lOMoAR cPSD| 70487227
Deep Learning PRACTICE TEST AND EXAM QUESTIONS
MSc in Computer Science and Engineering
MSc in Electrical and Computer Engineering
Version A
Instructions
• Make sure that your test has a total of 10 pages and is not missing any sheets, then write your
full name and student n. on this page (and your number in all others).
• The test has a total of 19 questions, with a maximum score of 100 points. The questions have
different levels of difficulty. The point value of each question is provided next to the question
number.
• Please provide your answer in the space below each question. If you make a mess, clearly
indicate your answer.
• The exam is open book and open notes. You may use a calculator, but any other type of
electronic or communication equipment is not allowed.
• Good luck.
Part 1 Part 2 Part 3, Pr. 1 Part 3, Pr. 2 Total
32 points 18 points 25 points 25 points 100 points
Part 1: Multiple Choice Questions (32 points)
In each of the following questions, indicate your answer by checking a single option.
, lOMoAR cPSD| 70487227
1. (4 points) An RNN-based sequence-to-sequence model with an attention mechanism translates
an input sentence of M words into an output sentence with N words. How does the number of
computational operations (algorithmic complexity) increase as a function of M and N?
◻O(M + N)
∎O(MN)
◻O(max(M,N)2)
◻O(MN)
Solution: The correct option is O(MN), since for each of the N generated words we need to attend
to M representations for the source words.
2. (4 points) A model is trained for 30 epochs with gradient descent and it leads to the following plot
for its training and test losses:
Which of the following statements is a plausible explanation for what could be happening?
◻The model is underfitting the training data.
∎The model is overfitting the training data.
◻The model generalizes well to unseen examples.
◻None the above.
Solution: The training error is decreasing, while the test error is increasing, which suggests that
the model is overfitting the training data.
3. (4 points) A neural network is overfitting its training data. What strategies could mitigate this?
∎Increase the dropout probability.
◻Decrease the amount of training data.
Deep Learning Page 2 of 10
Deep Learning PRACTICE TEST AND EXAM QUESTIONS
MSc in Computer Science and Engineering
MSc in Electrical and Computer Engineering
Version A
Instructions
• Make sure that your test has a total of 10 pages and is not missing any sheets, then write your
full name and student n. on this page (and your number in all others).
• The test has a total of 19 questions, with a maximum score of 100 points. The questions have
different levels of difficulty. The point value of each question is provided next to the question
number.
• Please provide your answer in the space below each question. If you make a mess, clearly
indicate your answer.
• The exam is open book and open notes. You may use a calculator, but any other type of
electronic or communication equipment is not allowed.
• Good luck.
Part 1 Part 2 Part 3, Pr. 1 Part 3, Pr. 2 Total
32 points 18 points 25 points 25 points 100 points
Part 1: Multiple Choice Questions (32 points)
In each of the following questions, indicate your answer by checking a single option.
, lOMoAR cPSD| 70487227
1. (4 points) An RNN-based sequence-to-sequence model with an attention mechanism translates
an input sentence of M words into an output sentence with N words. How does the number of
computational operations (algorithmic complexity) increase as a function of M and N?
◻O(M + N)
∎O(MN)
◻O(max(M,N)2)
◻O(MN)
Solution: The correct option is O(MN), since for each of the N generated words we need to attend
to M representations for the source words.
2. (4 points) A model is trained for 30 epochs with gradient descent and it leads to the following plot
for its training and test losses:
Which of the following statements is a plausible explanation for what could be happening?
◻The model is underfitting the training data.
∎The model is overfitting the training data.
◻The model generalizes well to unseen examples.
◻None the above.
Solution: The training error is decreasing, while the test error is increasing, which suggests that
the model is overfitting the training data.
3. (4 points) A neural network is overfitting its training data. What strategies could mitigate this?
∎Increase the dropout probability.
◻Decrease the amount of training data.
Deep Learning Page 2 of 10
