A Multi-task Model for Infant Crying Detection and Reasoning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Infant Health Monitoring

Background:

• Infant crying provides critical physical and mental health information.
• Current models struggle with generalization due to limited data and single-task focus.

Purpose of the Study:

• To develop a multi-task Infant Crying Detection and Reasoning (ICDR) model.
• To enhance infant cry analysis by integrating detection and reasoning tasks.
• To improve model generalization for unknown infants and reasons.

Main Methods:

• Proposed a multi-task learning framework (ICDR) with a shared feature extractor and two classifiers.
• Utilized data augmentation by mixing datasets from associated tasks.
• Incorporated inductive bias as a regularization term to mitigate overfitting.

Main Results:

• ICDR model outperformed single-task models on four infant crying datasets.
• Achieved improved F1-scores for both cry detection and reasoning tasks.
• Demonstrated effectiveness with both Feedforward Neural Network (FNN) and Convolutional Neural Network (CNN) as shared extractors.

Conclusions:

• Multi-task learning effectively leverages related task data to boost infant cry analysis.
• The ICDR model shows significant improvements in generalization for infant cry detection and reasoning.
• This approach offers a more robust solution for understanding infant vocalizations.