Class-Incremental Learning: A Survey.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models excel in closed-world scenarios but struggle with emerging classes.
• Class-Incremental Learning (CIL) aims to continuously learn new classes without performance degradation.
• Catastrophic forgetting is a major challenge in CIL, where models lose knowledge of previously learned classes.

Purpose of the Study:

• To comprehensively survey recent advancements in Class-Incremental Learning (CIL).
• To provide a unified and empirical evaluation of 17 CIL methods on benchmark tasks.
• To highlight the impact of memory budget on CIL evaluation and advocate for fairer comparison protocols.

Main Methods:

• Systematic review and categorization of existing CIL techniques.
• Empirical evaluation of 17 CIL algorithms on standard image classification datasets.
• Analysis of CIL performance considering varying memory budget constraints.

Main Results:

• Categorization of CIL methods based on their approaches to mitigate catastrophic forgetting.
• Empirical insights into the strengths and weaknesses of different CIL algorithms.
• Demonstration of how memory budget significantly influences comparative performance of CIL methods.

Conclusions:

• Existing CIL methods offer diverse strategies to combat catastrophic forgetting.
• Fair comparison of CIL algorithms necessitates aligning memory budgets during evaluation.
• Standardized evaluation protocols are crucial for advancing the field of incremental learning.