🤖 machine learning

AnomalyMatch: Discovering Rare Objects of Interest with Semi-supervised and Active Learning

The paper introduces AnomalyMatch, a scalable semi-supervised and active learning framework that effectively discovers rare objects in large datasets with severe label scarcity by combining FixMatch-based training with user verification, achieving high precision and AUROC scores on astronomical and natural image benchmarks.

Original authors: Pablo Gómez, Laslo E. Ruhberg, Maria Teresa Nardone, David O'Ryan

Published 2026-06-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pablo Gómez, Laslo E. Ruhberg, Maria Teresa Nardone, David O'Ryan

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 librarian in a library with billions of books, but you are looking for just a handful of very specific, rare stories. The problem is, you only have a tiny list of what those rare stories look like (maybe just five or ten examples), and you don't have time to read every single book in the library one by one.

This is the exact challenge astronomers face today. New telescopes are taking pictures of billions of galaxies, but the "interesting" ones—like galaxies that are merging or have strange shapes—are incredibly rare. Finding them by hand is impossible.

This paper introduces AnomalyMatch, a smart computer tool designed to solve this problem. Here is how it works, explained simply:

1. The "Smart Intern" Approach (Semi-Supervised Learning)

Imagine you hire a smart intern (the AI model) to find these rare books.

The Problem: You can't give the intern a textbook with thousands of examples of the rare stories because they don't exist in large numbers yet.
The Solution: You give the intern a tiny stack of the rare books (5–10 examples) and a massive pile of regular books (millions of unlabelled images).
How it learns: The intern looks at the regular books and tries to guess which ones might be rare based on the tiny stack. If the intern is very confident about a guess, it treats that guess as a "practice example" and learns from it. This allows the intern to learn from the massive pile of regular books without needing a label for every single one.

2. The "Human-in-the-Loop" (Active Learning)

The intern isn't perfect. Sometimes they might think a regular book is rare, or miss a rare one. This is where the Active Learning part comes in.

The Process: The intern sorts the library and puts the "most likely rare" books at the top of a list.
The Check: A human expert (you) looks at the top of that list. You say, "Yes, that one is rare," or "No, that's just a smudge on the lens, ignore it."
The Feedback: You feed this correction back to the intern. The intern immediately updates its brain and re-ranks the library.
The Result: After just a few rounds of this "check and correct" cycle, the intern becomes incredibly good at finding the rare items, often finding 76% to 94% of them in the very top 1% of the list it generates.

3. The "Specialized Glasses" (EfficientNet)

To see the details in these astronomical images, the tool uses a pair of "specialized glasses" called EfficientNet. Think of this as a high-powered microscope that has already been trained on millions of everyday pictures (like cats, cars, and trees). Because it already knows how to see shapes and patterns, it only needs a little bit of extra training to spot the weird, rare galaxies.

4. What They Tested

The team tested this tool in three different "libraries":

MiniImageNet: A standard computer vision library with pictures of everyday objects (like guitars and pianos). They tried to find just one type of object (e.g., only "Hourglasses") among thousands of other things. The tool was very successful.
GalaxyMNIST: A library of galaxy pictures with four distinct shapes. They tried to find one specific shape among the others. Again, the tool worked very well.
Galaxy Zoo (The Real Test): They tested it on a real dataset of galaxies where humans had already voted on which ones looked "odd." They compared their tool to another famous tool called Astronomaly. AnomalyMatch performed just as well as Astronomaly, proving it can handle real-world, messy data.

5. Why This Matters

The paper highlights a few key takeaways:

Less Work for Humans: You don't need to label thousands of images. Starting with just 5 to 10 examples is enough to get great results.
Speed: The tool is fast enough to scan hundreds of millions of images on a single computer graphics card.
Scalability: It is built to be integrated into the European Space Agency's (ESA) data platforms, meaning it's ready to help astronomers handle the flood of data coming from future telescopes like Euclid and the Vera C. Rubin Observatory.

In short, AnomalyMatch is a tool that lets a computer learn to spot the "needle in the haystack" by learning from a few examples and asking a human for quick help when it gets confused, making the search for rare cosmic discoveries much faster and easier.

Technical Summary: AnomalyMatch

Problem Statement
The identification of anomalies—rare and unusual outliers—in large-scale image datasets is a critical challenge in astronomy and computer vision. While next-generation observatories like Euclid and the Vera C. Rubin Observatory will release unprecedented volumes of data (billions of galaxies), the scarcity of labelled examples for rare objects (often <0.1% of a dataset) renders traditional supervised machine learning methods infeasible. Existing approaches face significant hurdles: unsupervised methods often struggle with the complexity of high-dimensional image data, while supervised methods require large labelled datasets that do not exist for many rare morphologies. Furthermore, citizen science initiatives, while valuable, face challenges in defining consistent anomaly criteria and cannot scale to manually inspect the vast data volumes of modern surveys. The core problem is developing a scalable, efficient framework that can detect rare anomalies with minimal initial labelled data while leveraging the abundance of unlabelled data and incorporating expert human feedback to refine detection.

