Multi-task learning exposes a model to several tasks during training instead of optimizing it for only one narrow objective. Related tasks can share internal representations, allowing information learned from one activity to improve performance or generalization on another.
The tasks and their sampling weights must be balanced carefully. Useful transfer can improve efficiency and breadth, while conflicting tasks can compete for model capacity or cause one objective to dominate. Evaluation should therefore report performance for each important task, not only one combined score.
Acronyms and aliases
multi-task training synonymmultitask learning variant
General terms
Frequently asked questions
Why train one model on several tasks?
Related tasks can reinforce shared skills and produce a more general model, sometimes using training data and parameters more efficiently than separate models.
Can multi-task learning hurt performance?
Yes. Tasks can interfere when their objectives conflict or their data is poorly balanced, so task mixtures and per-task results need careful monitoring.
