35

1A
D2

87
Paper / Subject Code: 42371 / Deep Learning

62

34
3F

37
22

AC

35
26-Dec-2023 10:30 am - 01:30 pm 1T01877 - B.E. Computer Science &

D2

87
7C

62

4
71
Engineering (Artificial Intelligence & Machine Learning) (R-2019-20 C Scheme)

53
3F
22
80

AC
73

23
D2
7C
73
(Sem VII) / 42371 - Deep Learning QP CODE: 10043892

F8

71

C6
E7

22
80

3
Duration: 3Hrs. Maximum Marks : 80

A
2

87
7C
CB

D
77

1
F

7
2
0
NB:

09

3
E

3
2
8

1A
2

7
7C
CB

3
43

2D
(1) Question No. 1 is compulsory.

8
7

F
E7

37
80
53

09

23
(2) Attempt any three questions out of the remaining five.

C2

87
B

3
23

D
7

07
C
(3) All questions carry equal marks.

34

3F
E7

2
C6

C2
38
5

0
(4) Assume suitable data, if required and state it clearly.

2
CB
3
1A

43

2D
77
62

07
3

09
37

C2
C

8
5

2
CB

3
Q1. ATTEMPT ANY FOUR. [20]

3
A

43
87

2D
7
2

07
1

7
6

53
Design AND gate using Perceptron.
3F
a.

09
7

C2
C

38
3

CB
3
A
D2

43
7
Suppose we have N input-output pairs. Our goal is to find the parameters

77
b.

07
F8

71

C6

53

09
22

that predict the output y from the input x according to some function y= xw.

BE

8
23

3
23
1A

43
7
7C

7
9C
D

8
Calculate the sum-of squared error function E between predictions y and

7
6

3
F

7
2
80

BE
C

5
3

30
3
2
inputs x. The parameter w can be determined iteratively using gradient

3
3
1A
D2

7
7C
73

77
2

C
34
8
descent. For the calculated error function E, derive the gradient descent

C6
F
7

09
37
22
0

BE
5
3
BE

23
A
D2

43
87
update rule w←w-α .
7C
3

9C
77

1
9C

C6

3
F

7
2
0

35
3
E

30
3
2
8

c. Explain dropout. How does it solve the problem of overfitting?

30

1A
D2

87
7C
B

34
7
9C
34

d. Explain denoising auto encoder model.

6
F
7

37
22
0

35
3
E

8
35

30

Describe sequence learning problem.

A
e.

87
7C
CB

62
2D
7

1
62

34

3F
E7

37
80

AC
9

C2
AC

35

87
CB

Q2. a. Explain Gated Recurrent Unit in detail. [10]

43

2D
77

71
62

07
71

3F
3

b. What is an activation function? Describe any four activation functions. [10]

E

73
2
AC

8
5

0
73

2
7C
CB

73
3

43

2D

F8
62
F8

71

E7

80
Q3. a. Explain CNN architecture in detail. Suppose, we have input volume of [10]
3

09

23
C2
AC

5
3

73

3
3
D2

43

32*32*3 for a layer in CNN and there are ten 5*5 filters with stride 1 and

2D
77
2

07
C
F8

71

C6

53

9
22

pad 2; calculate the number of parameters in this layer of CNN.

BE

C2
8
3

0
73

3
3
1A
D2

43
7C

7
b Explain early stopping, batch normalization, and data augmentation. [10]
62

07
9C
F8

7
3
7
22
80

BE
C

38
5
3

0
3

3
1A
D2

43
7
7C

Q4 a. Explain RNN architecture in detail. [10]

73

77
2

9C
F8

C6

53
E7

37
22
80

b. Explain the working of Generative Adversarial Network.

BE

[10]
3

30
23
A
D2

87
7C
CB

73

9C
34
1

6
3F

Q5 a. Explain Stochastic Gradient Descent and momentum based gradient [10]

E7

7
22
80

C
09

30
3

3
1A
D2

87
7C
CB

73

descent optimization techniques.

43

34
6
F
E7

37
22
80

b. Explain LSTM architecture.

AC
9

35
3
30

D2

87
7C
CB

73

62

[10]
71
4

3F
E7

2
80
53

AC
9

73
C2

Q6 a. Describe LeNET architecture. [10]

30

2
CB

73
3

2D

71
62

07
4

F
E7

b. Explain vanishing and exploding gradient in RNNs. [10]

53

73
2
C

8
30

2
7C
CB

73
3
1A

2D

F8
62

*****************************
E7

80
53

23
C2
C

30

CB

73
3
1A

2D
62

07
4

7
53

9
7

BE

C2
C

8
30
73

3
3
1A

7
62

07
9C
4
8

7
53
3F

BE
C

38
0
73

3
1A
D2

43

77
62

9C
F8

53
7

BE
C
23

30
73

3
1A

62

9C
D

4
F8

3
7
22

5
23

30
73

3
1A
7C

2
D

34
F8

6
7
22
80

AC

35
23

73
7C
73

62
D

F8

71
22
80

AC
23

73
7C
73

71
3F
7

22
80
BE

73
2
7C
73

2D

F8

43892 Page 1 of 1
9C

80

23
BE

2
30

7C
73

2D
9C

0
BE

C2
38
30

309CBE773807C22D23F87371AC623534
77

07
9C