PLUS adds a class-specific signal to help AI spot unfamiliar images

The difficult case: unfamiliar but similar

Image classifiers may encounter inputs outside the categories represented in their training data. Out-of-distribution, or OOD, detection aims to flag those inputs. The paper distinguishes far-OOD images, which are semantically distant from familiar categories, from near-OOD images, which resemble them visually and semantically.

In their November 4, 2026 preprint, Yonsei University researchers Wonseok Shin and Songkuk Kim argue that this distinction matters geometrically. Far-OOD inputs often activate fewer feature directions characteristic of their predicted class. Near-OOD inputs can activate those directions much like familiar images, making broad measures of confidence or feature atypicality less informative. Their proposed response is PLUS: Per-class Low-dimensional sUbspace Signal.

A compact view of each class

PLUS works with a frozen, pretrained classifier. It uses normalized features from the layer immediately before classification and trains a separate sparse autoencoder on in-distribution training features. That autoencoder learns a dictionary of feature directions; the authors use it as a geometric tool, not as evidence of human-interpretable concepts.

The dictionary contains 6,144 neurons, with at most 32 active for each input. For each class, PLUS selects the 16 neurons with the highest average activation across that class’s training examples. Their decoder directions define a compact subspace—a restricted set of coordinates for examining the class. Orthonormalizing those directions supplies perpendicular, unit-length axes.

Crucially, the selected neurons stay fixed for each class. Individual images activate different sparse-autoencoder neurons, so using those changing selections directly would not provide consistent class-specific coordinates. PLUS estimates a mean and covariance within each fixed subspace, using at most 300 training images per class.

Adding evidence, not replacing the detector

For a new image, PLUS uses the classifier’s predicted class and measures deviation within that class’s subspace. Its squared Mahalanobis distance accounts for the class’s usual variability: departures along normally stable directions receive more weight than departures along variable ones. Components outside the selected subspace are discarded.

PLUS then standardizes both this signal and the existing detector’s score using correctly classified in-distribution validation examples. It adds half the standardized PLUS signal to the standardized base score. The weight, 0.5, is fixed across all reported backbones and datasets; no auxiliary OOD examples are needed for fitting or calibration. The classifier remains unchanged, although the separate autoencoder requires training.

What the benchmarks show

The experiments use OpenOOD v1.5 with ImageNet-1K as the familiar distribution. Near-OOD datasets are SSB-hard and NINCO; far-OOD datasets are iNaturalist, Textures and OpenImage-O. Eighteen detectors are evaluated across seven backbones, spanning convolutional, transformer and hybrid architectures.

The authors report higher average near-OOD AUROC in all 126 detector–backbone combinations. AUROC measures how well scores rank unfamiliar inputs against familiar ones across thresholds; it is not classification accuracy. For example, adding PLUS to MDS++ raises near-OOD AUROC from 79.87 to 81.09 on ViT-B/16, and from 78.92 to 80.86 on Swin-T.

The averages do not mean every individual dataset improves. All evaluated combinations improve on SSB-hard, but one detector on Wide-ResNet50-2 loses 0.23 AUROC points on NINCO. Far-OOD AUROC improves in 321 of 378 individual comparisons, with reported declines no larger than approximately 1.8 points. Textures concentrates the trade-off, particularly on ViT-B/16 and Swin-T.

What the additional analyses establish

Ablations support contributions from both the selected subspace and its class-specific covariance. Random subspaces provide essentially no near-OOD gain, while shared or isotropic metrics recover only part of the improvement. Adding PLUS also improves near-OOD AUROC in all 714 tested two-detector ensemble configurations, supporting its usefulness beyond simply pairing existing detectors.

The supplementary comparisons qualify the sparse autoencoder’s advantage. Per-class principal component analysis leads on five of seven backbones; the SAE construction leads on only one. Its strength is consistency: its largest shortfall from the best tested subspace is 0.93 AUROC points, versus 1.91 for per-class PCA. Related classes also reuse selected neurons more often than unrelated classes, providing shared structure for analysis without establishing semantic interpretability.

Costs and limits

On cached Swin-T features, autoencoder fitting reportedly takes 7–9 minutes on an NVIDIA RTX A5000. Reported signal computation takes 0.4–0.7 microseconds per image, while stored class statistics require 32–126 MiB, depending on feature dimension.

The fixed fusion weight is a default, not a demonstrated optimum. Smaller weights can reduce far-OOD costs, and sensitivity varies by backbone. Evidence remains limited to ImageNet-based image classification: smaller datasets, massive label spaces and object detection are proposed extensions. The authors also leave rigorous conditions for subspace separability to future work. These are preprint findings, not peer-reviewed validation or a guarantee of reliable deployment.