Training-Free Human-in-the-Loop Anomaly Detection via Memory Bank Correction

Ayusha Abbas; Saram Abbas; Kabita Adhikari

Source-linked AI summary

Training-Free Human-in-the-Loop Anomaly Detection via Memory Bank Correction

Ayusha Abbas, Saram Abbas, Kabita Adhikari

arXiv:2608.17775v1Aug 18, 2026cs.LG

TL;DR

Cold-start anomaly detection offers little training data and no way for quality engineers to incorporate their judgments into deployed models. This paper introduces training-free PatchCore memory-bank correction and finds that ten golden samples plus operator corrections outperform hundreds of samples without corrections.

Problem
Cold-start production lines have almost no training data, while quality engineers cannot incorporate their judgments about false alarms and defects into existing anomaly detectors.
Method
The framework directly edits a PatchCore memory bank using reviewed false positives and false negatives without retraining, gradients, or original training data, evaluated on held-out images.
Results
Ten golden samples plus corrections outperform hundreds of samples without corrections, significantly improving 12 of 15 categories and harming none.
Takeaways & Limitations
The framework can produce correctly labelled deployment-time data without retraining at 43% of exhaustive-review labelling cost.
Takeaways & Limitations
Feedback is simulated as perfectly accurate, so live trials are needed before deployment, especially because one mistaken normal label can severely contaminate a small cold-start bank.

Abstract

from arXiv · show

Anomaly detectors are hardest to deploy exactly where training data is scarcest: a newly commissioned production line has a handful of verified "golden" samples and no machine-learning engineer on the factory floor. We present a training-free human-in-the-loop framework in which a domain expert corrects a PatchCore detector by direct memory bank editing: no retraining, no gradients, no original training data. A false-positive correction inserts the reviewed image's normal patches through a self-calibrating novelty gate admitting only those beyond the median pool-normal nearest-neighbour distance. From a bank built on only ten golden samples, operator corrections close a median 66% of the gap to an uncorrected fully trained bank (mean 80%, raised by three categories that overshoot parity), significantly improving 12 of 15 MVTec AD categories and harming none: ten samples plus corrections outperform hundreds of samples without them. On already-trained banks the headroom is smaller and concentrated where the bank undersamples normal appearance (gated: toothbrush +0.10, metal nut +0.09, zipper +0.05, screw +0.05), and no category except grid is significantly harmed. Evaluation uses a held-out protocol (20 splits per category, Holm-corrected Wilcoxon), because corrected images entering the bank inflate naive evaluation toward AUROC 1.0 by memorisation. Passive and active querying are statistically indistinguishable; a matched-label-budget control attributes gains to deployment-time label production at 43% of exhaustive-review cost; a defect-memory extension fails decisively. Feedback is simulated from ground truth; live expert trials, where mislabelling is costliest on small banks, remain future work.

I. INTRODUCTION · A. Contributions

The paper addresses cold-start and deployment drift by letting quality engineers correct PatchCore directly through training-free memory-bank edits. Across all 15 MVTec AD categories, corrections substantially recover performance from ten golden samples while held-out evaluation and novelty gating limit memorisation and degradation.