Methodology: AnomalyMatch
The authors present AnomalyMatch, a semi-supervised anomaly detection framework designed to treat anomaly detection as a heavily imbalanced binary classification problem ("normal" vs. "anomaly"). The methodology integrates three primary components:

Semi-Supervised Learning (FixMatch): The core training strategy adapts the FixMatch algorithm (Sohn et al. 2020). It utilizes a small set of labelled data ( $D_L$ ) and a large pool of unlabelled data ( $D_U$ ). The model, based on an EfficientNet backbone (specifically EfficientNetLite0), is trained using a combined loss function:
- Supervised Loss: Standard cross-entropy on the labelled set.
- Unsupervised Loss: Consistency regularisation applied to unlabelled data. Weakly augmented images are used to generate high-confidence pseudo-labels (if the model's confidence exceeds a threshold $\tau=0.95$ ). These pseudo-labels are then used to penalise the model's predictions on strongly augmented versions of the same images.
- Imbalance Handling: The framework employs oversampling via a weighted random sampler to address the severe class imbalance, ensuring rare anomaly samples are adequately represented during training.
Active Learning Loop: To mitigate the scarcity of labelled anomalies, AnomalyMatch incorporates an iterative human-in-the-loop process. After initial training, the model scores the unlabelled pool. A user interface (GUI) allows domain experts to:
- Visually inspect high-scoring candidates.
- Verify true anomalies and correct false positives (e.g., distinguishing real anomalies from imaging artifacts like bad CCDs or defocused regions).
- Iteratively expand the labelled training set with these verified samples.
- Retrain the model on the expanded set.
Scalable Implementation: The system is built for high performance, capable of processing hundreds of millions of images on a single GPU. It supports diverse input formats (PNG, JPEG, TIFF, FITS) and integrates with astronomical data storage standards (HDF5, Zarr) and the ESA Datalabs platform. The GUI, built with ipywidgets, facilitates interactive exploration within Jupyter notebooks.

Key Contributions

Novel Pipeline: A semi-supervised anomaly detection pipeline combining FixMatch (streamlined for heavy class imbalance), EfficientNet, and active learning, significantly reducing the requirement for labelled data.
Interactive GUI: A user-friendly interface enabling experts to iteratively guide anomaly selection, correct model errors, and refine performance without direct backend interaction.
Integration: Seamless integration of established astronomy tools (Astropy normalisation, FITS support) and deployment within the ESA Datalabs ecosystem.
Comprehensive Benchmarking: Extensive evaluation on diverse datasets (miniImageNet, GalaxyMNIST, and Galaxy Zoo 2) demonstrating performance with extremely limited initial labels (5–10 anomalies).
Scalability: An implementation capable of analysing 100 million images within days, suitable for upcoming large-scale surveys.

Experimental Results
The framework was evaluated on three datasets under severe class imbalance:

miniImageNet: Treating specific object classes (e.g., Hourglass, Guitar) as anomalies (1% of data). Starting with only 5 labelled anomalies, AnomalyMatch achieved an average AUROC of 0.96 and AUPRC of 0.82. After three active learning cycles, it achieved 76% precision in the top 1% of ranked images.
GalaxyMNIST: A dataset of galaxy morphologies where one class was treated as the anomaly (25% of data). Starting with 10 labelled anomalies, the system achieved an average AUROC of 0.89 and AUPRC of 0.77. Precision in the top 1% reached 94%.
Galaxy Zoo - The Galaxy Challenge: A subset of ~65,000 galaxies with "odd" classifications. Compared against the established Astronomaly software, AnomalyMatch achieved comparable performance with an average AUROC of 0.83. The authors note that performance is sensitive to the definition of "anomaly" (voter fraction cutoffs) in this soft-label dataset.

Ablation Studies

Initial Labels: Increasing the initial number of labelled samples (from 100 to 1000) yielded diminishing returns, with performance metrics remaining robust even with minimal initial labels.
Training Iterations: The optimal number of training iterations per active learning cycle was found to be around 100. Longer training cycles (250–500 iterations) led to overfitting on the initial labelled data, reducing overall detection efficiency.

Significance and Claims
The paper claims that AnomalyMatch offers exceptional utility and scalability for anomaly discovery in domains characterised by severe label scarcity, such as astronomy. By combining semi-supervised learning with active learning, the framework allows domain experts to define and detect scientifically interesting anomalies rather than just general statistical outliers. The authors emphasize that the method effectively leverages unlabelled data to reduce the manual review burden, making it a practical tool for the era of data-rich sky surveys. The results underscore that specialised approaches combining human expertise with efficient machine learning are key to advancing anomaly detection in fields where labelled data is inherently rare. The paper concludes that while the method is highly effective, future work should explore explainable AI, multimodal data integration, and the impact of labelling errors to further enhance robustness.

1. The "Smart Intern" Approach (Semi-Supervised Learning)

2. The "Human-in-the-Loop" (Active Learning)

3. The "Specialized Glasses" (EfficientNet)

4. What They Tested

5. Why This Matters

More like this