Anomaly detectors rank unfamiliar examples for expert confirmation, and the confirmed examples teach a classifier to recognise a new class. Across five benchmarks from physics, medicine and computer vision, the members flagged by the detector are those the classifier learns first. In three systems, confirming the highest-ranked members of a new class gives lower recognition than confirming the same number of random members. Handing subsequent selection to the promoted classifier recovers part of this loss in two systems. A paired intervention further shows that replacing nearby known classes with distant ones sharply reduces the confirmations needed for recognition. These results distinguish finding a new class from selecting examples that teach its extent: within the tested classes, concentrating confirmations at the top of the anomaly ranking makes learning less efficient.
Figures & tables
Figure 1: Discovery and learning favour the same members of a new class. Top: an anomaly detector flags an unfamiliar population, experts confirm selected members, and the confirmed examples enter training as a new class. Recognition is measured at a fixed false-alarm rate; the label cost N(f) is the estimated number of confirmations needed to reach efficiency f . Bottom: probability that a held-out member is recognised at a 1% false-alarm rate after promotion with n random confirmations, separately for members flagged before any label (magenta) and members not flagged (dark). Curves average over withholds with at least three members in each group; parentheses give the number of withholds.
Figure 2: Detector-selected confirmations slow new-class learning. (a) Mean recognition at a 1% false-alarm rate after promotion with n random members (dark), the n highest-ranked members under the anomaly detector (magenta), or the n least anomalous members (light). Gravity Spy uses all twenty-two classes; CIFAR-100 uses thirteen classes with zero-label detectability between 0.1 and 0.9 ; ECG uses thirty-five withholds of eleven classes in geometry-selected sub-taxonomies. (b) Class-averaged differences from random selection with 95% bootstrap intervals over classes. Gravity Spy at five, ten and twenty confirmations averages three training seeds. Each ECG class enters once, through its median over the sub-taxonomies containing it.
Figure 3: Handing later confirmations to the classifier recovers part of the penalty. Recognition at a 1% false-alarm rate for the hand-off cohorts: eight Gravity Spy classes with zero-label detectability between 0.1 and 0.9 , thirteen CIFAR-100 classes, and thirty-five ECG withholds involving eleven classes. Following two detector-selected confirmations, the promoted classifier selects further members by sampling among its alerts (dashed) or by highest promoted-class probability (dotted). Random and detector-selected policies are evaluated on the same cohorts.
Figure 4: Known-class context changes new-class learning cost. (a) Estimated confirmations needed for 80% recognition at a 1% false-alarm rate for each Gravity Spy class, among its three nearest eligible known classes (dark) or its three farthest (magenta). Two further background classes are shared. Costs are computed separately in four replicates and averaged within each class. (b) CIFAR-100: best recognition through forty confirmations with semantic neighbours versus distant classes. Five further background classes and the confirmation draws are shared; ninety-nine of one hundred points lie above the diagonal.
Comparison
Result
Interpolated N80 , cheaper : dearer
18:2
Median N80 , near → far
4.3→1.6
Replicate cheaper : dearer counts
15:5,19:1,17:1,17:3
First measured crossing, cheaper : dearer
19:1
Median first crossing, near → far
6.2→2.2
Two frequent far classes excluded
20:0
Supplementary Table 1: Gravity Spy neighbour-swap controls on deduplicated data. Counts exclude ties. Costs are computed per replicate and averaged within class.
System
n
Random
Detector
Least
Alerts
Top
Margin
Gravity Spy (8 classes)
5
0.42
0.30
0.48
0.42
0.35
0.42
10
0.72
0.43
0.66
0.61
0.42
0.53
20
0.89
0.67
0.79
0.82
0.73
0.83
CIFAR-100 (13 classes)
5
0.75
0.68
0.69
0.74
0.67
0.75
10
0.81
0.68
0.75
0.77
0.68
0.80
20
0.82
0.72
0.77
0.80
0.70
0.83
Supplementary Table 2: Mean recognition at a 1% false-alarm rate for the hand-off cohorts. “Least” selects the lowest anomaly scores; “alerts”, “top” and “margin” follow the two detector-selected confirmations. Randomised policies average three draws. The thirty-five ECG withholds involve eleven classes and are not independent class replicates.
