DANTE: A Reference-Guided Unsupervised Pipeline for Extended-Transient Anomaly Characterization in LIGO O4a
Luca Cirfeta · 2026
Read DANTE: A Reference-Guided Unsupervised Pipeline for Extended-Transient Anomaly Characterization in LIGO O4a by Luca Cirfeta (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-25702.txt.
Introduction
Gravitational wave detectors hear the universe and their own creaks: glitches, brief instrumental transients that mimic cosmic signals. Luca Cirfeta presents DANTE in 2026, an unsupervised pipeline for the LIGO O4a run that hunts extended transients with no labels at all. A pretrained vision transformer reads time frequency spectrograms patch by patch, giving fine maps of odd structure. Controlled injections expose a signal dilution barrier: pooling methods that recover long, spread out shapes go blind to anything under a second.
An adaptive mixture model that picks its own covariance shapes steadies the clustering of glitch families, which otherwise shifts with small samples. Recalibrating the background on O4a data itself resolved a domain shift puzzle: morphologies first flagged as novel turned out to be stationary artefacts of the new run. Remaining lone anomalies, Cirfeta concludes, still need multi channel checks before anyone calls them real.
Catalog record
| CREATOR | Luca Cirfeta |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 115 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.25702
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DANTE: A Reference-Guided Unsupervised Pipeline for Extended-Transient Anomaly Characterization in LIGO O4a Luca Cirfeta Abstract The analysis of gravitational-wave detector data during the fourth observing run (O4) requires robust methods to distinguish stationary instrumental noise from non-stationary transients (glitches). In this work, we present DANTE (Domain-Adaptive Network for Transient Evaluation), a pipeline designed to discover and triage novel non-stationary artifacts entirely without labels. We demonstrate that adapting a pre-trained Vision Transformer (DINOv2) to extract local patch embeddings from time-frequency spectrograms allows for high-resolution mapping of transient anomalies. We formalize the Signal Dilution Barrier via controlled injection tests, showing that while Multiple Instance Learning (MIL) Top- k k pooling recovers extended topologies, it is blind to sub-second morphologies. To address small-sample taxonomy instability, we introduce an adaptive Dirichlet Process Mixture Model (DPMM) that dynamically selects covariance structures. Finally, by implementing a native O4a background recalibration, we resolve the domain-shift problem, demonstrating consistency with the hypothesis that pervasive O4a morphologies (initially flagged as novel by historical references) are stationary artifacts. We conclude that unsupervised anomaly detection strictly requires native recalibration to filter domain-shift artifacts, while definitive classification of remaining unmodeled singletons requires multi-channel validation. I Introduction The international network of ground-based interferometric gravitational-wave (GW) detectors—Advanced LIGO Abbott et al. 2015 , Advanced Virgo Acernese et al. 2015 , and KAGRA Akutsu et al. 2021 —has entered an era of continuous, high-cadence astrophysical discoveries during the fourth observing run (O4). The detection of cosmic strains at the level of h\sim 10^{-23} , corresponding to spacetime distortions smaller than one-thousandth of a proton diameter, demands extraordinary instrumental sensitivity. This sensitivity, however, renders the detectors acutely susceptible to non-Gaussian, transient noise artifacts—commonly referred to as glitches —that originate from a broad spectrum of environmental, mechanical, and electronic sources Nuttall 2018 ; Davis et al. 2021 . Glitches overlap in time-frequency morphology with genuine astrophysical signals, including compact binary coalescences (CBCs), unmodeled burst sources, and continuous wave candidates, directly degrading the performance of matched-filter search pipelines and Bayesian parameter estimation frameworks Powell et al. 2015 ; Pankow et al. 2018 . I.1 The Detector Characterization Challenge for O4 Transient noise characterization— DetChar —is a cornerstone of data quality assurance for gravitational-wave observatories. The O4 run introduced significant instrumental upgrades relative to O3, including squeezed-light injection and increased laser power levels, which have historically generated new and previously undocumented glitch populations Tse et al. 2019 ; Capote et al. 2021 . The canonical supervised approach, exemplified by the Gravity Spy framework Zevin et al. 2017 ; Glanzer et al. 2023 , employs convolutional neural networks trained on curated labeled datasets from earlier observing runs. While this approach achieves high classification accuracy on known glitch types, it faces challenges when instrumental upgrades introduce previously undocumented glitch populations. Reference-guided novelty detection—classifying samples as anomalous without prior labels—offers a complementary approach. Our primary goal is to establish an unsupervised methodological framework on O4a data to identify unmodeled transient morphologies while controlling for the domain shift arising from instrumental