I. INTRODUCTION: Deployment accuracy degrades as normal appearance drifts, while periodic retraining requires labelled data, infrastructure, and ML expertise often unavailable on factory floors.Newly commissioned lines intensify the problem because they have almost no training data, although quality engineers can already identify false alarms and genuine defects.
I. INTRODUCTION: The framework applies a quality engineer’s binary normal-or-defect judgment as an immediate direct edit, requiring no retraining, ML engineer, or original training data.This is a training-free human-in-the-loop approach designed for deployment-time correction.
I. INTRODUCTION: PatchCore enables interpretable correction because its normal appearance is stored in an explicit patch-feature memory bank rather than buried in network weights.False-positive correction adds normal patches, whereas false-negative correction removes nearby patches to raise future scores for similar defects.
I. INTRODUCTION: A single simulated review session cleared held-out normal images while leaving a held-out defect flagged, demonstrating correction beyond the reviewed images.The figure reports 16–25 reviews for the displayed seeds, with category means of approximately 13 and 20.
A. Contributions: 66% median gap closure from ten golden samples was achieved relative to an uncorrected fully trained bank, with 12 of 15 categories significantly improved and none harmed.The mean gap closure was 80%, raised by three categories overshooting parity; ten samples plus corrections outperformed hundreds of samples without corrections.
A. Contributions: FP insertion reproduced the entire gain, while FN removal was a measured near no-op, showing that the mechanism teaches normality rather than memorising defects.Corrections use direct memory-bank edits without retraining, gradients, or training data.
A. Contributions: The novelty gate admits a patch only when its nearest-neighbour distance exceeds the median pool-normal distance, eliminating held-out degradation on five of six well-covered categories.The gate preserved every category that benefited across a sensitivity sweep, leaving grid as the single category still significantly harmed.
A. Contributions: Passive and active querying were statistically indistinguishable on every category, so no acquisition function is needed when bank edits act globally and the review budget exhausts the error pool.The evaluation used 15 MVTec AD categories, 20 independent splits per category, and Holm-corrected Wilcoxon tests; held-out scoring prevents reviewed-image memorisation from inflating AUROC toward 1.0.

II. RELATED WORK … D. Continual Learning and Test-Time Adaptation

The paper positions its deployment-time memory-bank correction at the intersection of unsupervised anomaly detection, human-in-the-loop learning, and active querying. It extends prior static, parameter-updating, and few-shot approaches with a correction layer applicable to memory-bank models.

II. RELATED WORK: The work combines unsupervised anomaly detection, human-in-the-loop machine learning, and active learning query strategies.
A. Unsupervised Anomaly Detection: Memory-bank methods dominate anomaly detection, with PatchCore refining SPADE through a greedy coreset of locally-aware patch features.PatchCore reports 99.6% mean AUROC on MVTec AD with a three-backbone ensemble at 320×320 resolution, while deployment-time adaptation remains future work.
B. Human-in-the-Loop Machine Learning: Prior human-in-the-loop anomaly-detection work uses expert feedback for supervised retraining, reweighting, re-optimisation, or bidirectional time-series detection.
C. Active Learning and Query Strategies: Although uncertainty sampling commonly outperforms passive random sampling, this regularity does not transfer to memory-bank correction.Under multi-seed held-out evaluation, uncertainty sampling fails to outperform random sampling.
D. Continual Learning and Test-Time Adaptation: Continual learning and test-time adaptation improve deployment performance through parameter updates, requiring regularisation against catastrophic forgetting.PatchCoreCL updates per-task memory banks as tasks arrive, whereas FR-PatchCore updates the memory bank during training via a negative cosine similarity loss.
D. Continual Learning and Test-Time Adaptation: Few-shot methods construct detectors from k samples, but treat the residual gap to full-training performance as fixed after deployment.Examples include RegAD, WinCLIP, GraphCore, and FastRecon; the paper’s cold-start evaluation instead begins with a bank built from ten samples.
D. Continual Learning and Test-Time Adaptation: The proposed method applies as a deployment-time correction layer on top of any memory bank model.

E. Positioning … B. Human-in-the-Loop Feedback Framework

The paper positions its approach among existing paradigms and defines a training-free PatchCore-based human-in-the-loop correction layer. The method uses explicit patch-memory editing, nearest-neighbour anomaly scoring, and iterative one-label feedback without retraining or gradients.

