Source-linked AI summary
PerMix: Anomalous sound detection by predicting the perturbation ratio of mixed samples
Jihoon Choi, Jung-Woo Choi
TL;DR
Because anomalous samples are scarce and variable, ASD commonly learns from normal data using mixup-based pretext tasks, which can mismatch labels and ignore mixing ratios. PerMix predicts perturbation ratios and uses dominant-class contrastive learning; it achieves state-of-the-art DCASE 2020 Task 2 performance with both lightweight and large-model backbones.
Problem
Mixup-based ASD pretext tasks can mismatch supervision when the labeled sample is a minority component and apply the same angular margin regardless of mixing ratio.
Source: §1 · p. 1Method
PerMix predicts the dominant class's perturbation ratio with a ratio-scaled angular margin and forms supervised contrastive positives by dominant class.
Source: §1 · p. 1, §2.1 · p. 2Results
PerMix achieves state-of-the-art DCASE 2020 Task 2 performance with both lightweight encoder and large audio-model backbones.
Source: §4 · p. 4Takeaways & Limitations
Its losses keep perturbed samples from being compressed toward the normal class center while pulling normal samples closer together.
Source: §4 · p. 4
Abstract
from arXiv · showhide
In anomalous sound detection based on self-supervised learning, the performance of the detection model largely depends on the selection of the pretext task through which the characteristics of normal data are learned. A common pretext task is to mix samples from different classes and classify the mixed sample as its target class with an angular margin loss. However, this supervision becomes mismatched when the target sample is the minority component of the mixture and the margin remains fixed irrespective of the mixing ratio. To address these issues, we propose a pretext task that directly predicts the perturbation ratio of mixed samples by reformulating the logit so that the angular margin is scaled by the perturbation ratio. In addition, we adopt a supervised contrastive loss modified for mixed samples, in which positive pairs are formed according to the dominant class. Experiments show that our approach achieves state-of-the-art results on the DCASE 2020 Task 2 benchmark. Our code is available at https://github.com/kaistss/PerMix.
1. INTRODUCTION
Because anomalous samples are scarce and highly variable, ASD commonly uses self-supervised learning on normal data, including classification-based tasks. Existing mixup approaches can mismatch labels and apply fixed angular margins across mixing ratios; PerMix addresses both issues.
1. INTRODUCTION
- Self-supervised learning trained exclusively on normal data is the predominant ASD paradigm because anomalous samples are scarce and highly variable.Classification-based self-supervision is widely used, with learned representations reported to disentangle machine characteristics in noisy environments.Source: §1 · p. 1
- Classification-based ASD models classify metadata such as machine IDs or operating conditions, often using angular margin losses to improve class compactness and separation.Later work added lightweight encoders or fine-tuned large audio models, while mixup-based approaches include NAMix and OS-SCL.Source: §1 · p. 1
- Mixup methods can assign the first sample's label even when it is a minority component and apply a fixed angular margin regardless of mixing ratio.These choices mismatch the supervision and penalize mixtures equally despite differing perturbation amounts.Source: §1 · p. 1
- PerMix predicts perturbation from the original class, scales the angular margin by perturbation ratio, and forms contrastive positives by dominant class.The training-objective changes apply to both lightweight encoders and large audio models; experiments report state-of-the-art DCASE 2020 Task 2 results.Source: §1 · p. 1
2. PROPOSED METHOD
PerMix constructs anchor-dominant mixtures of normal samples, learns a perturbation-sensitive dominant-class objective alongside dominance-based contrastive learning, and uses nearest-neighbor cosine distance for inference.
