Recurrent Network Expansion for Class Incremental Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Class incremental learning (CIL) is crucial for adaptive vision intelligence.
• Current network expansion (NE) methods face challenges like feature diffusion, parameter growth, feature confusion, and classifier bias.

Purpose of the Study:

• To introduce a novel dynamic structure, recurrent network expansion (RNE), to address limitations in existing CIL methods.
• To improve feature utilization and reduce model complexity in CIL.

Main Methods:

• Proposed RNE establishes sequential feature transfer among task experts via a shared module.
• New task experts adjust based on received features, preventing diffusion and focusing on key areas.
• RNE employs compression by replacing task experts with lightened versions and uses a decoupled classifier with pseudo-feature generation to mitigate confusion and bias.

Main Results:

• RNE significantly reduces parameters while maintaining performance.
• The method effectively alleviates feature confusion and corrects classifier bias.
• Achieved state-of-the-art (SOTA) performance on CIFAR-100, ImageNet-100, Food-101, and ImageNet-1K datasets.

Conclusions:

• RNE offers an effective solution for class incremental learning by improving feature handling and model efficiency.
• The proposed approach demonstrates superior performance in both standard and challenging CIL scenarios.