A learning paradigm in which a single model is trained simultaneously on multiple related tasks to exploit shared representations and improve generalization across tasks.

Principle

Jointly optimize a model so that shared parameters capture common structure across tasks while task-specific components handle differences; leverage inductive transfer to reduce overfitting and improve data efficiency.

Demonstration

Train one neural network with shared encoder layers and separate output heads for part-of-speech tagging, named-entity recognition, and syntactic chunking so the encoder learns linguistic representations useful to all three tasks.

Misapplication

Forcing very dissimilar tasks (e.g., medical image segmentation and music genre classification) into a single shared model without architecture or loss design, producing negative transfer where performance on one or more tasks degrades.

Consequence

When tasks are suitably related and losses balanced, models typically show better generalization, require fewer parameters than multiple independent models, and can learn useful features from limited data.

Reversal

Separate single-task models trained independently, each with dedicated parameters and no shared representation between tasks.

Boundary

Applies when tasks share representational needs or input modalities; less appropriate when tasks are unrelated, have conflicting optimization objectives, or when task-specific data dominate. Requires careful task weighting, architecture design, and validation to avoid dominance by one task.

Semantic Tension

Overlaps with transfer learning and multi-objective optimization: transfer learning focuses on sequential reuse of parameters; multi-objective treats tasks as simultaneous objectives. Multi-task learning emphasizes concurrent shared representation learning across tasks.

Synthesis

Multi-task learning unifies multiple related predictive objectives into one training process so shared components learn common structure while task-specific parts preserve differences, yielding parameter-efficient models that can generalize better when tasks are compatible and carefully balanced.