Reproduce “Cut out and Replay: A Simple yet Versatile Strategy for Multi-Label Online Continual Learning” on OpenResearch

About this paper

Xinrui Wang, Shao-yuan Li, Jiaqiang Zhang, Songcan Chen

The paper introduces CUTER (CUT-out-and-Experience-Replay), a strategy for Multi-Label Online Continual Learning that addresses catastrophic forgetting, missing labels, and class imbalance by identifying and replaying label-specific image regions. This approach achieves superior performance across benchmark datasets, notably improving average mAP by up to 5.83 percentage points compared to previous state-of-the-art methods on PASCAL VOC.

Key insights

• Label-specific region identification and feature learning are crucial for effectively mitigating missing labels and co-occurrence bias in MOCL, offering a more granular approach than whole-image methods.
• Pre-trained vision models exhibit inherent zero-shot localization capabilities, which can be quantitatively assessed using an annotation-free protocol based on the Fiedler value of feature map graphs.
• A low-rank regularization term (nuclear norm) applied to the feature map's adjacency matrix can effectively maintain and improve a model's localization ability, preventing its degradation during continual learning.

Summary by alphaXiv.