Reproduce “ShortcutBreaker: Low-Rank Noisy Bottleneck and Frequency Filtering Block for Multi-Class Unsupervised Anomaly Detection” 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
Peng Tang, Xiaobin Hu, Tingcheng Li, Yang Nan, Tobias Lasser, Hongwei Bran Li
A framework called ShortcutBreaker tackles the "identity shortcut" issue in multi-class unsupervised anomaly detection (MUAD) by integrating a Low-Rank Noisy Bottleneck (LRNB) and a Frequency-Filtering Unit (FFU). The approach demonstrated superior anomaly detection and localization on benchmarks like MVTec-AD, ViSA, Universal Medical, and Real-IAD, achieving up to 99.8% image-level AUC and 98.5% pixel-level AUC on MVTec-AD, alongside improvements in computational efficiency.
Key insights
- Matrix rank inequality provides a robust theoretical basis for preventing identity shortcuts, leading to a wider anomaly score gap between normal and anomalous samples.
- Processing features in the frequency domain offers inherent global modeling capabilities, effectively mitigates identity collapse more reliably than self-attention, and improves computational efficiency.
- The combination of local (preventing self-copying) and global (randomly impairing cross-token copying) masking within the FFU is crucial for comprehensively suppressing shortcut learning in ViT features.
- Dual injection of the LRNB across multiple decoder stages further amplifies the perturbation, leading to enhanced anomaly discrimination.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.