🤖 machine learning

TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection

The paper introduces TailedCore, a novel unsupervised anomaly detection framework that employs a TailSampler to independently handle tail class samples and noisy data, thereby overcoming the performance trade-off between noise robustness and tail class sensitivity in long-tail, contaminated datasets.

Original authors: Yoon Gyo Jung, Jaewoo Park, Jaeho Yoon, Kuan-Chuan Peng, Wonchul Kim, Andrew Beng Jin Teoh, Octavia Camps

Published 2026-05-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yoon Gyo Jung, Jaewoo Park, Jaeho Yoon, Kuan-Chuan Peng, Wonchul Kim, Andrew Beng Jin Teoh, Octavia Camps

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a quality control inspector at a massive factory. Your job is to spot defective products on a conveyor belt. This paper, TailedCore, tackles a very specific, messy real-world problem where your training data is a nightmare in two ways:

The "Long-Tail" Problem: You have thousands of examples of common items (like standard screws), but only a handful of examples for rare items (like a specific, custom-made gear). In data terms, the common items are the "head" of the distribution, and the rare ones are the "tail."
The "Noisy" Problem: Even the "good" examples you have are dirty. Some of the "normal" screws in your training pile actually have tiny scratches or defects on them.

The Dilemma: The "Too Strict vs. Too Lenient" Trap

The authors noticed that existing AI models get stuck in a terrible trade-off, which they call the "Tail-versus-Noise Dilemma."

The "Too Lenient" Model: If you teach a model to be very careful about rare items (the tail), it starts getting confused. It sees the tiny scratches on the "normal" items and thinks, "Oh, this is a defect!" It becomes too sensitive and flags normal things as broken.
The "Too Strict" Model: If you teach a model to ignore those tiny scratches (to handle the noise), it becomes too tough. It stops paying attention to the rare items. It looks at the custom gear and says, "I've never seen this before, so I'll just ignore it," missing the actual defects on the rare items.

It's like trying to learn a new language where you have a dictionary with 10,000 pages of common words but only one page of rare words, and that one page has some typos. If you focus on the typos, you miss the rare words. If you focus on the rare words, you get confused by the typos.

The Solution: TailedCore

The authors built a new system called TailedCore that solves this by treating the rare items and the noisy items as two separate problems. They do this using a clever two-step process involving a new tool they invented called TailSampler.

Step 1: The "Class Size Predictor" (TailSampler)

Imagine you are in a crowded room where people are grouped by the team they work for. Most teams have hundreds of people, but a few teams only have one or two people. You don't know who belongs to which team, and some people in the big teams are wearing "defect" badges (noise).

TailSampler is like a super-smart observer who looks at how people stand near each other.

It assumes that people from the same team stand at similar angles to each other.
It calculates how many people are likely in a specific "cluster" just by looking at the geometry of the crowd.
If it sees a tiny cluster (only 1 or 2 people), it knows, "Ah, this is a rare team!"
Crucially, it can tell the difference between a "rare team" and a "lonely person with a defect badge."

This allows the system to exclusively pick out the rare items without accidentally grabbing the noisy, defective ones from the common groups.

Step 2: The "Memory Bank" (TailedCore)

Once TailSampler has identified the rare items, TailedCore builds a special "memory bank" for the AI to use during inspection. This memory bank is a hybrid:

The Clean Part: It takes the common items, filters out the ones with scratches (noise), and stores the clean versions.
The Rare Part: It takes the rare items identified by TailSampler and adds them to the memory.

Now, when the AI inspects a new product:

If it's a common item, it compares it to the clean, filtered memory.
If it's a rare item, it has the specific examples of that rare item in its memory to compare against.

The Result

By separating the "rare" from the "noisy," TailedCore doesn't have to choose between being too strict or too lenient. It gets the best of both worlds.

In their experiments, they tested this on standard industrial datasets (like MVTec AD and VisA) that they artificially made "noisy" and "unbalanced." The results showed that TailedCore consistently outperformed the current state-of-the-art models. It was much better at spotting defects on rare products without getting confused by the dirty data on common products.

In short: TailedCore is a new way to teach an AI to spot factory defects when the training data is messy and unbalanced, by using a smart "head-counting" trick to isolate the rare items before the AI even starts learning.

Technical Summary: TailedCore

Problem Statement
The paper addresses a practical yet challenging scenario in unsupervised anomaly detection: training a single model on a multi-class dataset that suffers from two simultaneous complications. First, the dataset exhibits a long-tail distribution, where "head" classes (many-shot) are data-rich, but "tail" classes (few-shot) have very few samples (e.g., 1, 4, or 20). Second, the training data is noisy, meaning the "normal" training set is contaminated with defective samples (anomalies).

