Multitask Learning - Page 2

Multitask learning trains a model to perform more than one task, often by sharing an encoder or intermediate layers while using separate output heads. Learning related objectives together can improve generalization, reduce the amount of task-specific data needed, and produce reusable representations. For example, a language system might learn classification, entity detection, and question answering within one training program. Benefits are not automatic: tasks can compete for capacity, produce gradients that interfere, or appear at very different frequencies. Effective designs balance losses, sampling, architecture, and evaluation so strong performance on a dominant task does not hide deterioration elsewhere. Unrelated or poorly labeled tasks may create negative transfer.