Reproduce “Mitigating Long-Tailed Anomaly Score Distributions with Importance-Weighted Loss” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jungi Lee, Jungkwon Kim, Chi Zhang, Sangmin Kim, Kwangsun Yoo, Seok-Joo Byun
ELROILAB Inc. researchers introduced Importance-Weighted Loss (IWL), a method to mitigate long-tailed anomaly score distributions in anomaly detection, a common issue when "normal" data is diverse. IWL reshapes anomaly scores towards a Gaussian distribution using importance sampling and Box-Cox transformation, leading to an average AUROC improvement of 0.043 on image datasets and AUPR gains up to 0.12 on real-world hyperspectral imaging datasets compared to conventional methods.
Key insights
- Despite theoretical expectations for Gaussian anomaly score distributions, real-world normal data diversity frequently leads to long-tailed, skewed distributions, degrading anomaly detection performance.
- Dynamically re-balancing the contribution of diverse normal instances during training, through importance weighting, can effectively mitigate bias caused by long-tailed distributions without requiring explicit sub-class labels.
- Aligning the anomaly score distribution towards a Gaussian target during training ensures a more balanced representation of all normal patterns, including minority ones, improving overall anomaly detection accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.