Researchers showed that anomaly detection can work well even without training the network. The PRISM framework feeds data through a fixed block of random neural features followed by a simple Mahalanobis distance score that uses normal‑reference statistics. In tabular data, PRISM‑T achieved an average 86.4 % AUROC and 65.9 % AUPRC across 47 benchmark datasets — the best mean AUROC and the second‑best mean AUPRC among state‑of‑the‑art methods. In visual data, PRISM‑V matched 90.4 % image‑level and 96.4 % pixel‑level AUROC on 27 categories of the MVTec AD and VisA datasets. Both models reached these results without any weight optimization or pretrained representations.

PRISM‑T uses one PRISM block consisting of a linear transform, instance and batch normalizations, and a LeakyReLU activation. The Mahalanobis score uses a λ of 10⁻³. PRISM‑V interleaves five such blocks with blur‑pool downsampling, concatenates the mid‑level blocks, and produces a dense anomaly map.

When compared to methods that rely on large pretrained models, PRISM‑V performs competitively on image‑level detection. It attains 92.9 % image‑level AUROC and 96.2 % pixel‑level AUROC on MVTec AD, but it falls short of the best pretrained‑feature methods such as PatchCore (98.9 % image AUROC) and WinCLIP+ (97.2 % image AUROC). PRISM‑V also uses only 0.33 M parameters, less than 0.5 % of the 68.9 M of PatchCore or the 208 M of WinCLIP+. However, its pixel‑level localization is lower; the AUPRO score is 80.2 % compared to 98.4 % for PatchCore.

These findings show that random nonlinear transformations can provide a strong baseline for anomaly detection. They also allow quick deployment with minimal computational cost, but for tasks that demand very precise pixel‑level localization, pretrained features still offer an advantage.