Berbagi Parameter Keras dalam Kuis Jaringan Multi-Tugas

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. What does hard parameter sharing mean?
   • Each task has a separate full network with a similarity penalty
   • Tasks share labels but no model parameters
   • One output head predicts every task without task identity
   • Tasks share a common hidden trunk and use task-specific output heads
2. How does hard sharing differ from soft parameter sharing?
   • Soft sharing keeps separate task parameters and encourages them to remain similar
   • Soft sharing always uses one shared trunk
   • Hard sharing stores no parameters
   • The two terms mean the same architecture
3. Why can hard sharing reduce overfitting for related tasks?
   • Each task receives more independent parameters
   • Shared layers learn features that must serve multiple tasks
   • The model avoids training data
   • All tasks are forced to use identical output labels
4. How are task losses commonly combined for joint training?
   • By choosing one task and ignoring the others
   • By averaging labels across tasks
   • As a weighted sum in a shared objective
   • By using only the smallest loss
5. What can happen if one task’s loss or gradients dominate?
   • All task heads disappear
   • The tasks become independent
   • The model automatically balances every gradient
   • The shared trunk may prioritize it while other tasks are under-served
6. When is hard parameter sharing most suitable?
   • When tasks can benefit from useful shared representations
   • When every task is unrelated and incompatible
   • When a task must never influence another
   • When there is only one task
7. What does a task-specific head usually do?
   • Stores all training examples
   • Maps shared features to that task’s output
   • Balances the gradients of every other task
   • Selects which hidden trunk to use per token
8. What should be monitored when training a multi-task model?
   • Only the total parameter count
   • Only the largest task’s accuracy
   • Per-task performance and whether shared updates help or harm each task
   • The number of output heads alone