2. PROPOSED METHOD
- Each normal mini-batch sample is embedded and mixed with R randomly selected peers; Dirichlet(0.5) weights assign the largest coefficient to the anchor.Source: §2 · pp. 1–2
2.1. Perturbation Ratio Prediction
- The dominant-class ratio λ_i defines the perturbation amount; PerMix scales the angular margin accordingly, reducing the effective margin and class-center pull as λ_i falls.The dominant-class ratio includes weights from same-class samples; the remaining mixture ratio comes from other classes.Source: §2.1 · p. 2, §2 · p. 2
2.2. Dominance-based Contrastive Learning
- The dominance-based contrastive loss forms positive pairs from mixed samples sharing the same dominant class.Source: §2.2 · p. 2
2.3. Training and Inference
- Training jointly minimizes the perturbation and dominance-based contrastive losses; inference scores each test sample by cosine distance to its nearest training-embedding neighbor in memory.Source: §2.3 · p. 2
3. EXPERIMENTAL RESULTS
Experiments compare PerMix with existing ASD models on DCASE 2020 and evaluate generalization on DCASE 2022. PerMix leads the reported DCASE 2020 development-set averages and achieves the highest harmonic mean on both DCASE 2022 sets.
3.1. Experimental Setup
- Experiments use DCASE 2020's 41 machine-type/ID classes and DCASE 2022's 250 machine-type/attribute classes, with lightweight and fine-tuned BEATs backbones.The models are evaluated with AUC and pAUC; pAUC covers the low false-positive-rate region [0, 0.1].Source: §3.1 · pp. 2–3
3.2. Performance Comparison
- 95.89% average AUC with TFSTgram-MFN and 96.56% with BEATs make PerMix the best overall on the DCASE 2020 development set, ahead of TFSTgram (OS-SCL).Table 1 reports AUC and pAUC for each machine type.Source: Table 1 · p. 3, §3.2 · p. 3
- On ToyConveyor, PerMix keeps normal samples in compact clusters and separates anomalies that OS-SCL maps into normal clusters.The comparison uses BEATs embeddings; ToyConveyor signals are dominated by background noise.Source: Figure 1 · p. 3, §3.2 · p. 3
- Replacing ArcFace with perturbation-ratio loss improves average AUC and pAUC with both backbones, while adding dominant-class contrastive loss to it yields the best performance.The losses are complementary: perturbation-ratio loss separates mixed samples according to perturbation but does not pull normal samples together; dominant-class contrastive loss helps compact them.Source: §3.2 · p. 3, Table 2 · p. 4
- PerMix improves development-set AUC over OS-SCL by 1.61 pp with TFSTgram-MFN and 1.93 pp with BEATs; its BEATs pAUC gain is 3.49 pp.On the evaluation set, BEATs PerMix improves AUC by 2.09 pp over OS-SCL.Source: §3.2 · pp. 3–4, Table 3 · p. 4
- On BEATs, ArcFace, NAMix, and OS-SCL have comparable training costs and similar overall mean performance of around 92%, averaging AUC and pAUC across development and evaluation sets.Source: §3.2 · p. 4, Table 4 · p. 4
- With DCASE 2022 hyperparameters unchanged, BEATs PerMix achieves the highest harmonic mean on both development and evaluation sets among the compared models.The comparison includes a lightweight encoder, a GMM-based method, and BEATs-LoRA.Source: Table 5 · p. 4, §3.2 · p. 4
4. CONCLUSION
PerMix predicts the dominant class’s perturbation ratio in mixed normal samples, adjusting the angular margin and using dominance-based contrastive loss to address mixed-sample supervision mismatch. It achieves state-of-the-art performance on DCASE 2020 Task 2 with both lightweight and large audio model backbones.
4. CONCLUSION
- PerMix achieves state-of-the-art performance on DCASE 2020 Task 2 with both lightweight encoder and large audio model backbones.Source: §4 · p. 4
- Its perturbation-ratio prediction loss adjusts the angular margin to the dominant-class ratio, preventing perturbed samples from being compressed toward the normal class center.Source: §4 · p. 4
- Its dominance-based contrastive loss resolves mixed-sample supervision mismatch while pulling normal samples closer together.Source: §4 · p. 4