Multi-task learning (MTL) involves training a model to perform multiple related tasks simultaneously, sharing knowledge across tasks. For example, a network can learn both sentiment analysis and text classification by sharing parameters in initial layers while dedicating task-specific heads in the output.
MTL improves data efficiency, reduces overfitting, and leverages complementary information from related tasks. Techniques like hard parameter sharing (common backbone layers) and soft parameter sharing (task-specific layers) balance shared and task-specific learning.
Applications include autonomous driving (e.g., detecting lanes and objects) and NLP (e.g., question answering and sentiment analysis). Task weighting and loss balancing ensure that all tasks contribute effectively to model performance.