Multi-task learning (MTL) is a training paradigm in which a single model is simultaneously trained on multiple related tasks, leveraging commonalities and differences across tasks to improve generalization. Instead of training separate models for each task, an MTL setup might have shared layers (learning representations common to tasks) and then task-specific output layers. For example, a network might share a trunk that processes an image, then branch into one head that performs object detection, another for segmentation, and another for depth estimation. The hypothesis is that the shared representation benefits from extra supervision signals from all tasks, acting as a regularizer (the model must find features useful for all tasks). MTL often yields better performance on each task than training separate models, especially when one task has limited data and can benefit from another task’s data. It also tends to yield more compact models (one model for many things). A challenge is balancing tasks (some might dominate the loss) and ensuring tasks are indeed synergistic (some combinations might conflict).