Weakly Supervised Learning for Anomaly Detection in the Gamma-Ray Sky ✛ ✦
Context
The Fermi-LAT catalog of active galactic nuclei and pulsars contains thousands of sources, many of which remain unidentified. Traditional machine learning classification methods require detailed signal modeling, reducing sensitivity when an object's true nature deviates from training templates. Weakly supervised learning, in contrast, uses only class proportion information within mixed samples, enabling anomaly detection without explicit signal assumptions. This approach is particularly valuable in particle astrophysics, where the nature of dark matter or hypothetical particles like axions remains unknown.
Methods
The study employs a background-versus-mixture variant of weak supervision. The background sample consists of well-identified astrophysical sources, such as pulsars or active galactic nuclei, while the mixed sample contains both background objects and potentially anomalous spectra in unknown proportions. A classifier (gradient-boosted decision tree ensembles) is trained to distinguish these two samples, effectively extracting signal features without individual labels. Input features are low-level: fluxes in seven (for the pulsar-AGN and dark matter tasks) or 50 (for axion oscillations) energy bins, obtained through Fermi's spectroscopy and photometry. To boost background statistics, synthetic spectra generated by normalizing flows—a generative model trained on real data—were used. Performance was compared against a fully supervised benchmark.
Results
For the pulsar versus active galactic nuclei separation task, the weakly supervised classifier achieved a true positive rate (TPR) of 0.66–0.82 depending on signal fraction, approaching the supervised benchmark (TPR 0.82) at higher fractions. In the search for dark matter subhalos with mass 80 GeV in the b-quark annihilation channel, with a signal fraction of 10% and an imbalanced background-to-mixture ratio of 2:1, the TPR was about 0.4 at a false positive rate (FPR) of ~0.07, while the supervised method gave TPR 0.83 at FPR 0.02. For the axion-photon oscillation scenario with strong coupling (gaγ = 500×10⁻¹¹ GeV⁻¹) and an effective signal fraction of 3%, weak supervision yielded TPR 0.49 and FPR 0.08, whereas the supervised method at 30% signal achieved TPR 0.74. In all cases, accuracy remained high (0.90–0.97) due to background class dominance. A key finding: as the signal fraction in the mixed sample increases, TPR rises but FPR may also climb, reflecting a trade-off inherent to the background-versus-mixture scheme.
Implications
Weakly supervised learning does not replace full model-dependent searches but serves as a flexible tool for preliminary candidate selection. It reduces reliance on precise knowledge of the signal spectrum, which is especially crucial when searching for unexpected manifestations of dark matter or effects beyond the Standard Model. The method allows ranking sources by degree of anomalousness, flagging subsets for subsequent detailed analysis, including multi-wavelength observations and hypothesis testing. This complements traditional likelihood-based approaches and expands the toolkit for indirect searches of new physics in astrophysical data.
Future development
Direct application of the method to real Fermi-LAT data is planned, with careful handling of selection effects and validation of the background distribution assumption. Methodologically, exploring deeper architectures such as attention-based neural networks and integrating with simulation-based inference for quantitative parameter constraints after candidate detection is promising. Extensions to data from other observatories and to searches for more subtle spectral features are also envisioned.
Impact
The method will impact gamma-ray astronomy, particle physics, and interdisciplinary research at the intersection of machine learning and astrophysics. It may also prove useful in other fields requiring the identification of rare anomalies in large datasets without a clear signal model.
Next steps
Application to unidentified sources in the Fermi-LAT catalog, with systematic control and validation of the background assumption. Development of methods combining weak supervision with Bayesian approaches to assess candidate credibility.
Key open problems
The work touches on several open problems in modern physics: the nature of dark matter, the existence of axions and axion-like particles, and the identification of populations of unidentified gamma-ray sources that could point to new astrophysical classes or exotic processes.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The study applies weakly supervised learning (background-mixed scheme) to Fermi-LAT data to search for anomalous gamma-ray sources. The classifier is trained to distinguish a background sample from a mixed one using only class proportions, and is tested on tasks of separating pulsars and active galactic nuclei, searching for dark matter subhalos, and axion oscillations. The results are compared with a fully supervised benchmark, showing moderate effectiveness at a high false-positive rate.
