Transfer Learning Quiz

Answers and explanations appear as you take the quiz.

1. Which learned starting point can transfer learning reuse for a target task?
   • Only the source dataset filename
   • A requirement to erase all learned parameters
   • A guarantee that the original output labels are the new labels
   • Pretrained parameters or representations
2. In 'feature extraction' style transfer learning, you typically:
   • Freeze most layers and train only a new output layer
   • Retrain the entire network at a high learning rate
   • Throw away the pretrained weights
   • Use no neural network at all
3. Why does the illustrated fine-tuning workflow recommend a small learning rate for pretrained weights?
   • To make the pretrained weights mathematically unable to change
   • To guarantee that every target example is memorized
   • To limit disruptive changes to the existing representation
   • To replace the need for a target loss
4. Which target setup can make fixed feature extraction worth comparing with other adaptation strategies?
   • A task defined only by unrelated output filenames
   • Limited target data with a task related to the source task
   • A requirement to discard every pretrained representation
   • A guarantee that the source task and target task must have identical labels
5. What concern arises when the target input distribution differs from the source pretraining distribution?
   • The reused representation may fit the target task poorly
   • The model necessarily stops containing parameters
   • All target examples become source training examples
   • Every evaluation score must improve
6. What did the cited image-network transfer study distinguish about features across layers?
   • Every early feature transfers perfectly to every possible task
   • Later layers contain only random unused numbers
   • All network layers necessarily encode identical features
   • Earlier features were often more general and later features more source-task specialized in its setup
7. In the documented Keras fine-tuning workflow, why keep base-model BatchNormalization layers in inference mode?
   • To make all pretrained weights update at the same rate
   • To avoid disrupting their stored moving mean and variance statistics
   • To replace the target task's output head
   • To force new moving statistics from every target batch regardless of the workflow
8. After changing layer trainable settings in the stated Keras workflow, what refreshes the compiled training configuration?
   • Changing only an output filename
   • Resetting all pretrained parameters to random values
   • Recompiling the model
   • Keeping the old compiled configuration without modification