CS 7643 Quiz 3
Study online at https://quizlet.com/_goc4o6
1. What is modeling error in rela- Given a particular NN architecture, the actual model that
tion to NN architecture? represents the real world may not be in that space.
2. What happens to modeling er- It decreases
ror as model complexity in-
creases?
3. What happens to optimization It decreases
error as model complexity in-
creases?
4. What is estimation error in ma- The difference between the predicted values and the actual
chine learning? values on a test set.
5. Can minimizing training error No.
guarantee good performance
on a test set?
6. What is optimization error in Even if your NN can perfectly model the world, your algo-
neural networks? rithm may not find good weights that model the function.
7. What happens to modeling er- It reduces.
ror as model complexity in-
creases?
8. What is transfer learning? The reuse of features learned on a very large dataset for a
completely new task.
9. What is the first step in transfer Train on a very large dataset.
learning?
10. What should you do with the Replace it and retrain it.
last fully connected layer when
using transfer learning?
, CS 7643 Quiz 3
Study online at https://quizlet.com/_goc4o6
11. What is done after initializing Continue training on the new dataset.
the network with weights from
a large dataset?
12. What are the two options for You can either retrain all weights (finetune) or freeze certain
training the network after ini- layers.
tializing it for transfer learning?
13. What is the effect of freezing Freezing reduces the number of parameters that need to be
layers in transfer learning? learned.
14. When does transfer learning When the source domain and target domain share similar
perform well? feature representations.
15. When does transfer learning It does not work very well if the target task's dataset is very
not work well? different from the source.
16. What is the recommendation if It is better to just train on the new data.
you have enough data in the
target domain that is different
from the source?
17. What is the purpose of freez- To learn parameters in the fully connected layers while
ing early layers in a CNN during keeping the early layers unchanged.
transfer learning?
18. What is the architecture of 2x(CONV=>MAXPOOL=>NORM)=>3xCONV=>MAX-
AlexNet? POOL=>3xFC
19. What activation function is pri- ReLU
marily used in AlexNet?
20. PCA-based data augmentation
Study online at https://quizlet.com/_goc4o6
1. What is modeling error in rela- Given a particular NN architecture, the actual model that
tion to NN architecture? represents the real world may not be in that space.
2. What happens to modeling er- It decreases
ror as model complexity in-
creases?
3. What happens to optimization It decreases
error as model complexity in-
creases?
4. What is estimation error in ma- The difference between the predicted values and the actual
chine learning? values on a test set.
5. Can minimizing training error No.
guarantee good performance
on a test set?
6. What is optimization error in Even if your NN can perfectly model the world, your algo-
neural networks? rithm may not find good weights that model the function.
7. What happens to modeling er- It reduces.
ror as model complexity in-
creases?
8. What is transfer learning? The reuse of features learned on a very large dataset for a
completely new task.
9. What is the first step in transfer Train on a very large dataset.
learning?
10. What should you do with the Replace it and retrain it.
last fully connected layer when
using transfer learning?
, CS 7643 Quiz 3
Study online at https://quizlet.com/_goc4o6
11. What is done after initializing Continue training on the new dataset.
the network with weights from
a large dataset?
12. What are the two options for You can either retrain all weights (finetune) or freeze certain
training the network after ini- layers.
tializing it for transfer learning?
13. What is the effect of freezing Freezing reduces the number of parameters that need to be
layers in transfer learning? learned.
14. When does transfer learning When the source domain and target domain share similar
perform well? feature representations.
15. When does transfer learning It does not work very well if the target task's dataset is very
not work well? different from the source.
16. What is the recommendation if It is better to just train on the new data.
you have enough data in the
target domain that is different
from the source?
17. What is the purpose of freez- To learn parameters in the fully connected layers while
ing early layers in a CNN during keeping the early layers unchanged.
transfer learning?
18. What is the architecture of 2x(CONV=>MAXPOOL=>NORM)=>3xCONV=>MAX-
AlexNet? POOL=>3xFC
19. What activation function is pri- ReLU
marily used in AlexNet?
20. PCA-based data augmentation
