Glossary term · · 3 min read
Task Weighting
A model reads “My parcel hasn't arrived; I need it tomorrow” and learns to choose a department and recognize urgency. Each task receives a loss, a number assessing the error in its answer: lower is better. Suppose the department loss is 0.20 and the urgency loss is 20, because their scores use different scales. Simple addition gives 20.20. Does the larger number mean urgency should outweigh the other task so strongly in the joint score?
Task Weighting sets the coefficients by which we multiply individual task losses before combining them into a learning objective. They can be chosen manually or learned under an additional rule. This is part of Multi-task Learning, learning several tasks that use a shared part of a model. Kendall, Gal and Cipolla, §3, Equation 1 describe the general weighted sum.

The same change, a different evaluation
Multiply the urgency loss by 0.01, keeping the department coefficient at 1. The joint score is now 0.20 + 0.01 × 20 = 0.40.
Consider a proposed change to the model's settings that raises the department loss to 0.22 and lowers the urgency loss to 19. Does the joint objective regard it as an improvement?
| Urgency coefficient | Before the change | After the change | Joint objective's evaluation |
|---|---|---|---|
1 | 20.20 | 19.22 | Improvement |
0.01 | 0.40 | 0.41 | Deterioration |
0.02 | 0.60 | 0.60 | No change |
The trade-off has changed: a smaller urgency error does not always offset a larger department error. These numbers are our own example, and the table compares one proposed move; it does not simulate training. A lower score after rescaling does not mean better answers. Nor should the strength of an update be inferred from loss magnitude alone: what matters is its sensitivity to changes in settings, described by the gradient.
Where do the coefficients come from?
Urgency in this message may be ambiguous because the data do not always reveal the actual deadline. One method estimates a shared level of such uncertainty for each task and uses it to weight losses. In Kendall et al. §3.1–3.2, this is uncertainty weighting. An additional penalty prevents the trivial “solution” of silencing the tasks. It was studied on images; our message is an analogy, not a result of that paper.
A limitation: choosing coefficients does not guarantee improvement on both tasks. Their proposed directions of change may still oppose each other; Gradient Surgery corrects precisely such directions. We check the effect of the chosen weighting separately for each task on a Validation Set, examples not used to learn the model's settings.
The text and illustration were prepared with AI assistance. The numbers and message are our own educational example.