Deep multi-task learning: a review of concepts, methods, and cross-domain applications

Abstract Multi-task learning (MTL) is a machine learning method that has witnessed exponential traction due to its ability to solve multiple tasks simultaneously. Joining training-related tasks in MTL improves both predictive performance and data efficiency. The shared knowledge across tasks enables MTL models to outperform traditional single-task models. The effectiveness of MTL depends on several factors, including how tasks are chosen, how relationships between them are structured, and the choice of shared and task-specific model components. In literature, several architectures and strategies for multi-task learning have been successfully applied across different fields such as natural language processing, computer vision, healthcare, and others. However, MTL has some challenges, such as task interference, data availability, optimal task selection, and hyperparameter tuning. Recent research has introduced new techniques to address these issues, such as adaptive task balancing and parameter sharing mechanisms, although each approach has its limitations. In this review, we provide a comprehensive examination of the multi-task learning concept, and the strategies used in several different domains. It also explores the key factors influencing the success of MTL and provides a detailed classification of research efforts applying MTL in a wide range of fields, highlighting existing achievements and challenges.