Medboost sources stuvia
CS 231N
Convolutional Neural Networks for
Visual Recognition Spring 2019 Sample
Midterm Exam
May 7, 2019
Full Name:
Question Score
True/False (20 pts)
Multiple Choice (40 pts)
Short Answer (40 pts)
Total (100 pts)
Welcome to the CS231N Midterm Exam!
• The exam is 1 hour 15 minutes.
• No notes or electronic devices are allowed.
I understand and agree to uphold the Stanford Honor Code during this exam.
Signature: Date:
1
, Medboost sources stuvia
1 True / False (20 points)
Fill in the circle next to True or False, or fill in neither. Fill it in completely like this: ●.
No explanations are required.
Scoring: Correct answer is worth 2 points. To discourage guessing, incorrect answers are worth -1 points.
Leaving a question blank will give 0 points.
1. Your Neural Network is not gradient checking. It could be that it’s because you’re using the AdaGrad
update instead of Vanilla SGD.
◯ True
◯ False
SOLUTION:
False. Gradient check happens before multiplication with learning rates.
2. If the input to a ConvNet is a zero image (all zeros), then the class probabilities will come out uniform.
◯ True
◯ False
SOLUTION:
False. May have nonuniform bias terms.
3. Turning off L2 weight regularization will likely lead to higher accuracy on the training set.
◯ True
◯ False
SOLUTION:
True. Turning off regularization increases chances of overfitting, which may give us higher accuracy on
the training set.
4. It’s sufficient for symmetry breaking (i.e. there will be no symmetry in the weights after some updates)
in a Neural Network to initialize all weights to 0, provided that the biases are random.
◯ True
◯ False
SOLUTION:
True. Non-zero asymmetric biases contribute to non-zero asymmetric upstream gradients, which when
combined with asymmetric layer input gives us asymmetric gradients for the weights.See the following
diagrams for a 2 layer neural network:
2
CS 231N
Convolutional Neural Networks for
Visual Recognition Spring 2019 Sample
Midterm Exam
May 7, 2019
Full Name:
Question Score
True/False (20 pts)
Multiple Choice (40 pts)
Short Answer (40 pts)
Total (100 pts)
Welcome to the CS231N Midterm Exam!
• The exam is 1 hour 15 minutes.
• No notes or electronic devices are allowed.
I understand and agree to uphold the Stanford Honor Code during this exam.
Signature: Date:
1
, Medboost sources stuvia
1 True / False (20 points)
Fill in the circle next to True or False, or fill in neither. Fill it in completely like this: ●.
No explanations are required.
Scoring: Correct answer is worth 2 points. To discourage guessing, incorrect answers are worth -1 points.
Leaving a question blank will give 0 points.
1. Your Neural Network is not gradient checking. It could be that it’s because you’re using the AdaGrad
update instead of Vanilla SGD.
◯ True
◯ False
SOLUTION:
False. Gradient check happens before multiplication with learning rates.
2. If the input to a ConvNet is a zero image (all zeros), then the class probabilities will come out uniform.
◯ True
◯ False
SOLUTION:
False. May have nonuniform bias terms.
3. Turning off L2 weight regularization will likely lead to higher accuracy on the training set.
◯ True
◯ False
SOLUTION:
True. Turning off regularization increases chances of overfitting, which may give us higher accuracy on
the training set.
4. It’s sufficient for symmetry breaking (i.e. there will be no symmetry in the weights after some updates)
in a Neural Network to initialize all weights to 0, provided that the biases are random.
◯ True
◯ False
SOLUTION:
True. Non-zero asymmetric biases contribute to non-zero asymmetric upstream gradients, which when
combined with asymmetric layer input gives us asymmetric gradients for the weights.See the following
diagrams for a 2 layer neural network:
2
