NPTEL Online Certification Courses

Indian Institute of Technology Kharagpur

Deep Learning
Assignment- Week 8
TYPE OF QUESTION: MCQ/MSQ
Number of questions: 10 Total mark: 10 X 1 = 10
______________________________________________________________________________

QUESTION 1:
The input image has been converted into a matrix of size 256x256x3 and 4 kernel/filter of size
7x7 with a stride of 2 and no padding. What will be the size of the convoluted matrix?

a. 127x127x3
b. 128x128x4
c. 124x124x3
d. 125x125x4

Correct Answer: d

Detailed Solution:

The size of the convoluted matrix is given by CxC where C=((I-F+2P)/S)+1, where C is the
size of the Convoluted matrix, I is the size of the input matrix, F the size of the filter matrix
and P the padding applied to the input matrix. Here P=0, I=256, F=7 and S=2. There the
answer is 125x125, since there are 4 such kernels.
______________________________________________________________________________

QUESTION 2:
The figure below shows image of a face which is input to a convolutional neural net and the
other three images shows different levels of features extracted from the network. Can you
identify from the following options which one is correct?

a. Label 3: Low-level features, Label 2: High-level features, Label 1: Mid-level

features
 NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur

b. Label 1: Low-level features, Label 3: High-level features, Label 2: Mid-level

features
c. Label 2: Low-level features, Label 1: High-level features, Label 3: Mid-level
features
d. Label 3: Low-level features, Label 1: High-level features, Label 2: Mid-level
features

Correct Answer: d

Detailed Solution:

Convolutional NN will try to learn low-level features such as edges and lines in early layers
then parts of faces of people and then high-level representation of a face.

______________________________________________________________________________

QUESTION 3:
Which of the following statement is False about ReLU layer?

a. ReLU has expression f(x)=max(0,x)

b. The derivative of ReLU is 1 if x>0; o otherwise
c. Implementation of ReLU has more computational cost than tanh or sigmoid
d. ReLU activation function was introduced in AlexNet architecture.

Correct Answer: c

Detailed Solution:

Option c comes from the direct formula. Both tanh and sigmoid have higher computational
cost

____________________________________________________________________________

QUESTION 4:
For a transfer learning task, which layers according to you can be more generally transferred to
another task?

a. Higher layers
b. Lower layers
c. Task specific
d. Cannot comment
 NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur

Correct Answer: b

Detailed Solution:

Lower layers are more general features (for eg: can be edge detectors) and thus can be
transferred well to other task. Higher layers on the other hand are task specific.

______________________________________________________________________________

QUESTION 5:
Suppose your input is a 256 by 256 color (RGB) image, and you use a convolutional layer with
100 filters that are each 5x5. How many parameters does this hidden layer have (with bias)

a. 2501
b. 2600
c. 7500
d. 7600

Correct Answer: d

Detailed Solution:

As we have an RGB Image so each filter would be 3D, whose dimension is 5 * 5 * 3 = 75

We also have a bias so total parameters for each filter = 75+1 = 76

Now we have 100 such filters. total number of parameters= (5 * 5 * 3 + 1)* 100 = 7600

______________________________________________________________________________

QUESTION 6:
Statement 1: Adding more hidden layers will solve the vanishing gradient problem for a 2-layer
neural network

Statement 2: Making the network deeper will increase the chance of vanishing gradients.

a. Statement 1 is correct
b. Statement 2 is correct
c. Neither Statement 1 nor Statement 2 is correct
d. Vanishing gradient problem is independent of number of hidden layers of the
neural network.
 NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur

Correct Answer: b

Detailed Solution:

As more layers using certain activation functions are added to neural networks, the
gradients of the loss function approaches zero, making the network hard to train. Thus
statement 2 is correct.

____________________________________________________________________________

QUESTION 7:
Which of the following is false about CNN?

a. Output should be flattened before feeding it to a fully connected lyer

b. There can be only 1 fully connected layer in CNN
c. We can use as many convolutional layers in CNN
d. None of the above
Correct Answer: b

Detailed Solution:

Direct from classroom lecture

______________________________________________________________________________

QUESTION 8:
Let us consider a Convolutional Neural Network having three different convolutional layers in
its architecture as:

Layer-1: Filter Size – 3 X 3, Number of Filters – 10, Stride – 1, Padding – 0

Layer-2: Filter Size – 5 X 5, Number of Filters – 20, Stride – 2, Padding – 0

Layer-3: Filter Size – 5 X5 , Number of Filters – 40, Stride – 2, Padding – 0

Layer 3 of the above network is followed by a fully connected layer. If we give a 3-D
image input of dimension 39 X 39 to the network, then which of the following is the input
dimension of the fully connected layer.
 NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur

a. 1960
b. 2200
c. 4563
d. 13690

Correct Answer: a

Detailed Solution:

the input image of dimension 39 X 39 X 3 convolves with 10 filters of size 3 X 3 and takes
the Stride as 1 with no padding. After these operations, we will get an output of 37 X 37 X
10.

Output of layer 2 would be 17x17x20

Ouput of layer 3 would be 7x7x40. Flattening this gives 1960.

______________________________________________________________________________

QUESTION 9:
Consider a CNN model which aims at classifying an image as either a rose, a marigold, a lily or
orchid (consider the test image can have only 1 of the images at a time) . The last (fully-
connected) layer of the CNN outputs a vector of logits, L, that is passed through a ____
activation that transforms the logits into probabilities, P. These probabilities are the model
predictions for each of the 4 classes.

Fill in the blanks with the appropriate option.

a. Leaky ReLU
b. Tanh
c. ReLU
d. Softmax

Correct Answer: d

Detailed Solution:

Softmax works best if there is one true class per example, because it outputs a probability
vector whose entries sum to 1.
 NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur

____________________________________________________________________________

QUESTION 10:
Imagine you're training a CNN for Autonomous driving vehicle to distinguish between
pedestrian, bicycle, bike and cars in images. You have two options:

Option 1: A shallow CNN with just one convolutional layer having 1024 kernels and a few fully
connected layers.

Option 2: A deeper CNN with 4 convolutional layers having 128 kernels in first layer,128 kernels
in second layer, 256 kernels in third layer and 512 kernels in fourth layer and then fully
connected layers.

Both options use the same total number of kernels convolutional layer (=1024)

Choose the incorrect statement:

a. Option 1 will give higher inference speed since it can operate convolutions in
parallel whereas option 2 can’t be parallelized as results of next convolutional
layers are dependent on past layers
b. Option 2 , Deeper CNN, with multiple layers, can perform hierarchical feature
extraction, thereby has higher representational power and accuracy
c. Option 1 , Shallow CNN, with one convolutional, can extract 1024 features from
image, thereby has higher representational power and accuracy
d. Option 2 , Deeper CNN, with multiple layers, can extract more abstract features
that depend on features of shallower layers and therefore has higher
representational power and accuracy

Correct Answer: c

Detailed Solution:

Shallow CNN have low representation power as they fail to perform hierarchical feature
extraction and abstract feature synthesis from signal. Higher number of feature at lower
level doesn’t guarantee higher discriminative features.

______________________________________________________________________________

************END*******