Multi-Task Learning Quiz

Answers and explanations appear as you take the quiz.

1. How is a multi-task learning model set up during training?
   • Training one model on several related tasks with shared representations
   • Training a separate model for every task
   • Using only unlabeled data for a single task
   • Removing all hidden layers from a network
2. In a typical hard-parameter-sharing MTL network, what is shared and what is separate?
   • Everything is separate per task
   • A shared backbone with task-specific output heads
   • Only the loss function is shared
   • Only the input data, never the layers
3. Why can multi-task learning improve generalization?
   • It eliminates the need for any data
   • It guarantees zero training error
   • Signals from related tasks regularize the shared representation and reduce overfitting
   • It removes the need for task-specific heads
4. What is 'negative transfer' in multi-task learning?
   • When adding tasks speeds up training
   • When the model has too few parameters
   • When all tasks share the same labels
   • When one task's learning hurts performance on another due to conflicting signals
5. How is the overall loss usually formed in multi-task learning?
   • As a weighted sum of the individual task losses
   • By taking only the largest task loss
   • By ignoring all but one task
   • By multiplying the losses together
6. Why are large language models considered inherently multi-task?
   • They can only perform one fixed task
   • A single network handles translation, summarization, coding, and Q&A
   • They use a different model for each prompt
   • They never share parameters across tasks
7. In soft parameter sharing, how are task models usually related?
   • They use one identical model with no task-specific outputs
   • They share no information and cannot exchange features
   • They keep separate parameters but add a constraint or penalty encouraging similarity
   • They share only the training labels
8. What does PCGrad do when two task gradients conflict?
   • It discards every task except the largest one
   • It changes the task labels
   • It replaces the shared backbone with separate datasets
   • It projects away conflicting gradient components before combining updates