Semi-Supervised Active Learning for Sound Classification in Hybrid Learning Environments.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Signal Processing

Background:

• Supervised learning for sound classification requires extensive labeled data, which is often scarce.
• Data scarcity leads to models with poor generalization capabilities.
• Minimizing human annotation is crucial for efficient model training.

Purpose of the Study:

• To propose an efficient method combining confidence-based Active Learning and Self-Training.
• To minimize the need for human annotation in sound classification model training.
• To evaluate the proposed method in both pool-based and stream-based scenarios.

Main Methods:

• Instances are pre-processed by calculating classifier confidence scores.
• Low-confidence instances are sent to human annotators.
• High-confidence instances are automatically labeled by the machine.
• The approach integrates Active Learning and Self-Training strategies.

Main Results:

• The proposed method significantly reduces the number of required labeled instances compared to Passive Learning, Active Learning, and Self-Training.
• A 52.2% reduction in human-labeled instances was achieved in both pool-based and stream-based scenarios.
• The approach demonstrated efficacy in a sound classification task with 16,930 instances.

Conclusions:

• The combined Active Learning and Self-Training approach effectively addresses the challenge of limited labeled data in sound classification.
• This method offers a practical solution for building robust sound classification models with reduced annotation effort.
• The findings highlight the potential for significant efficiency gains in machine learning model development.