Core Concepts
Investigating Pixel-Level Anomaly Detection in Continual Learning to enhance real-world applications.
Abstract
Abstract:
Investigates Pixel-Level Anomaly Detection in Continual Learning.
Adapts state-of-the-art techniques for the CL setting.
Introduction:
Anomalies defined as deviations from normal data.
Unsupervised techniques crucial for label-free learning.
Related Work:
Three families of CL techniques: rehearsal-based, regularization-based, architecture-based.
Continual Learning Approach:
Replay approach effective for reducing Catastrophic Forgetting.
Anomaly Detection Methods:
DRAEM, STFPM, EfficientAD, Padim, PatchCore, CFA, FastFlow tested and adapted for CL setting.
Experimental Setting:
MVTec Dataset used for evaluation.
Metrics include AUC ROC, f1 score, PR AUC, AU PRO.
Results:
PatchCore shows optimal performance with minimal forgetting.
Conclusions and Future Work:
Integration of AD techniques into CL framework successful.
Stats
"A significant advantage of the unsupervised techniques is that they do not require labeled data to learn from."
"The related literature suggests that the Experience Replay approach appears to be the most effective and practical solution to reduce Catastrophic Forgetting."
"PatchCore emerges as the optimal choice, achieving a f1 pixel-level score of 0.58."
Quotes
"Experience Replay is expected to have low forgetting and be computationally efficient."
"Memory Bank-based approaches tend to require more memory than other methods."
"PatchCore method demonstrates no sign of forgetting."