Skip to main content
Back to timeline
arXivSource publication:

No training required: PRISM pairs frozen random networks with Mahalanobis scoring to top 47 tabular datasets at 86.4% mean AUROC

Synopsis

Anomaly detection has largely been treated as a representation-learning problem, with gains coming from trained features or pretrained backbones. PRISM combines frozen randomly initialized neural features with Mahalanobis scoring against normal references, reaching the highest mean AUROC of 86.4% and the second-highest mean AUPRC of 65.9% across 47 ADBench tabular datasets, and mean image-level and pixel-level AUROCs of 90.4% and 96.4% across 27 visual categories in MVTec AD and VisA, all without optimizing neural weights or using pretrained representations.

Source-provided article image: How Far Can Untrained Neural Networks Go in Anomaly Detection? From Tables to Pixels
arXiv

Interpretation

Anomaly detection can be competitive without learning representations at all: PRISM treats a randomly initialized, frozen neural network as a feature transform and scores Mahalanobis distance using means and covariances estimated from normal samples, sharing this idea across tabular and image modalities. Earlier untrained-neural-network anomaly detection was validated on only a small set of tabular datasets, and its visual experiments were limited to flattened MNIST and F-MNIST with image-level scores only, leaving open whether it could cover diverse tabular data or pixel-level localization. On the 47 ADBench tabular datasets, PRISM-T reaches a mean AUROC of 86.4% and a mean AUPRC of 65.9%, ranking first and second among compared methods; it shows no significant difference from leading baselines including DTE-NP, LUNAR, AnoDDAE, TCCM, K-DSM and DenoiserAD on either metric.

The same random-feature-plus-Mahalanobis recipe extends from tables to pixels: PRISM-V delivers both image-level detection and location-wise localization across 27 visual categories in MVTec AD and VisA. Earlier untrained visual experiments scored whole images only and did not perform spatially coherent pixel-level localization on industrial defect benchmarks. PRISM-V reaches image-level AUROCs of 92.9% and 87.3% on MVTec AD and VisA, pixel-level AUROCs of 96.2% and 96.7%, and AUPROs of 85.9% and 73.1%, with only 0.33M parameters; a gap remains relative to pretrained-feature methods, smallest in pixel-level AUROC.

The random transform alone shifts the score gap between anomalous and normal samples, and the nonlinear activation amplifies that effect. Direct Mahalanobis scoring in observation space gives tabular mean AUROC of 82.4% and AUPRC of 56.8%, and visual image-level AUROC of 77.7% with AUPRO of 62.4%. Removing the activation and keeping only the random linear transform raises tabular AUROC to 83.2% and AUPRC to 59.3%; adding LeakyReLU brings them to 86.4% and 65.9%, while visual image-level AUROC and AUPRO rise to 90.4% and 80.2%.

Perspective

Evaluation covers 47 ADBench tabular datasets plus 15 MVTec AD categories and 12 VisA categories; tabular experiments use a semi-supervised setting that builds references from normal samples only, and visual experiments follow each benchmark's official one-class protocol. The method targets researchers and engineers who need an anomaly-detection baseline or want to gauge the gain from learned representations, with one global configuration per modality chosen on three preselected development datasets and then fixed.

Visually, PRISM-V still trails pretrained-feature methods, varies across categories such as Screw, Capsules and Macaroni2, and in low-AUPRO categories its anomaly responses spread into normal regions. The theoretical analysis indicates that a general improvement in standardized score separation under nonlinear transformation is hard to guarantee, so that conclusion rests on empirical observation in both modalities.