Multi-Task Learning (MTL) is a training paradigm in which a single model is jointly optimized across multiple related tasks simultaneously, sharing representations and parameters to exploit commonalities and complementary structure among tasks. In Wi-Fi sensing and crowd analysis, MTL matters because it improves generalization and data efficiency — a shared backbone trained on tasks such as activity recognition, localization, and person counting simultaneously can learn richer, more transferable feature representations than task-specific models trained in isolation, which is especially valuable when labeled data is scarce or domain shift is a concern. Key variants include hard parameter sharing, where all tasks share a common feature extractor with task-specific output heads, and soft parameter sharing, where each task maintains its own parameters that are regularized to remain similar, with more recent approaches incorporating attention-based or gradient-balancing mechanisms to handle conflicting task gradients.
Dictionary term