E. Positioning: Table I compares the approach with existing paradigms across four key desiderata.
III. METHODOLOGY: The methodology covers PatchCore, HITL correction with a novelty gate, querying strategies, AUCC, and experimental protocols.
A. PatchCore Baseline: PatchCore represents normal appearance with locally-aware patch features from a frozen ImageNet-pretrained WideResNet-50 memory bank.Feature maps from backbone levels 2 and 3 are combined into 1,536-dimensional descriptors at 784 spatial positions.
1) Features and Memory Bank:: 1% of training patches are retained by greedy k-center coreset subsampling, producing banks from N = 470 to 3,065 entries.The uncompressed union can contain some 200,000 vectors for a typical 200–300 image category.
2) Anomaly Scoring:: At inference, the detector extracts 784 patch features and scores each image using the maximum patch-level distance to the memory bank.Each patch distance is the mean over its k = 9 nearest bank entries.
2) Anomaly Scoring:: The maximum patch score targets small local defects, while the same nine distances provide anomaly localisation and mean aggregation reduces sensitivity to one bank entry.Retrieval uses a ball-tree index.
2) Anomaly Scoring:: An image is anomalous when s(x) ≥τ, with τ set to the 90th percentile of normal-training scores and recomputed after every correction round.The implementation uses a single WideResNet-50 at 224×224 without score reweighting and applies the mean-of-nine statistic consistently across baselines and corrected banks.
B. Human-in-the-Loop Feedback Framework: The post-deployment correction layer edits only the explicit memory bank, leaving the backbone, training data, and gradients untouched.Each round scores images, presents one error-pool image to an expert, receives one binary label, and applies the corresponding bank correction until the round limit or pool exhaustion.

1) False Positive Correction: Memory Bank Expansion: · 2) Novelty-Gated Insertion: · 3) False Negative Correction: Memory Bank Surgery:

The framework corrects false positives by expanding or novelty-gating the memory bank, while false-negative correction surgically removes locally isolated bank patches. Corrections are applied immediately, with bounded bank size and explicit safeguards against redundant insertion and criterion mismatch.

1) False Positive Correction: Memory Bank Expansion:: A false positive signals an underrepresented region of normal appearance space, where patch features lie anomalously far from the current memory bank.The detector flags a normal image when its anomaly score meets or exceeds τ.
1) False Positive Correction: Memory Bank Expansion:: 784 patch features from each corrected false positive are added directly to the memory bank to fill the missing normal-appearance region.Each correction increases |M| by 784 entries.
1) False Positive Correction: Memory Bank Expansion:: Repeated expansion is capped at Nmax through random subsampling, preserving coreset coverage while bounding nearest-neighbour query cost.The cap is applied whenever insertion would leave |M| above Nmax, and removing it was not a sensitive control condition.
2) Novelty-Gated Insertion:: Novelty gating inserts a candidate patch only when its nearest-neighbour distance exceeds τnov, the median distance of review-pool normal patches to the baseline bank.The threshold is fixed before correction, uses pool data only, and requires no tuning.
2) Novelty-Gated Insertion:: On a toothbrush false positive, the gate rejected patches within existing coverage and inserted only the genuinely novel patches.The gate is permissive for undersampled banks and conservative for well-covered banks.
3) False Negative Correction: Memory Bank Surgery:: A false negative identifies an anchor patch by maximum 1-NN distance, then removes a local bank gap so similar defects receive higher future scores.The removal size kremove = 5 balances gap size against over-removal.
3) False Negative Correction: Memory Bank Surgery:: False-negative correction can behave as a near no-op because the 1-NN anchor criterion differs from the image score’s mean-of-nine d9 criterion.Removal targets the most locally isolated patch, not necessarily the patch that determined the image score.

C. Querying Strategies

Querying strategies change only the order in which errors are reviewed, not the correction mechanism. The framework compares prevalence-proportional random review with uncertainty sampling near the decision threshold.

C. Querying Strategies: Querying strategies determine only the order in which errors are reviewed, not the correction mechanism.The evaluation compares passive and active querying.
C. Querying Strategies: Passive querying samples uniformly at random from the error pool, requiring no computation beyond scoring and preserving false-positive/false-negative prevalence.It samples FP and FN errors in proportion to their prevalence.
C. Querying Strategies: Active querying uses uncertainty sampling to select the image whose score is closest to the decision threshold.Its rationale assumes boundary-adjacent samples are most informative when updates are local.
C. Querying Strategies: The boundary-adjacent informativeness assumption fails for memory bank correction because its updates are global rather than local.The paper refers to empirical evidence in Section IV and global updates in Section V.

