Reproduce “Improvements of Discriminative Feature Space Training for Anomalous Sound Detection in Unlabeled Conditions” on OpenResearch

About this paper

Takuya Fujimura, Ibuki Kuroyanagi, Tomoki Toda

The research improves discriminative Anomalous Sound Detection (ASD) performance in scenarios lacking meta-information labels by enhancing feature extraction and introducing various pseudo-labeling techniques. These methods collectively boosted detection accuracy, with pseudo-labeling reducing performance gaps by over 30% in some unlabeled conditions and external pre-trained models or triplet learning often outperforming baselines.

Key insights

• Combining multi-resolution spectrograms captures complementary information, leading to more robust feature representations for ASD.
• Subspace loss effectively enhances feature learning by promoting robust individual network training, offering a parameter-efficient alternative to existing auxiliary losses.
• Pseudo-labeling strategies can substantially recover performance in unlabeled conditions, often approaching or sometimes exceeding ground-truth label performance, though the optimal method varies by machine and environment.

Summary by alphaXiv.