What is multitask training?

Definition

In multitask training, examples from several tasks contribute to a shared model. For a recommendation system, those tasks might include finding relevant items, ranking candidates, and responding to language requests. Shared learning can help the model reuse patterns across tasks.

The benefit is not automatic. Tasks may reinforce one another, compete for capacity, or require different data quality. Comparing multitask and single-task versions helps show whether the shared approach improves the outcomes that matter.

ELI5

Multitask training is like practicing several related skills with one learner. An AI model sees examples for more than one kind of job and tries to learn what those jobs have in common.

For example, the same model might practice finding songs from a description and choosing which songs to recommend next. A test can show whether learning both helps or hurts either job.

Frequently asked questions

Why train one model on several tasks?

The tasks may share useful patterns, reducing the need for a completely separate model for each kind of work.

Can multitask training make performance worse?

Yes. If tasks conflict or are poorly balanced, gains on one task can come with weaker results on another.

Videos explaining multitask training

  1. Yves Raimond and Jacqueline Wood in blue shirts around the headline “SPOTIFY'S LLM RECS” on a black background.