PADP: progressive and adaptive data pruning for efficient incremental learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Data pruning is crucial for efficient model training.
• Existing methods are ill-suited for dynamic incremental learning environments.
• Incremental learning requires adaptive strategies due to changing data distributions.

Purpose of the Study:

• To develop a progressive and adaptive data pruning method for incremental learning.
• To address the limitations of fixed pruning rates in dynamic settings.
• To enhance model performance and reduce training costs in incremental learning.

Main Methods:

• Proposed PADP (Progressive and Adaptive Data Pruning) method.
• Introduced instant difficulty and difficulty variation scores for sample evaluation.
• Implemented a class-balance retention mechanism to ensure class representation.

Main Results:

• PADP outperforms existing data selection methods on CIFAR-100 and Tiny-ImageNet.
• Achieved up to 52.90% reduction in training time with maintained or improved accuracy.
• Demonstrated generalization across multiple incremental learning frameworks.

Conclusions:

• PADP offers an effective and practical solution for data pruning in incremental learning.
• The method dynamically adapts to changing data and model states.
• PADP significantly reduces computational costs without compromising model performance.