Topics · How it works
Multi-task learning
Contents
Multi-task learning trains one computer model to make several related predictions together; for creators seeking to understand Instagram recommendations, the evidence presented concerns Meta research and Facebook, not confirmed use on Instagram123. A typical model shares an initial processing stage across prediction goals, then uses separate parts for each goal, reducing the computing resources needed to provide predictions3. Meta’s research examines how to share useful information selectively between goals and balance learning across them, rather than relying only on information common to every goal23. The evidence presented does not establish Instagram-specific controls or effects on creators’ reach123. The latest research discussed is Meta’s November 2024 MultiBalance paper, which concerns recommendation models rather than offering Instagram-specific creator guidance3.
Learning several predictions together
Meta research and patents. These may not describe how Instagram works today.
Meta’s November 2024 research describes a typical multi-task learning model as having a shared processing part followed by separate parts for individual prediction goals3. Meta’s research explains that the shared part turns input information into a form that each goal’s separate part can use to learn its own prediction3. Meta’s April 2023 AdaTT research also describes sharing knowledge between tasks as a way to improve how accurately the model predicts outcomes2.
Sharing useful information without hurting other goals
Meta research and patents. These may not describe how Instagram works today.
Meta’s AdaTT research says multi-task learning needs selective information sharing because some prediction goals are more closely related than others, making information common to every goal insufficient2. In Meta’s research example, predicting whether someone will share a video can draw heavily on knowledge about likes while also using different information from the more plentiful examples of people watching videos2. Meta’s MultiBalance research addresses a different problem: balancing how the shared model learns across goals so that learning one does not harm another3.
What the evidence establishes for Instagram
Meta research and patents. These may not describe how Instagram works today.
Meta’s engineering team documented multi-task learning in Facebook News Feed in January 2021, providing a concrete application outside Instagram1. In that Facebook account, Meta’s engineering team describes predicting which content will matter to each person and identifies multi-task learning among the techniques powering predictions about meaningful interactions and content quality1. Meta’s November 2024 MultiBalance research reports tests in Meta’s ads and feeds recommendation system—the system that helps choose which ads and posts people see3.