Patch-Level DINOv2 Scoring for Gravitational-Wave Glitch Detection: Breaking the Signal Dilution Barrier via Vector-Quantized Local Feature Indexing

This paper introduces a patch-level scoring architecture using frozen DINOv2 and vector-quantized local feature indexing to overcome the signal dilution limitations of global CLS token metrics, thereby enabling unsupervised detection and topological localization of diverse gravitational-wave glitches in LIGO O4a data.

The Gist

The Big Problem: The "Needle in a Haystack" Effect

Imagine you are looking at a giant, 37-by-37 grid of tiles (1,369 tiles total) that represents a snapshot of sound from a gravitational wave detector. Most of the tiles are just "static" or background noise.

Sometimes, a real signal (a "glitch" or a gravitational wave) appears, but it only covers a tiny few tiles—maybe just 5 or 10 of them.

The Old Way (The "Global Average" Mistake):
Previously, the computer tried to understand the whole image by taking the "average" of all 1,369 tiles and squishing them into a single summary number (called a [CLS] token).

• The Analogy: Imagine you have a bucket of water. You drop a single drop of red dye into it. If you take a sample from the bucket and mix it, the water looks barely pink. The red dye is so diluted by all the clear water that you can't tell it's there.
• The Result: Because the signal was so small compared to the background noise, the computer's "average" completely ignored the glitch. It was mathematically blind to anything smaller than 5% of the image.

The New Solution: The "Top-K" Detective

The authors, led by Luca Cirfeta, realized they needed to stop looking at the "average" and start looking at the specific, weird tiles.

1. Zooming In (Patch-Level Scoring):
Instead of squishing the whole image into one number, they kept all 1,369 individual tiles separate. They treated each tile as its own little clue.

2. The "Dictionary of Normal" (Vector-Quantized Index):
To know what a "glitch" looks like, the computer needs to know what "normal" looks like. The authors built a massive dictionary (a reference index) containing 1,216 examples of what normal noise looks like, broken down by different shapes and patterns.

• The Analogy: Imagine a librarian who has memorized the exact texture of every normal page in a library. If you hand them a page, they can instantly compare it to their mental dictionary.

3. The "Top-K" Strategy:
When a new image comes in, the computer compares every single tile against its dictionary. It asks: "Which tiles look the most different from normal?"

• Instead of averaging everything, it picks the top 68 most suspicious tiles (this number, $k=68$, was found to be the sweet spot for the specific signals they were hunting).
• It calculates a score based only on those top 68 weird tiles, ignoring the 1,300+ normal ones.
• The Analogy: Instead of asking, "Is the whole room noisy?" (which might be "no" because most of the room is quiet), the detective asks, "Are there any specific people in this room shouting?" If even one person is shouting, the answer is "Yes, there is an anomaly."

What They Found

The team tested this new method on real data from the LIGO detector (specifically from May 2026).

• The "Spiral" Signal: For signals that spread out over a medium area (like a "SpiralBurst"), the new method worked perfectly. It could clearly separate the signal from the noise, whereas the old method saw nothing.
• The "Blip" Signal: For extremely tiny, split-second signals (like an "AsymBlip"), the new method still couldn't see them.
  • Why? The signal was so small it didn't even fill up a single tile on the grid. It was like trying to see a single grain of sand through a telescope that only has a resolution of a beach ball. The paper calls this the "Spatial Diffraction Limit."
• The "Heat Map" (Saliency Map): The authors also created a visual map that highlights exactly where the weird tiles are.
  • Important Note: The paper warns that this map is for visualization only, not for making final decisions. Sometimes, random noise can look like a "hot spot" just by chance. The map helps humans see where to look, but the computer's "Top-68 score" is what actually decides if a signal is real.

The Bottom Line

The paper claims to have solved a specific mathematical problem where computer vision models were "diluting" small signals by averaging them with background noise. By switching from a "global average" approach to a "find the top weird tiles" approach, they successfully detected signals that were previously invisible to the system.

However, they admit this isn't a magic bullet for everything: if a signal is smaller than the grid's smallest tile, it still cannot be seen. The goal now is to use this new "Top-K" scoring to help computers find new, unknown types of glitches in future data.

Technical Summary

Technical Summary: Patch-Level DINOv2 Scoring for Gravitational-Wave Glitch Detection

1. Problem Statement: The Signal Dilution Barrier

The characterization of non-Gaussian transient noise ("glitches") in gravitational-wave interferometers is essential for maximizing the astrophysical reach of the Advanced LIGO and Virgo network. While supervised frameworks like Gravity Spy excel at classifying known morphologies, they lack the ability to detect novel anomaly populations. Previous unsupervised approaches utilizing Vision Transformers (ViT), specifically DINOv2, faced a critical structural limitation identified in prior work (Cirrfa 2026b): the Signal Dilution Effect.

Standard DINOv2 architectures process spectrograms by dividing them into a $37 \times 37$ grid (1,369 patches) and aggregating these into a single global [CLS] token via average pooling. For short-duration transients (e.g., AsymBlip or SpiralBurst) that occupy less than 5% of the spectrogram grid, the anomaly signal is mathematically diluted by the background noise covering the remaining 95% of the grid. Consequently, the global similarity metric fails to distinguish these events from noise, resulting in a Boolean Recall of 0.00 even at high signal-to-noise ratios (SNR > 400).

