential purposes: the main benchmarks establish comparative performance, the ablations isolate mechanisms, and the supplementary comparisons expose where lightweight inference-time methods remain less competitive.

A training-data-free detector still needs calibration data

Anti-coh scores an image using the frozen classifier’s internal features and classification weights. It requires neither original training-set statistics nor test-time gradients.

That reduces specific data-access and inference requirements, particularly when an operator has inherited a pretrained model without the data used to develop it.

However, the reported implementation selects both the percentile parameter $p$ and regularization strength $\lambda$ using validation AUROC on examples containing familiar and unfamiliar images (Appendix F.2). The detector is training-data-free at inference, not tuning-data-free during development.

For an organization replacing a vision backbone, Cognaptus infers a concrete evaluation workflow:

  1. Rebenchmark the detector after each backbone change. Test familiar inputs and representative unfamiliar categories rather than assuming the earlier detector remains calibrated.
  2. Compare operating-point errors, not only AUROC. Set an acceptance threshold using the organization’s tolerance for rejecting supported images and accepting unsupported ones.
  3. Test difficult near-OOD categories separately. An unfamiliar category resembling a supported one may be more consequential than an easily distinguished far-OOD image.
  4. Define exception handling before production adoption. A proposed inference gateway could route flagged images for recapture, abstention, or another verification path rather than automatically using the classification.
  5. Compare the cost of alternative detectors. Where training-set feature statistics are accessible, a representation-based approach may justify its additional data and infrastructure requirements.

This is a proposed operating procedure, not an evaluated deployment from the paper. The study does not measure downstream intervention quality, false-acceptance costs, or the effect of routing decisions on a live service.

The decision is whether the detection layer survived the upgrade

Anti-coh demonstrates that a lightweight OOD detector can improve markedly when its calculations respect the internal feature structure of the classifier it monitors. It also shows that correcting activation magnitudes and moderating conflicting class contributions address different problems.

The strongest evidence supports better benchmark discrimination than SCALE across the evaluated ImageNet architectures. The CIFAR-100 results, representation-based comparisons, and difficult near-OOD cases prevent treating that result as universal.

For operators, the decisive lesson concerns model changes: a replacement backbone requires renewed evaluation of the unfamiliar-image detector, even when the application’s predictions and interface still look familiar. Anti-coh expands the available options when original training data are inaccessible, but the decision to accept or escalate an image must still be grounded in representative validation and the costs of errors in the actual workflow.

Cognaptus: Automate the Present, Incubate the Future.


  1. Donghwan Kim and Youngwoong Jin and Hyunsoo Yoon (2026). Anti-coherence Regularized Absolute Scaling for Out-of-Distribution Detection. arXiv:2610.32255. https://arxiv.org/abs/2610.32255 ↩︎