AI Paper Atlas

Research map / Label-efficient and robust learning / OOD and anomaly detection

HARMONY: Hierarchical Anchor Retrieval on Manifold for Oblivious-source Acoustic Anomaly Detection

Yuhang Zhang, Yang Li, Mingyuan Li, Qiang Wu

NeurIPS 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Anomaly detection with contaminated data

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HARMONY, a metadata-free universal acoustic anomaly detection method that uses hierarchical anchor retrieval on a manifold to monitor mixed machine populations without machine IDs.

Problem

Acoustic anomaly detection usually needs per-machine or per-domain training with explicit metadata, but machine identifiers are often incomplete, unreliable, or unavailable in deployment; the goal is a single model that handles mixed machine populations without IDs.

Approach

Argues the main failure mode of metadata-free universal ASD is not a weak detector but the lack of machine-specific reference structure; uses hierarchical anchor retrieval on a manifold to recover the relevant reference context for each oblivious input (based on the title and partial abstract only).

Measured effects reported in the paper

What was changedWhereMeasuredEffect
remove the learned projection (use raw BYOL-A features) (ablation)HARMONY · DCASE 2020AUC (%): 62.40 -> 90.00Improved
add spectral anchor initialization (ablation)HARMONY · DCASE 2020AUC (%): 81.40 -> 90.00Improved
add soft assignment (ablation)HARMONY · DCASE 2020AUC (%): 87.60 -> 90.00Improved
replace flat search with hierarchical IVF retrieval (ablation)HARMONY · DCASE 2020AUC (%): 90.00 -> 89.80; latency 42.3 ms -> 5.4 msMixed
train per-region detectors on HARMONY-routed partitions instead of unified trainingPatchCore · DCASE 2020AUC (%): 62.10 -> 91.80Improved
train per-region detectors on repartitioned dataAudio-MAE · DCASE 2020AUC (%): 86.10 -> 92.40Improved
per-region training increases cost for trained detectorsAudio-MAE · DCASE 2020training time (h): 14.2 -> 108.0Hurt
apply HNSW to the HARMONY representation instead of HARMONY routing (ablation)HARMONY · DCASE 2020AUC (%): 89.80 -> 89.71No clear effect
use vMF soft prototypes instead of HARMONY (ablation)HARMONY · DCASE 2020 (routed)AUC (%): 89.80 -> 85.71Hurt
add projection to raw BYOL-A features (geometry metrics) (ablation)HARMONY · DCASE 2020quantization error: 0.842 -> 0.181Improved
contaminate training data with anomalies (ablation)HARMONY · DCASE 2020AUC (%): 89.65 -> 83.31Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy