NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur
Deep Learning
Assignment- Week 11
TYPE OF QUESTION: MCQ/MSQ
Number of questions: 10 Total mark: 10 X 1 = 10
______________________________________________________________________________
QUESTION 1:
Which of following can be a target output of semantic segmentation problem with 4 class?
a.
0 1 0 1 1 0 0 0 1 0 0 0
0 1 0 0 0 0 1 0 0 1 0 0
1 0 0 0 0 0 0 0 0 0 0 0
I II III IV
b.
0 1 0 1 0 0 0 0 1 0 1 0
0 1 0 0 1 0 1 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0 1 1
I II III IV
c.
0 1 0 1 0 0 0 0 1 0 0 0
0 1 0 0 0 1 1 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0 1 1
I II III IV
d.
0 1 0 1 0 0 0 0 1 0 0 0
0 1 0 1 0 0 1 0 0 0 0 0
1 1 0 0 1 0 0 0 1 0 1 1
I II III IV
Correct Answer: c
Detailed Solution:
Target output should be one hot encoded vector at every pixel location. It should one if the pixel
belongs to that particular class otherwise 0.
NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur
QUESTION 2:
Suppose you have a 1𝐷 signal 𝑥 = [1,2,3,4,5] and a filter 𝑓 = [1,2,3,4], and you perform stride
2 transpose convolution on the signal 𝑥 by the filter 𝑓 to get the signal 𝑦. What will be the
signal 𝑦 if we don’t perform cropping?
a. 𝑦 = [1,2,5,8,9,14,13,20,19,26,3,4]
b. 𝑦 = [1,2,3,4,5,4,3,2,1]
c. 𝑦 = [1,2,5,8,9,14,13,20,17,26,15,20]
d. 𝑦 = [0,0,5,8,9,14,13,20,19,26,0,0]
Correct Answer: c
Detailed Solution:
1 2 3 4 5
1 1
2 2
1*3+2*1=5 3 1
4*1+2*2=8 4 2
3*2+1*3=9 3 1
4*2+2*3=14 4 2
3*3+1*4=13 3 1
4*3+2*4=20 4 2
3*4+1*5=17 3 1
4*4+2*5=26 4 2
15 3
20 4
______________________________________________________________________________
NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur
QUESTION 3:
What are the different challenges one face while creating a facial recognition system?
a. Different illumination condition
b. Different pose and orientation of face images
c. Limited dataset for training
d. All of the above
Correct Answer: d
Detailed Solution:
Please refer to the lecture of week 11.
______________________________________________________________________________
QUESTION 4:
Fully Connected Convolutional network or FCN became one of the major successful network
architectures. Can you identify what are the advantages of FCN which makes it a successful
architecture for semantic segmentation?
a. Larger Receptive Field
b. Mixing of global feature
c. Lesser computation required
d. All of the above
Correct Answer: d
Detailed Solution:
Please refer to the lecture of week 11. It has a larger receptive field by using strided Convolution
layer, also it mixes global feature and number of computations are reduced as we are down
sampling the image resolution.
____________________________________________________________________________
NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur
QUESTION 5:
In a Deep CNN architecture, the feature map before applying a max pool layer with (2x2) kernel
is given bellow.
12 6 15 9
19 2 7 18
14 2 17 6
3 5 19 2
After few successive convolution layers, the feature map is again up-sampled using Max Un-
pooling. If the following feature map is present before Max-Unpooling layer, what will be the
output of the Max-Unpooling layer?
5 6
8 13
a.
0 0 0 0
5 0 0 6
8 0 0 0
0 0 13 0
b.
5 5 6 6
5 5 6 6
8 8 13 13
8 8 13 13
c.
5 0 6 0
0 0 0 0
8 0 13 0
0 0 0 0
d. None of the above
Correct Answer: a
Detailed Solution:
NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur
Please refer to lectures of week 11.
____________________________________________________________________________
QUESTION 6:
What could be thought as disadvantage of Fully Convolutional neural network for semantic
segmentation addressed by other researchers?
a. It has a fixed receptive field, so the object with lesser size than the receptive
filed will be missed by the network.
b. Down Sampling the image dimension over the depth makes the feature map
sparse.
c. It requires lot of computation.
d. None of the above
Correct Answer: a
Detailed Solution:
The fixed receptive filed is the disadvantage of Fully convolution network for semantic
segmentation addressed by “Learning Deconvolution Network for Semantic Segmentation”
paper.
____________________________________________________________________________
QUESTION 7:
What will be the dice coefficient of following two one hot encoded vector? (|A|=no of 1 bit)
A 1 0 1 0 0 0 1 1 1 0 0 1 0 1
B 1 0 0 0 0 1 1 1 0 0 0 1 0 0
a. 0.83
b. 0.41
c. 0.67
d. 0.90
Correct Answer: c
Detailed Solution:
No of 1 bit in A=7
No of 1 bit in B =5
NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur
Overlapping 1 bit =4
2∗|𝐴∩𝐵| 2∗4
Dice Coefficient = |𝐴|+|𝐵|
= = 0.67
5+7
______________________________________________________________________________
QUESTION 8:
What will be the value of dice coefficient between A and B?
(Consider, |A|= sum of all elements.)
a. 0.93
b. 0.77
c. 0.11
d. 0.89
Correct Answer: a
Detailed Solution:
|A|=7.82
|B|=8
0 0 0 0
A∩B= 0 0 0 0
0.89 0.85 0.88 0.91
0.99 0.97 0.95 0.97
|A∩B|=7.42
NPTEL Online Certification Courses
Indian Institute of Technology Kharagpur
Dice Coefficient = 2*|A∩B|/(|A|+|B|)=2*7.42/ (8+7.82) =0.93
______________________________________________________________________________
QUESTION 9:
In FaceNet, why the L2 normalization layer is used?
a. To constrain the embedding function in a d-dimensional hyper-sphere.
b. For regularization of weight vector, i.e. L2 regularization.
c. For getting a sparse embedding function.
d. None of the above.
Correct Answer: a
Detailed Solution:
Using the L2 normalization layer we impose the constrain that ||f(x)||22 =1 . This will constrain
the embedding function to live on the d-dimensional hyper-sphere.
______________________________________________________________________________
QUESTION 10:
What is the use of Skip Connection in image denoising networks?
a. Helping de-convolution layer to recover an improved clean version of image.
b. Back propagating the gradient to bottom layers, which makes the training easy.
c. To create the direct path between convolution layer and the corresponding
mirror de-convolution layer.
d. All of the above.
Correct Answer: d
Detailed Solution:
Please refer to lecture of week 11.
______________________________________________________________________________
______________________________________________________________________________
***********END*******
