Multi-Task Learning
Multi-Task Learning refers to a method in which a single computer program learns multiple tasks at the same time instead of each one separately. The tasks share a common body of knowledge, which often makes the program better than if it had been trained on only one task.
Computer programs that learn from examples are called AI models. Normally, such a model is trained for exactly one task, such as recognizing traffic signs in photos. Multi-Task Learning takes a different approach: a single model learns several tasks at the same time. On the same photo, for example, it should recognize signs, find pedestrians, and estimate distance. The decisive point is that the tasks share a common part of the model. What the model learns for one task therefore also benefits the others.
When tasks help each other
The most important reason for Multi-Task Learning is data scarcity. For many tasks there are only a few well-labeled examples, because labeling costs people and time. A model that learns from just 500 examples often simply memorizes them. If it is additionally trained on a related task with 50,000 examples, it has to find more general patterns. These general patterns then also help with the small task.
A second reason is simply efficiency. Three separate models need three times the memory, three times the computing time, and three times the maintenance. A shared model handles everything in a single pass. In a self-driving car, this is not a luxury but a requirement: the onboard computer has limited computing power and still has to answer many questions at the same time.
But there is also the reverse case. If the tasks fit together poorly, they pull the model in different directions. Experts then speak of negative transfer: joint training makes the results worse instead of better. Multi-Task Learning is therefore not an automatic win, but a decision that has to be tried out.
Shared trunk, separate heads
The usual setup is called a shared trunk with multiple heads. The trunk is the large part of the model that all tasks use together. It processes the input, for example an image, into a kind of intermediate description. From there, a separate small head branches off for each task. This head translates the intermediate description into the concrete answer, for example which sign is visible.
During training, each task gets its own error measure. It measures how far off the answer was. These individual values are combined into an overall score, usually with different weights. The model is then adjusted so that this overall score becomes smaller. The shared trunk thus receives feedback from all tasks at the same time.
The weighting is the trickiest point. If one task gets too much weight, it dominates the training and the others wither. A comparison helps: someone studying for three subjects at once has to divide their time sensibly. Twelve hours of math and ten minutes of English is technically multitasking, but not good multitasking. There are methods that adjust the weights automatically, but a perfect solution does not exist.
From search engines to language models
Multi-Task Learning is most visible in recommendation systems. A video platform predicts several things at once for each video: whether you click, how long you watch, whether you rate it. These predictions come from a shared model because they are closely related. The camera software in smartphones and driver-assistance systems works in a similar way.
Large language models, such as those behind chatbots, are also a case of Multi-Task Learning, albeit in a special form. They are not trained on separate task heads, but instead learn translating, summarizing, and coding all as text prediction. The tasks simply appear as instructions within the text itself. This training on many types of instructions is called instruction tuning.
In company announcements and trade articles, the term usually appears as an argument for cost savings. When someone writes that a model handles ten tasks at once, they often mean lower operating costs. Multi-Task Learning should not be confused with multimodality. Multimodal means that a model processes different types of input, such as text and image. Multi-task means that it answers different questions.