System
B
Detector
Alerts
Hand-off
Gravity Spy
5
0.23 (3.8)
0.31 (3.1)
0.23 (3.8)
(8 classes)
10
0.42 (7.8)
0.42 (6.2)
0.49 (7.0)
20
0.57 (15.6)
0.66 (12.9)
0.68 (14.4)
40
0.74 (29.1)
0.79 (25.2)
0.88 (31.3)
CIFAR-100
5
0.54 (3.2)
0.54 (1.7)
0.54 (3.2)
(13 classes)
10
0.67 (5.9)
0.67 (3.4)
0.72 (6.5)
Supplementary Table 3: Mixed-stream confirmation. Mean recognition of the target class at a 1% false-alarm rate, with the mean number of target members among the confirmations in parentheses. Randomised policies average three draws.
This paper considers a practical few-shot anomaly detection (FSAD) setting, termed discriminative FSAD, where a limited number of both normal and anomalous examples are available as references during inference. Existing FSAD methods rely on normal-only references through normality matching, ignoring the discriminative clues in anomalous references, while directly fitting both references can overfit to the seen anomalies. We introduce IDEAL, an intrinsic deviation learning framework that leverages both reference types to learn intrinsic deviation patterns characterizing generalizable abnormality as deviations from normality. IDEAL decomposes the learning process into two novel components: 1) a Normal Variation Eraser to suppress nuisance normal variations that may lead to noisy deviations from normality, thereby highlighting anomaly-relevant deviation representations; 2) an Intrinsic Deviation Encoder to decompose these denoised deviation representations into intrinsic deviation vectors capturing the most discriminative orthogonal deviation directions. At inference, IDEAL scores query-to-normal deviations preserved after projection onto the learned intrinsic deviation vectors, enabling generalization for both seen and unseen anomalies. Extensive experiments on eight real-world datasets show that IDEAL generalizes effectively to unseen anomalies and consistently outperforms existing state-of-the-art FSAD methods. Code and data are available at https://github.com/mala-lab/IDEAL.
Huan Wang, Jun Shen, Jun Yan +1
Singapore Management University, Singapore · University of Wollongong, Australia
Anomaly detection (AD) is often framed as a representation-learning problem, where strong performance is expected to require learned features or AD-specific optimization. However, this view can overlook the intrinsic separability of anomalies from normal references, even in the absence of learned representations. We therefore investigate how far untrained neural networks can go in anomaly detection. To this end, we propose PRISM, a framework combining fixed random nonlinear features with normal-reference Mahalanobis scoring for tabular detection and dense visual localization. Across 47 tabular datasets in ADBench, PRISM-T achieves the highest mean AUROC (86.4%) and the second-highest mean AUPRC (65.9%) among the compared methods. On MVTec AD and VisA, covering 27 visual categories, PRISM-V achieves mean image-level and pixel-level AUROCs of 90.4% and 96.4%, respectively. Notably, both models achieve these results without neural weight optimization or pretrained representations. A theoretical analysis further characterizes how nonlinear transformations affect the mean and covariance contributions to the Mahalanobis score gap, while empirical analyses examine the resulting score separability. These findings establish PRISM as both a competitive anomaly detector and a prior-to-learning baseline for assessing the added value of learned representations in anomaly detection.
Hyojun Yi, Geunyoung Kim, Hogeon Seo +1
Department of Artificial Intelligence and Robotics, Sejong University, Seoul, Republic of Korea · Applied AI Section, Korea Atomic Energy Research Institute, Daejeon, Republic of Korea
Detecting anomalies in link streams that represent various kinds of interactions is an important research topic with crucial applications. Because of the lack of ground truth data, proposed methods are mostly evaluated through their ability to detect randomly injected links. In contrast with most proposed methods, that rely on complex approaches raising computational and/or interpretability issues, we show here that trivial graph features and classical learning techniques are sufficient to detect such anomalies extremely well. This basic approach has very low computational costs and it leads to easily interpretable results. It also has many other desirable properties that we study through an extensive set of experiments. We conclude that detection methods should now target more complex kinds of anomalies.
Matthieu Latapy, Stephany Rajeh
CNRS, LIP6 Sorbonne Universit´e Paris, France · Efrei Research Lab Efrei Paris Panth´eon-Assas Universit´e Villejuif, France