upgrades between O3b and O4a. DANTE is designed to provide candidate topological anomalies to downstream offline DetChar protocols for formal multi-channel physical classification, operating alongside human-in-the-loop and supervised vetting. I.2 Limitations of Standard Vision Transformer Approaches Contemporary unsupervised GW anomaly detection has increasingly leveraged self-supervised representation learning, particularly Vision Transformers (ViTs) Dosovitskiy et al. 2021 pre-trained on large-scale natural image corpora. The DINOv2 framework Oquab et al. 2024 ; Darcet et al. 2024 is particularly attractive: its training via distillation with no labels produces patch-level feature representations with strong emergent semantic properties and spatial locality, which transfer non-trivially to the time-frequency geometry of Q-transform spectrograms. However, standard ViT deployments introduce a methodological obstacle: the signal dilution effect . Because the global [CLS] token aggregates information from all 1,369 spatial patches via an implicit attention-pooling mechanism, the localized anomaly signal is heavily diluted by the overwhelming majority of stationary background patches. This architectural constraint severely limits the sensitivity to sub-second morphologies (e.g., Blips). We formally characterize this barrier and its mathematical bounds via the Mock Data Challenge detailed in Section V.1 . I.3 The Curse of Dimensionality in Small-Sample Anomaly Taxonomy A second, orthogonal challenge arises in the taxonomy stage, where detected anomalies must be grouped into morphologically coherent families. The principal mathematical challenge is that fitting an unrestricted Gaussian Mixture Model in \mathbb{R}^{384} with n<100 samples is algebraically ill-posed. The condition number of the sample covariance matrix \hat{\boldsymbol{\Sigma}}\in\mathbb{R}^{384\times 384} scales as \mathcal{O}(d/n) , inducing numerical instability: full covariance matrix estimation requires d(d+1)/2=73{,}920 parameters for d=384 , far exceeding the sample size. This produces rank-deficient covariance matrices, singular factorization errors, and severe overfitting of the mixture boundaries Bishop 2006 . Non-linear dimensionality reduction approaches such as UMAP McInnes et al. 2018 can project embeddings into a low-dimensional manifold. However, to preserve global density fields required for DPMM stability, we favor a linear projection approach. I.4 Contributions of This Work In this paper, we present DANTE ( D omain- A daptive N etwork for T ransient E valuation) Pipeline V2, a robust processing pipeline that addresses both limitations simultaneously. Our principal contributions are: 1. Mitigating the Signal Dilution Effect via Patch-Level MIL: We provide a baseline demonstration of the patch-level Multiple Instance Learning (MIL) mechanism to mitigate the signal dilution effect common in global-pooling ViT architectures. By focusing on the Top- k most anomalous patches, the pipeline biases detection toward extended mid-band transients (e.g., >1 second). We explicitly delegate the implementation of an adaptive multi-scale k -sweep to future iterations (V3). 2. Adaptive Taxonomy under Small-Sample Constraints: We formulate an adaptive PCA dimensionality floor combined with a conditional DPMM covariance framework, dynamically selecting full , tied , or diag covariance structure as a deterministic function of sample size n , eliminating matrix singularity and density distortion without requiring any non-linear manifold embedding. 3. Cross-Session Connectivity: We introduce a hierarchical cosine similarity clustering framework operating on archived 384-dimensional MIL feature vectors, enabling automated linkage of morphologically related transients across independent sessions without requiring human-in-the-loop validation. 4. Production Infrastructure for O4a Continuous Streams: We describe the complete engineering architecture—state-aware resumption, SWMR-enabled HDF5 archival, batched GPU inference, and structured provenance logging—enabling uninterrupted deployment on raw LIGO strain data acquired via the Gravitational Wave Open Science Center (GWOSC) GWOSC 2023 . It is critical to note that DANTE is explicitly designed as an extended-duration transient detector (targeting anomalies >1 second). As demonstrated in our Mock Data Challenge, sub-second glitches (e.g., Blips) suffer from extreme signal dilution in the 32-second global spatial pooling mechanism, yielding a near-zero recall. Consequently, DANTE does not replace supervised, short-window classifiers like Gravity Spy, but rather operates as a complementary offline archival system for uncovering extended, unmodeled morphological families. The remainder of this paper is organized as follows. Section II reviews related work in ViT-based GW anomaly detection and unsupervised clustering methods. Section III presents the full two-layer architecture of DANTE Pipeline V2 in mathematical detail. Section IV describes the experimental setup, data provenance, and engineering infrastructure. Section V reports detection and taxonomy results on LIGO O4a data. Section VI discusses the astrophysical implications, operational limits, and roadmap. Section VII concludes. II Related Work II.1 Supervised Glitch Classification The Gravity Spy framework Zevin et