Reproduce “Memory Efficient Continual Learning for Edge-Based Visual 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
Manuel Barusco, Lorenzo D'Antoni, Davide Dalle Pezze, Francesco Borsatti, Gian Antonio Susto
An investigation from the University of Padova explores continually learning visual anomaly detection on edge devices, demonstrating that the PaSTe architecture improves performance and reduces catastrophic forgetting, while introducing compressed replay techniques that significantly cut memory usage at varying performance costs.
Key insights
- PaSTe's architectural optimizations, specifically its frozen and shared initial layers, inherently contribute to reduced catastrophic forgetting and improved performance in continual learning scenarios.
- Storing intermediate feature maps instead of raw images in replay memory can reduce memory footprint by 50% without compromising pixel-level anomaly detection performance.
- A clear trade-off exists between the aggressiveness of feature compression techniques (quantization) and the precision of pixel-level anomaly detection accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.