Introduction
A medical student doesn't learn anatomy, physiology and pathology in total isolation — each subject helps the others. Knowing how the heart works makes heart disease easier to understand.
Multi-task learning (MTL) trains one model on several related tasks at the same time, so that what it learns for one task helps the others. Examples:
- a self-driving car's vision model that detects cars, pedestrians, lanes and traffic signs with one shared network;
- a recommendation model that predicts both "will the user click?" and "will they buy?";
- a language model that does tagging, sentiment and question answering together;
- a medical model predicting several diagnoses from the same scan.
Benefits:
- better accuracy, especially for tasks with little data — they borrow strength from related tasks;
- less overfitting — the shared part must be useful for all tasks, which acts like regularization;
- efficiency — one model to train, store and run instead of several.