Multi-task learning trains a single model to handle several related tasks at once, sharing part of its internal representation across all of them rather than building a separate model per task. When the tasks are genuinely related, this sharing acts like a built-in regularizer: patterns the model discovers for one task help it generalize on the others, which is especially valuable when data for any single task is limited.
The upside: positive transfer
If two tasks share underlying structure, recognizing objects and estimating their depth in an image, for instance, a shared representation can learn that structure once and apply it to both outputs. The model effectively gets more useful training signal per example than either task would provide alone.
The downside: negative transfer
When tasks are unrelated or actively pull the shared representation in conflicting directions, forcing them to share hurts both. The model ends up as a compromise that fits neither task as well as a dedicated model would have. This failure mode is called negative transfer, and it is the main risk multi-task learning has to manage.
Task similarity is the deciding factor
Whether sharing helps or hurts comes down almost entirely to how similar the tasks' underlying patterns actually are. Highly related tasks reinforce each other. Weakly related tasks see little effect either way. Actively conflicting tasks drag each other down.
Try it yourself
The Multi-Task Learning Lab lets you dial a task-similarity slider and watch a shared model's performance on two toy regression tasks shift from mutually beneficial to actively harmful, compared against training two separate models.
🧪 Try it yourself: the Multi-Task Learning Lab simulation lets you experiment with everything described above directly in your browser.