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