D. Evaluation Metric: Area Under the Correction Curve (AUCC) … 2) Category Selection: Stratified by Separation Index:

The evaluation uses AUCC to measure how efficiently human correction improves AUROC, while experiments span all 15 MVTec AD categories and stratify interpretation by separation index. This setup emphasizes early gains and category difficulty, from overlapping to fully separated score distributions.

D. Evaluation Metric: Area Under the Correction Curve (AUCC): AUCC measures correction efficiency by integrating AUROC gains over the review budget rather than comparing only before-and-after AUROC.For round r, the per-round gain is gr = max(0, AUROCr − AUROC0).
D. Evaluation Metric: Area Under the Correction Curve (AUCC): AUCC rewards reaching the same final AUROC with fewer human reviews and remains comparable across baselines.It normalizes cumulative gain against immediate, sustained perfect correction and adapts area-under-the-learning-curve evaluation to correction.
1) Dataset:: Experiments use MVTec AD, comprising 5,354 images across 15 categories and 73 defect types.The benchmark includes five textures and ten objects, with defect-free training images and mixed-class test images.
2) Category Selection: Stratified by Separation Index:: All 15 MVTec AD categories are evaluated with both passive and active querying strategies.The separation index σ is used to predict correction utility and interpret results across categories.
2) Category Selection: Stratified by Separation Index:: The separation index σ is the normalized distance between normal and anomalous score distributions, measured in normal-class standard deviations.Using σnormal rather than pooled variance reflects the separation relative to the normal score distribution and PatchCore’s thresholding basis.
2) Category Selection: Stratified by Separation Index:: A high separation index denotes an easy category with well-separated scores, whereas a low index denotes a hard category with overlapping distributions.Representative categories range from heavily overlapping toothbrush scores to fully separated leather scores.
2) Category Selection: Stratified by Separation Index:: Baseline AUROCs average 0.9342 and range from 0.8139 for toothbrush to 1.0000 for bottle and leather.Table II orders all 15 categories by separation index and reports their baseline AUROCs and dataset statistics.

3) Implementation Details and Bank Size Caps: … A. Full 15-Category Results Overview

The study uses a frozen WideResNet-50 PatchCore implementation with capped memory banks and evaluates correction through strictly held-out, seed-matched protocols. Across 15 MVTec AD categories, gains concentrate in four undersampled-normal-manifold categories, while several others remain flat or decline systematically.

3) Implementation Details and Bank Size Caps:: The frozen WideResNet-50 extracts stages 2–3 features from 224×224 ImageNet-normalised images, using k = 9 Euclidean nearest neighbours with a ball-tree index.The backbone uses timm ImageNet-pretrained weights and remains frozen throughout.
3) Implementation Details and Bank Size Caps:: Nmax = 12,000 patches bounds scoring cost without materially changing results; removing the cap altered only transistor (+0.021) significantly and doubled scoring cost.When exceeded, each bank is randomly subsampled back to Nmax; the index is rebuilt after every correction.
4) Held-Out Evaluation Protocol:: Held-out AUROC measures generalisation because only review-pool images may receive corrections, while the 30% held-out set is rescored but never corrected or used for querying or thresholding.Using corrected images for evaluation would drive AUROC toward 1.0 through memorisation rather than generalisation.
4) Held-Out Evaluation Protocol:: 20 independent split seeds support per-category testing with paired held-out deltas and two-sided Wilcoxon signed-rank tests, with Holm correction across categories.Reported results are means over seeds with standard deviations, and passive querying uses a different random review order per seed.
5) Cold-Start Protocol:: Cold-start experiments begin with k ∈{10, 20} golden training normals, retain matched 70/30 splits, and extend the correction budget to 90 rounds.Seed matching enables recovery fractions and median reviews to parity with the fully trained reference bank.
IV. EXPERIMENTAL RESULTS: The experimental results cover the full 15-category benchmark, then examine the novelty gate, cold-start performance, strategy equivalence, robustness, label economics, and FP/FN asymmetry.The full benchmark precedes analyses explaining which categories and deployment conditions drive the results.
A. Full 15-Category Results Overview: +0.146 AUROC was achieved by toothbrush, with metal nut at +0.098, zipper at +0.054, and screw at +0.040; these four categories gained on 15–20 of 20 seeds.The gains occur where false positives reflect an undersampled normal manifold; five near-ceiling categories were unaffected, while six were flat-to-negative, losing up to 0.047 AUROC.
A. Full 15-Category Results Overview: Adjusted p < 10−4 for toothbrush, metal nut, and zipper, and adjusted p = 0.032 for screw, while grid, pill, tile, and transistor were significantly negative.These Holm-corrected Wilcoxon results confirm systematic positive effects for all four winners and systematic degradation for four flat-to-negative categories.

