A kaleidoscope of noise — that’s one way to describe the data stream pouring out of the LIGO interferometers. Built by Rainer Weiss, Kip Thorne, and Barry Barish, these detectors capture ripples in spacetime from merging black holes, but at the same time they register everything: the chatter of ground vibrations, the hum of electrical interference, the rumble of a passing truck. Each such disturbance is a glitch — a brief noise burst, sometimes indistinguishable from the sought-after gravitational-wave signal. As sensitivity increased in the O4a run, the kaleidoscope spun faster: patterns grew more varied, and the usual classifiers, trained on old shapes, stopped recognizing new ones.

Enter Machine Learning, specifically the DANTE V3 algorithm, which works like a multi-focus lens. Instead of a single fixed window, it scans the data on four timescales — from 0.5 to 4 seconds. On each scale, a spectrogram is built — a time–frequency portrait of the signal. A specially trained transformer (DINOv2) breaks this portrait into patches and, using a Multiple Instance Learning mechanism (focusing on the most anomalous fragments), picks out the most suspicious ones while suppressing the background. But the main trap is noise manifold drift: detectors constantly change their “habits,” so what was normal yesterday looks anomalous today. To avoid mistaking the kaleidoscope’s own evolution for unusual patterns, DANTE deploys a two-layer defense. First, Domain Shift Defence: each candidate is compared to an “atlas” of 1216 typical noise patterns collected during quiet periods. Then, physical vetoes: if a burst in the gravitational-wave channel coincides with fluctuations in seismometers or magnetic sensors (PEM channels), it’s flagged as environmental.

From 84 runs of the two detectors (H1 and L1), 10 372 unique candidates were extracted. After filtering through the domain shift defense, the landscape changed dramatically: only 28.3% of events remained; the rest turned out to be products of the drifting background. Cluster analysis found no compact families — all survivors merged into a fuzzy cloud, indistinguishable from random fluctuations. Two events stood apart. The first (H1, GPS 1369305276.0) was firmly rejected: its coherence with the control system channel reached 0.987. The second outlier (L1, GPS 1382955228.0) passed all tests: none of the seven PEM channels showed a significant connection, and the partner detector H1 saw no counterpart. Yet the lack of a gravitational-wave template prevents us from attributing an astrophysical origin — it is cautiously classified as an uncatalogued instrumental transient. The upper limit on the rate of such never-before-seen glitches came to 5.83 events per year for H1 and 5.63 for L1.

The DANTE V3 story makes it clear: without grounding in real noise context, unsupervised algorithms tend to multiply entities. A protocol where every stage is calibrated against empirical data, not analytical assumptions, becomes a reliable filter for future gravitational-wave observations. The next step is a causal version operating in real time: a sliding background index will allow the pipeline to be deployed in live monitoring of O4b and subsequent runs. Thus the kaleidoscope of noise ceases to be chaos and turns into a map on which genuine spacetime ripples emerge ever more sharply.