multi-task learning
/ MUL-tee-task LER-ning /
A medical student who studies anatomy, physiology, and pharmacology together often understands each one better than if they crammed them in isolation, because the subjects reinforce one another. Multi-task learning applies that idea to machines: instead of training a separate model for each job, you train one model to do several related jobs at once, letting them share what they learn.
Concretely, the model has a common core — shared layers that build a general understanding of the input — and then separate "heads," one per task, branching off the top. During training it learns from all the tasks together, so the shared core is shaped by every task's data at the same time. Predicting a face's age, gender, and emotion from one shared image representation is a textbook case. The shared part acts like a built-in form of regularization: features useful across many tasks tend to be the genuinely meaningful ones, which can reduce overfitting and squeeze more value out of limited data.
It shines when the tasks are related and labeled data for any single one is scarce — the tasks effectively lend each other examples. But it is not free. If the tasks pull in different directions, they interfere and all of them can suffer ("negative transfer"), and balancing how much each task counts in the combined loss is fiddly. A bundle of unrelated tasks is usually better off as separate models.
A self-driving system uses one shared vision backbone feeding three heads at once: detect other cars, find lane lines, and estimate distance. Because all three rely on understanding the road scene, learning them together makes the shared backbone sharper than any single task would, and it runs faster than three separate networks.
One shared core, several task heads — related jobs reinforce each other.
Multi-task learning helps only when tasks genuinely share structure. Force unrelated tasks together and they fight over the shared weights — each ends up worse than if trained alone.