2. Methodology: Patch-Level Vector Quantization and Top-k Scoring

To overcome the signal dilution barrier, the authors propose an architectural shift from global token aggregation to dense, patch-level analysis. The methodology consists of three core components:

2.1. Patch-Level Feature Extraction

Instead of relying on the global [CLS] token, the model extracts the 1,369 individual patch tokens ($P_i \in \mathbb{R}^{384}$) directly from the final transformer block. These tokens undergo strict L2-normalization to ensure they reside on the unit hypersphere, facilitating cosine similarity calculations.

2.2. Vector-Quantized (VQ) Reference Index

To manage the computational intractability of searching 1,369 high-dimensional vectors against a massive dataset, the authors employ Spherical Vector Quantization.

• Construction: Using 19 known morphological classes from the Gravity Spy O3b dataset, patch tokens are clustered using MiniBatchKMeans ($K=64$ centroids per class).
• Result: This creates a compact, spatially invariant dictionary of 1,216 prototypical centroids ($19 \times 64$) representing the known structural space. This index ensures perfect reproducibility across hardware iterations.

2.3. Top-k Order Statistics Scoring

The core innovation is the replacement of global averaging with a Top-k Novelty Scoring mechanism.

1. Local Anomaly Calculation: For each patch in an incoming spectrogram, the algorithm computes the anomaly score ($a_i$) as the inverse of the maximum cosine similarity against the VQ dictionary.
2. Top-k Aggregation: The anomaly scores are sorted in descending order. The global novelty score is defined as the mean of the top-$k$ values:
   $$\text{Novelty} = \frac{1}{k} \sum_{j=1}^{k} a_{(j)}$$
3. Optimization: An empirical sweep determined $k=68$ as the optimal statistic for SpiralBurst morphologies, which occupy approximately 5% of the grid (~74 patches). This prevents the re-introduction of signal dilution by excluding the majority of background patches from the score.

2.4. Topological Saliency Maps

To address spatial localization without the artifacts introduced by the VQ index (which loses positional information), the authors decouple the visualization tool from the detector. A Topological Saliency Map is generated by comparing patch tokens coordinate-by-coordinate against a "Background Median Matrix" derived from 78 null noise segments. This provides a non-discriminative visualizer for post-hoc interpretation.

3. Key Contributions

1. Architectural Resolution: The first demonstration of a patch-level scoring architecture that successfully mitigates the Signal Dilution Effect in gravitational-wave time-frequency data.
2. Vector-Quantized Indexing: A scalable method for compressing high-dimensional patch manifolds into a reproducible reference index ($K=64$ per class) suitable for streaming applications.
3. Top-k Scoring Algorithm: A novel scoring mechanism that isolates the most anomalous structural components, mathematically mapping the detection statistic to the physical topological area of the anomaly.
4. Micro-MDC on Real Data: The first patch-level Mock Data Challenge (MDC) performed on real LIGO O4a strain data (session 20260524), demonstrating statistically significant separation where global approaches failed completely.

4. Experimental Results

The authors conducted a Micro-MDC injecting three morphologies (AsymBlip, SpiralBurst, HarmonicComb) into LIGO O4a L1 data.

• SpiralBurst (Mid-Band): The patch-level approach achieved a Kolmogorov-Smirnov (KS) statistic of 0.963 at the optimal $k=68$, indicating a statistically significant separation ($p < 0.01$) between glitch and noise distributions. This contrasts with the global [CLS] approach, which yielded a Recall of 0.00.
• HarmonicComb (Broadband): The method achieved extreme separability (KS > 0.97) across the entire $k$-sweep, recovering signals that were previously undetectable by global pooling.
• AsymBlip (Ultra-Short): The study confirmed a spatial diffraction limit. For transients occupying only ~15 patches (significantly smaller than the ViT patch size), the KS statistic remained non-significant ($p > 0.5$) regardless of $k$. This confirms that signals smaller than the patch footprint remain mathematically unresolved by this architecture.
• Saliency Validation: The Topological Saliency Map correctly localized Scattered Light and injected SpiralBurst signatures. However, analysis of the Max/Mean ratio revealed that background noise can produce localized similarity spikes comparable to injected signals. This confirms the saliency map functions as a topological visualizer rather than a binary detector.

5. Significance and Claims

The paper claims to provide a statistically robust resolution to the signal dilution barrier inherent in applying frozen Vision Transformers to gravitational-wave spectrograms. By abandoning global average pooling in favor of Vector-Quantized patch-level indexing and Top-k scoring, the framework enables the detection of spatially extended morphologies that were previously invisible to unsupervised models.

The authors emphasize that this approach does not claim to solve the detection of ultra-short transients (sub-patch events) but successfully isolates the topological footprint of mid-band and broadband anomalies. The framework is presented as a necessary precursor for Dirichlet Process Mixture Models (DPMM) to discover unmodeled transient populations in LIGO O4a data. The work establishes that patch-level scoring is a prerequisite for effective anomaly detection in high-resolution time-frequency data, transforming the detection paradigm from a blind global average to a targeted topological isolation.