A Multi-task Model for Infant Crying Detection and Reasoning

Clinical studies have shown that infant crying is a crucial signal containing physical and mental information, such as hunger and pain, which can provide valuable insights into infants' pathology and demand. However, existing studies either focused on infant crying detection or reasoning (needs/diseases), where the limited data and label types hinder the model's generalization to unknown infants and reasons. To this end, we propose a multi-task Infant Crying Detection and Reasoning (ICDR) model for both the tasks of cry detection and reasoning, which utilizes a shared extractor to extract the deep representations and incorporates two classifiers for different tasks. In this way, ICDR can augment data by mixing datasets from associated tasks and introducing inductive bias as the regularization term to mitigate overfitting. Extensive experiments were performed on four infant crying datasets, showing that ICDR outperforms its corresponding single-task model in both tasks, exhibiting a 0.31% and 6.12% improvement in F1-score for infant crying detection and reasoning when using FNN as the shared extractor, obtaining 0.11% and 1.38% improvement when using CNN as the shared extractor. These results demonstrate that multi-task learning can efficiently leverage data from related tasks to enhance the model's generalization for infant crying detection and reasoning.