B. Novelty-Gated Insertion Eliminates the Harm … 1) Toothbrush: Maximum Feedback Benefit:

Novelty-gated memory-bank correction removes nearly all category-level harm and delivers its greatest benefits in cold-start deployment, where ten golden samples recover much of full-training performance. Toothbrush exemplifies the mechanism, achieving a large held-out gain from only a few decisive corrections while avoiding memorisation-based evaluation.

B. Novelty-Gated Insertion Eliminates the Harm: Gating turns every flat-to-negative category except grid neutral or positive, while all four winners remain significantly positive.The ungated-to-gated changes are transistor −0.024 →+0.005, pill −0.023 →+0.007, cable −0.019 →+0.006, tile −0.013 → −0.001, capsule −0.009 →+0.003; winners are toothbrush +0.101, metal nut +0.089, zipper +0.051, screw +0.048.
B. Novelty-Gated Insertion Eliminates the Harm: The novelty gate self-calibrates by inserting fewer patches on well-covered banks and more on undersampled banks, although grid still degrades despite 230 inserted patches per correction.Tile averages 65 of 784 candidate patches per correction and transistor 78, versus 271 for metal nut and 236 for zipper.
C. Cold-Start Deployment: Near-Full Performance from Ten Samples: 80% mean recovery and 0.66 median recovery close the cold-start gap from only k = 10 golden samples, with significant held-out improvements in 12 of 15 categories and no significant harm.The initial banks contain just 78 patches; the median is computed over the 11 categories with a meaningful gap.
C. Cold-Start Deployment: Near-Full Performance from Ten Samples: The cold-start effect persists across protocol variations: k = 10 ungated achieves 0.89 mean recovery, while k = 20 achieves 0.85, and neither setting produces significant negatives.The corresponding significant-positive counts are 11 of 15 and 8 of 15; grid is neutral at +0.038 with k = 10 ungated.
1) Toothbrush: Maximum Feedback Benefit:: The held-out protocol is essential because corrected images can make pool AUROC reach 1.000 through memorisation rather than generalisation.The held-out set is never corrected, separating transfer to unseen data from direct bank memorisation.
1) Toothbrush: Maximum Feedback Benefit:: The toothbrush gain is genuine across operating points and concentrated in a few decisive corrections: one round contributes +0.139, while most rounds contribute nothing.Held-out corrected ROC curves dominate baseline curves at every operating point; one reviewed false positive adds 784 patches, of which 470 are inserted through the gate.
D. Key Category Analysis: Cold-start trajectories show metal nut crossing and exceeding its full-training baseline, zipper approaching parity, and screw remaining budget-limited.Metal nut crosses around round 50; zipper ends just below parity after an early dip; screw is still climbing at the 90-round limit.

2) Where Correction Fails: The Flat-to-Negative Group: … H. Label Economics: Is the Gain Just Extra Training Data?

Correction is systematically harmful only for grid among the flat-to-negative categories after novelty gating, while passive and active querying reach statistically indistinguishable outcomes. Multi-seed validation shows robust category-specific effects, and matched-budget results indicate that HITL correction approaches exhaustive labelling at substantially lower label cost.

