Active learning for bird sound classification via a kernel-based extreme learning machine.

Area of Science:

• Ornithology
• Bioacoustics
• Machine Learning
• Ecology

Background:

• Bird sound recognition is crucial for understanding evolution, biodiversity, and climate change impacts.
• The increasing volume of unlabeled bird sound data presents a significant challenge.
• Efficient methods for handling large unlabeled datasets are needed.

Purpose of the Study:

• To propose and evaluate two novel active learning (AL) methods for bird sound recognition.
• To reduce the reliance on extensive expert human annotation.
• To compare the performance of kernel-based extreme learning machines (KELM) against support vector machines (SVM) within these AL frameworks.

Main Methods:

• Development of sparse-instance-based active learning (SI-AL) and least-confidence-score-based active learning (LCS-AL).
• Integration of kernel-based extreme learning machines (KELM) with both AL methods.
• Comparative analysis against conventional support vector machines (SVM).

Main Results:

• KELM demonstrated superior performance compared to SVM across both SI-AL and LCS-AL paradigms.
• KELM achieved higher classifier capacity (60%-80% unweighted average recall) with limited human annotations.
• Significant improvements were observed: SI-AL (minimum 34.5% vs 59.0% for SVM) and LCS-AL (minimum 17.3% vs 28.4% for SVM).

Conclusions:

• The proposed SI-AL and LCS-AL methods, when combined with KELM, effectively reduce the need for manual annotation in bird sound recognition.
• KELM offers a more efficient classification approach than SVM, especially when dealing with limited labeled data.
• These findings contribute to advancing ornithological research through improved machine learning techniques for bioacoustic data analysis.