Where its strength is
The method does not require detailed signal modeling and can detect anomalies without assumptions about the nature of the source, which is valuable for finding unknown objects.
Where it could go next
- It might be interesting to look at the experience of using neural networks to search for anomalous quasars — there the idea of extracting rare objects from a large background was also used.based on: The Puzzle That Doesn't Fit: Neural Network Searches for Anomalous Quasars arXiv:2607.17728v1
- Nearby is a work on applying machine learning to analyze spectra in search of invisible gas; perhaps their approach to extracting weak signals against noise will be useful for improving sensitivity.based on: Seeing the Invisible Gas: Lasers and Machine Learning arXiv:2605.01306
- In the archive there is a work where a neural network searches for anomalies in the distribution of quasars associated with gravitational lenses; perhaps their method of detecting rare patterns is also applicable to gamma-ray sources.based on: Cosmic Masquerade: How a Neural Network Unmasks Gravitational Lenses arXiv:2607.03885
Worth trying
- Test the use of more complex architectures (e.g., neural networks) instead of gradient boosting to improve separation at small signal fractions.
- Check whether adding synthetic data generated by normalizing flows helps in other search scenarios (e.g., for axion-like particles with different masses).
- Explore the possibility of combining weak supervision with false-positive reduction methods, such as probability calibration.
Why this may matter
Such an approach could expand the possibilities of searching for new physics in astrophysical telescope data, especially when theoretical models are uncertain.
Nearby works in our archive
- The Puzzle That Doesn't Fit: Neural Network Searches for Anomalous Quasarssimilarity 0.67
- Seeing the Invisible Gas: Lasers and Machine Learningsimilarity 0.64
- JADE: The AI Detective That Reads the Cosmic Web in Light Trailssimilarity 0.63
- Dark Matter Sings in Two Voicessimilarity 0.63
- Cosmic Masquerade: How a Neural Network Unmasks Gravitational Lensessimilarity 0.62
- How Headphones Learned to See the Source of Noisesimilarity 0.62
The closest work in the archive sits at 0.67 by meaning — the usual level for our corpus.
Original abstract
as written by the authors arXiv:2607.07158
The $$\gamma$$-ray sky, as observed by the Fermi Large Area Telescope, contains a significant number of unassociated sources that may point to new astrophysical populations or more exotic phenomena. Machine-learning methods are widely used for source classification and searches for new physics, but most existing approaches rely on fully supervised training and therefore on explicit signal models. We explore weakly supervised classification as a less model-dependent strategy for analysing $$\gamma$$-ray source spectra. In a background-versus-mixture setup, classifiers are trained on samples with different signal admixtures rather than on fully labelled signal and background events. We study three representative scenarios: pulsar-active galactic nuclei separation as a controlled benchmark, the identification of dark-matter subhalos, and spectral irregularities induced by axion-photon oscillations. In each case we investigate the impact of signal fraction and sample composition on classification performance. Our results show that weak supervision can identify anomalous or signal-like subsets of data while reducing the reliance on detailed signal templates during training. In favourable cases, the method approaches the performance of fully supervised classifiers, while remaining applicable in situations where the signal model is uncertain or only partially specified. Weakly supervised learning therefore provides a complementary candidate-selection and anomaly-ranking strategy for $$\gamma$$-ray data analysis and searches for new phenomena.
The Fermi-LAT catalog contains 2,430 unidentified sources, more than any known class, offering a rich field for discovery.
The theme resonates with Blake Crouch's novel *Dark Matter*, where characters confront unpredictable manifestations of a hidden reality, much like scientists searching for signals from an invisible substance in gamma-rays.
Key numbers
- Number of active galactic nuclei in the 4FGL catalog: 4006
- Number of pulsars: 320
- Typical signal fraction in weakly supervised learning: 10%
- Dark matter mass benchmark: 80 GeV
- Axion-photon coupling constant gaγ for strong regime: 50×10⁻¹¹ GeV⁻¹