2) Where Correction Fails: The Flat-to-Negative Group:: Grid remains significantly harmed after gating (−0.031), while the novelty gate resolves five of six flat-to-negative categories by rejecting redundant insertions.Ungated grid degradation is −0.047 on 15/20 splits; its bank inserts 23 redundant patches.
E. Strategy Comparison: Passive and active querying show no significant difference on any category, with mean improvements agreeing within ±0.01 AUROC across all 15 categories.Held-out trajectories overlap at every round on 14 of 15 categories; metal nut differs in speed, not endpoint.
E. Strategy Comparison: Correction utility concentrates in low-σ categories with false-positive-dominated error pools, while passive and active AUCC agree on every category except metal nut.The metal nut exception reflects convergence speed rather than a different final destination.
F. Ablation Study: kremove Parameter: The kremove ablation produced AUCC 0.9984 and final AUROC 1.0000 for every tested k ∈{1, 3, 5, 9}.The toothbrush session contained only false-positive corrections, so kremove never triggered; this was a single-seed, pre-held-out, pool-only evaluation.
G. Multi-Seed Validation: Grid degrades on 15/20 seeds, confirming that the flat-to-negative failure is systematic rather than sampling misfortune.On hard categories, single-seed outcomes can vary widely; toothbrush improvements range from −0.03 to +0.31 AUROC.
H. Label Economics: Is the Gain Just Extra Training Data?: The HITL loop reaches held-out AUROC 0.972 with 12.6 labels on average, versus 0.901 for randomly labelled images and 0.974 for appending all pool normals.The exhaustive pool-normal ceiling costs ∼29 labels, compared with the HITL loop’s 12.6 labels.

I. Ablation Study: FP vs. FN Correction … B. Negative Result: Defect Memory Bank

The ablation shows that correction works through inserting reviewed normal patches: FP correction reproduces the improvement, whereas FN removal is ineffective or destructive. Re-anchoring removal and memorising defects do not fix this asymmetry, while query strategy remains largely unhelpful.

I. Ablation Study: FP vs. FN Correction: FP-only correction matches the combined loop on all 20 seeds, improving held-out AUROC by +0.146 while requiring 7.5 versus 12.6 reviews.Its AUCC is 0.709 versus 0.563 for the combined loop.
I. Ablation Study: FP vs. FN Correction: FN-only correction leaves held-out AUROC at baseline on 13 of 20 seeds and has a slightly negative residual mean of −0.013.The result is based on 8.8 removals per session; unadjusted p = 0.034.
1) Why the Anchor Metric Is Not “Fixed”:: Re-anchoring FN removal on the score-driving mean-of-nine patch remains destructive, collapsing AUROC from 0.811 to 0.54–0.56 for every non-zero removal size.The collapse is independent of anchor metric and removal count.
1) Why the Anchor Metric Is Not “Fixed”:: Because a false negative’s nearest normal neighbour supports other normal-looking patches, removing it from a 1% coreset has no safe size.The toothbrush coreset contains 470 patches, making the failure a structural coreset problem rather than magnitude tuning.
V. DISCUSSION: Memory-bank corrections are global, so inserting false-positive patches changes nearest-neighbour distances across all images and makes ordering less consequential than in local active learning.The review budget exhausts the eligible error pool under any ordering.
A. Why Querying Strategy Barely Matters: Centroid-distance querying matched passive on toothbrush, underperformed on metal nut, and was severely counterproductive on cable because of strong FP-selection bias.These were single-seed pool-protocol experiments predating held-out evaluation; better-than-random query strategies remain open.
B. Negative Result: Defect Memory Bank: A defect memory bank collapses held-out AUROC by −0.226 on pill and −0.299 on screw, with screw worse than the standard loop on 20/20 seeds.Both Holm p < 10−4; rejected images over-generalise defect regions toward normal appearance, supporting correction by teaching normality.

C. Bank Growth, and Why the Cap Is Not the Problem … A. Limitations

The framework’s gains come from correcting undersampled normal appearance rather than enlarging the bank, while global defects, grid, evaluation leakage, and study-scale limitations remain important constraints. Deployment is practical without retraining, and corrections produce deployment-time labels at lower cost than exhaustive review.

