Exploring multi-granularity balance strategy for class incremental learning via three-way granular computing.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Class incremental learning (CIL) enables continuous learning from data streams.
• Catastrophic forgetting remains a significant challenge in CIL.
• Existing episodic memory replay methods face buffer limitations, causing data imbalance.

Purpose of the Study:

• To propose a novel CIL method, MGBCIL, addressing data imbalance and catastrophic forgetting.
• To leverage multi-granularity balance strategies inspired by granular computing.
• To improve performance in incremental learning scenarios.

Main Methods:

• Introduced a multi-granularity balance strategy (MGBCIL) with batch, task, and decision-level approaches.
• Employed a weighted cross-entropy loss with smoothing for batch processing.
• Utilized contrastive learning and knowledge distillation for class separation and knowledge preservation.

Main Results:

• MGBCIL demonstrated superior performance over existing methods on CIFAR-10 and CIFAR-100 datasets.
• Achieved up to 9.59% accuracy improvement and 25.45% forgetting rate reduction in specific settings.
• Effectively mitigated catastrophic forgetting by balancing new and old class samples.

Conclusions:

• MGBCIL offers an effective solution for catastrophic forgetting in CIL.
• The multi-granularity balance strategy enhances learning stability and performance.
• This approach shows significant promise for real-world incremental learning applications.