Class-Incremental Learning: Survey and Performance Evaluation on Image Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Incremental learning is crucial for efficient, scalable AI systems, mimicking human learning.
• Catastrophic forgetting, a performance drop on prior tasks after learning new ones, is the primary challenge.
• Class-incremental learning (CIL) requires models to distinguish all classes without task identifiers, a recent advancement over task-incremental learning.

Purpose of the Study:

• To provide a comprehensive survey of current class-incremental learning methods for image classification.
• To conduct an extensive experimental evaluation of thirteen CIL methods.
• To analyze CIL performance under various conditions, including domain shifts and different network architectures.

Main Methods:

• Surveying existing literature on class-incremental learning techniques.
• Implementing and evaluating thirteen distinct CIL methods on multiple large-scale image classification datasets.
• Designing experimental scenarios to assess performance across varying domain shifts and network architectures.

Main Results:

• Extensive experimental evaluation of thirteen class-incremental learning methods.
• Analysis of method performance across multiple large-scale image classification datasets.
• Investigation into the impact of small and large domain shifts on CIL performance.

Conclusions:

• The survey and evaluation provide a benchmark for class-incremental learning methods in image classification.
• Understanding performance under different domain shifts is critical for robust CIL.
• Comparative analysis across network architectures informs future CIL research and development.