The authors identify a critical "tail-versus-noise trade-off" in existing methods. Models robust to noise (e.g., SoftPatch) tend to filter out rare samples, causing them to miss tail class information. Conversely, models effective at capturing tail classes (e.g., PatchCore via greedy sampling) often retain noise, leading to poor performance on head classes with contaminated training data. Existing literature fails to simultaneously handle class imbalance and pixel-level noise contamination without explicit class labels.

Methodology
The proposed solution, TailedCore, is a memory-based anomaly detection model designed to decouple the handling of tail classes from noise. The core innovation lies in a two-stage memory construction process:

Noise Discrimination: The model first constructs a base memory bank ( $M_{clean}$ ) by filtering out defective patches from the training set using outlier scoring (based on Local Outlier Factor), similar to SoftPatch. This ensures the base memory is free from noise but may lack tail class representation.
TailSampler (Few-Shot Sampling): To recover the missing tail class information, the authors introduce TailSampler, a novel few-shot sampler.
- Class Size Prediction: TailSampler estimates the "class size" (cardinality) of every training sample without ground-truth labels. It operates on the hypothesis that in the embedding space, the distribution of inter-class and intra-class similarities exhibits reflective symmetry.
- Adaptive Neighborhoods: For each sample embedding, an adaptive angle is calculated to define a neighborhood. The class size is estimated by taking the mode of the sizes of the neighborhoods of the neighbors (a "neighborhood of neighborhoods" approach), rather than a simple count. This provides a robust estimate of whether a sample belongs to a small (tail) or large (head) class.
- Elbow Detection: The system determines the maximum size threshold ( $K_{max}$ ) for a "tail class" by applying the elbow method to the distribution of estimated class sizes.
- Exclusive Sampling: Samples with estimated class sizes $\le K_{max}$ are identified as tail class instances. Their patch features are sampled exclusively to form a tail memory bank ( $M_{tail}$ ).
Memory Augmentation: The final memory bank for TailedCore is the union of the noise-clean base memory and the tail-specific memory: $M_{TailedCore} = M_{clean} \cup M_{tail}$ . This allows the model to be robust against noise while retaining the critical features of few-shot classes.

Key Contributions

Problem Formulation: The paper defines and benchmarks the specific setting of unsupervised long-tail noisy anomaly detection, a scenario common in industrial applications where production lines have varying rates and defective training data.
TailSampler: A novel class size predictor and few-shot sampler that relies on the reflective symmetry of embedding similarities to aggregate few-shot class instances exclusively, without requiring class labels.
TailedCore Model: A memory-based detector that successfully navigates the tail-versus-noise trade-off by independently handling tail class augmentation and noise discrimination.
Comprehensive Benchmarking: The authors provide extensive experiments on modified MVTecAD and VisA datasets with Pareto and step-shaped tail distributions, demonstrating that TailedCore outperforms state-of-the-art methods (including PatchCore, SoftPatch, HVQ, and foundation-model-based approaches like WinCLIP and AnomalyCLIP) in both image-level classification and pixel-level segmentation.

Results
Experimental results indicate that TailedCore consistently outperforms existing baselines across various tail distributions (Pareto, step-K=4, step-K=1) and noise ratios.

Performance: TailedCore achieves superior image-level and pixel-level AUROC scores, particularly in the "all" category which balances head and tail performance. For instance, on MVTecAD with step-K=1, TailedCore achieves an image-level AUROC of 94.44%, significantly outperforming PatchCore (80.01%) and SoftPatch (75.40%).
Ablation Studies: The study confirms that the class size predictor (TailSampler) is critical; removing the few-shot augmentation causes a drastic drop in tail class performance. The method also shows that the "majority vote" mechanism in class size estimation is more robust than simple averaging or top-1 selection.
Robustness: The model maintains high performance even as the noise ratio in the training set increases, whereas other methods degrade significantly.

Significance and Claims
The paper claims that TailedCore represents a significant stride in handling real-world anomaly detection scenarios where data is both imbalanced and contaminated. By resolving the inherent trade-off between noise robustness and tail class sensitivity, the work provides a robust foundation for industrial applications where companies must manage diverse product lines with varying production rates and imperfect data collection. The authors emphasize that their approach requires no supervised labels or complex prompt engineering, making it a practical solution for unsupervised settings. The code and data generation protocols are made available to facilitate further research in this domain.