C. Bank Growth, and Why the Cap Is Not the Problem: A 20-seed cap-removal control found no held-out delta changes except transistor’s (+0.021), refuting cap churn as the cause of failures.Toothbrush grows ∼14× from its 470-patch original bank because inserted patches cover new manifold.
D. What the Loop Cannot Fix: Cable’s cable_swap is globally defined and locally normal, while grid loses 0.031 AUROC despite inserting 230 patches per correction.Patch-level correction cannot encode global structure; grid is the method’s one open failure but is unharmed at cold start.
E. Practical Deployment Guidelines: The quality engineer reviews model outputs, contributes one binary label per image, and needs no ML knowledge, Python, training infrastructure, or ML team.Bank edits are CPU operations applied directly to the deployed model.
E. Practical Deployment Guidelines: +0.05 to +0.10 AUROC gated are the largest mature-bank gains, while cold-start gains reach +0.25; no category except grid is significantly harmed.The novelty gate should be enabled by default and performance should be confirmed with held-out checks rather than in-loop accuracy.
E. Practical Deployment Guidelines: Passive and active querying are statistically indistinguishable, while effective categories exhaust errors in 13–20 corrections.At one image per minute, this corresponds to roughly 13–20 minutes of review; monitoring a held-out sample is recommended.
VI. CONCLUSION: From ten golden samples, corrections close a median 66% of the gap to an uncorrected fully trained bank, significantly improve 12 of 15 categories, and harm none.The mean closure is 80%, and ten samples plus corrections outperform hundreds of samples without them.
VI. CONCLUSION: 43% of exhaustive-review labelling cost is sufficient because the framework produces correctly labelled deployment-time data without retraining.The contribution is deployment-time label production, not outperforming training data.
A. Limitations: Headline results average 20 pool/held-out splits per category, but some held-out sets contain only ∼13 images, making single-split AUROCs noisy.The evaluation uses a research implementation; production deployment would use approximate nearest-neighbour indexing as the bank grows.

B. Future Work • · C. Threats to Validity

Future work targets faster, richer, and broader correction loops, while validity is limited by simulated perfect feedback, cold-start sensitivity to operator error, single-dataset evaluation, and applicability mainly to memory-bank detectors. The paper therefore requires live expert validation before industrial claims and broader generalization.

B. Future Work •: Future extensions include batched correction for complex categories such as cable, pixel-level region marking, cross-category transfer, and integration with zero-shot detectors.These directions broaden correction beyond image-level labels and single-category memory banks.
B. Future Work •: Approximate nearest-neighbour indexing and incremental rescoring could reduce per-round latency to seconds, leaving approximately 1 minute per image of expert review as the bottleneck.Suggested implementations include FAISS and HNSW.
B. Future Work •: Live expert trials are the necessary next validation step because simulated feedback assumes always-correct experts and produces an upper bound.Real reviewers may disagree with ground truth, fatigue, and drift.
C. Threats to Validity: A single incorrect “normal” label can insert up to 784 defect patches, making imperfect feedback especially consequential for a 78-patch cold-start bank.Against a mature bank near the 12,000-patch cap, this represents only a few per cent contamination.
C. Threats to Validity: MVTec AD covers 15 diverse categories but omits factory conditions such as lighting variation, camera drift, and novel defect types, requiring proprietary-data revalidation.The findings should not yet be treated as general.
C. Threats to Validity: The correction mechanism applies to PatchCore and related coreset-based memory-bank detectors, not reconstruction-, flow-, or classifier-based detectors.The paper claims nothing beyond the memory-bank family.
C. Threats to Validity: 13–20 minutes is the reported correction-session duration, contrasting with formal retraining cycles that take weeks to months.This enables the person already reviewing errors to fix them immediately without an ML team.
DATA AND CODE AVAILABILITY: All experiments use the publicly available MVTec AD benchmark, and the implementation and reproduction scripts are available from the corresponding author on reasonable request.No proprietary or personal data were used.