Multitask Learning
Also known as: MTL, Joint Learning, Shared Representation Learning, Çok Görevli Öğrenme
Multitask Learning (MTL) is a machine learning paradigm in which a model is trained simultaneously on multiple related tasks, sharing representations across them to improve generalization. Introduced formally by Rich Caruana in 1997, MTL draws on the intuition that auxiliary tasks act as inductive bias, providing extra supervision signals that help the shared layers learn richer, more robust feature representations than single-task training would yield.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use MTL when you have multiple related prediction tasks that share input structure — for example, part-of-speech tagging and named entity recognition in NLP, or simultaneous detection and segmentation in computer vision. It is most effective when tasks share informative features and datasets are limited, since auxiliary tasks supply extra supervision. Avoid MTL when tasks are unrelated or when one task's gradients consistently harm others (negative transfer). Single-task baselines should always be established first; if a single task already has abundant data and strong performance, the MTL benefit may be marginal.
Strengths & limitations
- Improves generalization by leveraging complementary supervision signals from related tasks
- Acts as an implicit regularizer, reducing overfitting especially in low-data regimes
- Reduces total inference cost: one shared forward pass serves multiple task heads simultaneously
- Encourages learning of task-agnostic representations that can support downstream transfer
- Negative transfer can occur if tasks are insufficiently related, degrading performance on one or more tasks
- Task loss weighting (lambda_t) requires careful tuning and is problem-dependent
- Shared capacity may be insufficient when tasks have very different complexity or output structures
- Debugging is harder than single-task training because errors in one task can propagate through shared weights
Frequently asked
How is MTL different from transfer learning?
Transfer learning trains on a source task first and then adapts to a target task sequentially, typically freezing or fine-tuning pretrained weights. MTL trains on all tasks simultaneously from the start, with continuous two-way gradient flow between tasks. Both exploit shared representations, but MTL is preferred when tasks are available together, while transfer learning is preferred when the target task data arrives later or is scarce.
What is negative transfer and how can it be mitigated?
Negative transfer occurs when learning one task degrades performance on another, typically because their gradient directions conflict in the shared parameter space. Mitigation strategies include gradient projection (projecting conflicting gradients onto a common subspace), uncertainty-based loss weighting to down-weight dominant tasks, task grouping based on similarity metrics, and architectural separation of task-specific layers to limit the shared capacity that conflicting tasks can corrupt.
Should I always use equal loss weights across tasks?
Rarely. Equal weights implicitly assume all tasks have the same scale and importance, which is almost never true in practice. Common approaches include manual tuning on a validation set, uncertainty-based dynamic weighting (which treats each task weight as a learned noise parameter), and gradient normalization methods that scale task losses so their gradient magnitudes are comparable. Starting with equal weights and then tuning is a reasonable baseline strategy.
Sources
- Caruana, R. (1997). Multitask learning. Machine Learning, 28(1), 41–75. DOI: 10.1023/A:1007379606734 ↗
How to cite this page
ScholarGate. (2026, June 2). Multitask Learning. ScholarGate. https://scholargate.app/en/deep-learning/multitask-learning
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Curriculum LearningDeep learning↔ compare
- Knowledge DistillationDeep learning↔ compare
- Transfer LearningMachine learning↔ compare