Learning with few samples in deep learning for image classification, a mini-review.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels in many tasks but requires large datasets.
• Obtaining abundant data is often challenging in real-world applications.
• Few-shot learning (FSL) addresses data limitations by enabling rapid learning from few samples.

Purpose of the Study:

• To provide a comprehensive survey of recent few-shot classification methods.
• To define the few-shot classification problem and its challenges.
• To propose a novel taxonomy for organizing and understanding FSL techniques.

Main Methods:

• Classifying methods into four main categories: Data Augmentation, Metric-based, Optimization-based, and Model-based.
• Analyzing sample-level and task-level data augmentation strategies.
• Examining feature embedding and metric functions in metric-based methods.
• Comparing self-learning and mutual learning in optimization methods.
• Discussing memory-based, rapid adaptation, and multi-task learning in model-based methods.

Main Results:

• A structured taxonomy categorizing diverse few-shot classification approaches.
• Discussion on the applicability and effectiveness of each method in various scenarios.
• Comparative analysis of the advantages and disadvantages of different FSL techniques.

Conclusions:

• Few-shot classification is crucial for overcoming data scarcity in deep learning.
• The proposed taxonomy offers a clear framework for understanding and comparing FSL methods.
• Future research directions and prospects in few-shot learning are outlined.