Small-sample continual learning classification method with vaccine to update memory cells based on the artificial...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Immunology-inspired Computing

Background:

• Data-driven classification typically requires large datasets for robust pattern recognition.
• Small and unbalanced datasets often lead to poor classification accuracy in traditional methods.
• Batch learning methods lack continuous improvement capabilities with new data.

Purpose of the Study:

• To propose a novel continual learning classification method (SCLM) for small sample scenarios.
• To address limitations of traditional methods in handling insufficient and imbalanced data.
• To enhance classification performance through immune system-inspired learning mechanisms.

Main Methods:

• SCLM generates synthetic samples ('vaccines') by identifying group centers of training data.
• The method mimics immune B cell maturation and memory cell activation during training.
• In the test phase, SCLM updates memory cells by learning new samples for improved recognition.

Main Results:

• SCLM demonstrates superior classification performance compared to other methods with insufficient training samples.
• The data generation technique significantly boosts classification accuracy across various methods.
• Experiments on UCI datasets and reciprocating compressor fault diagnosis validate SCLM's effectiveness.

Conclusions:

• SCLM offers a viable solution for classification tasks with limited and imbalanced data.
• The immune-inspired approach enhances continual learning and classification accuracy.
• The proposed data generation strategy is beneficial for improving existing classification algorithms.