Multi-Task Learning

Learning Multiple Tasks

Share representations across tasks.

Hard Sharing

Encoder shared across tasks. Task-specific heads. Prevents overfitting through regularization.

Soft Sharing

Each task has own parameters. Regularization encourages similarity. LoRA: low-rank adaptation.

Challenges

Task weighting: how to balance? Negative transfer: tasks hurt each other. Imbalanced tasks.

Key Takeaways

1. Share encoder across tasks
2. Balance task weights
3. Can improve or hurt