Cognitive prototype learning: Towards self-supervised and semantic-aware few-shot open-set sound recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Acoustics

Background:

• Deep learning for environmental sound classification faces challenges with limited data and open categories.
• Current models lack human-like cognitive abilities for recognizing novel sounds with few examples.

Purpose of the Study:

• To develop a cognitive prototype learning framework for few-shot open-set recognition (FSOR).
• To bridge the gap between machine intelligence and human cognitive reasoning in sound classification.

Main Methods:

• A self-supervised perception module uses contrastive learning for acoustic patterns.
• A semantic context module derives embeddings from class names.
• A semantic prototype generation module fuses acoustic and semantic features for compact prototypes.

Main Results:

• The framework achieved 64.50% accuracy and 66.29% AUROC (5-way 1-shot) and 78.36% accuracy and 75.09% AUROC (5-way 5-shot) on the ESC-50 dataset.
• Outperformed existing FSOR methods in both closed-set classification and open-set detection.

Conclusions:

• The proposed cognitive prototype learning framework effectively addresses data scarcity and category openness in environmental sound classification.
• The integration of self-supervised learning and semantic guidance enhances prototype compactness and mitigates overfitting.