A survey on few-shot class-incremental learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models face challenges with real-time data scarcity.
• Few-shot class-incremental learning (FSCIL) trains models on new tasks with minimal data, risking catastrophic forgetting and overfitting.

Purpose of the Study:

• To provide a comprehensive survey of Few-shot Class-Incremental Learning (FSCIL).
• To synthesize existing research from both theoretical and applied perspectives.
• To offer novel categorizations and identify future research directions in FSCIL.

Main Methods:

• Categorization of theoretical FSCIL approaches into five subfields: traditional machine learning, meta-learning, feature-based, replay-based, and dynamic network structures.
• Review of over 30 theoretical and 20 applied research studies.
• Performance evaluation of recent theoretical research on benchmark FSCIL datasets.

Main Results:

• FSCIL research is categorized into five distinct theoretical approaches.
• FSCIL demonstrates significant applications in computer vision (image classification, object detection, segmentation), natural language processing, and graph analysis.
• Recent theoretical methods show promising performance on benchmark datasets.

Conclusions:

• FSCIL is crucial for enhancing the practicality and adaptability of deep learning models.
• The survey provides a structured overview of FSCIL methodologies, performance benchmarks, and application domains.
• Future research should focus on advancing FSCIL theory, exploring new problem setups, and expanding its applications.