SET - B

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

(COMMON TO INFORMATION TECHNOLOGY)
BATCH (2022 - 2026): III Year/ VI SEM
CCS355 NEURAL NETWORKS AND DEEP LEARNING (R-2021)
MODEL EXAMINATION
DATE:22-05-2025 TIME: 3 HRS TOTAL MARKS: 100

PART A: (10X2=20 Marks)

1. What are the advantages of neural networks over conventional computers? (C315.1,
PO1,PO3,PO5)
2. What are the design Parameters of ANN? (C315.1, PO1,PO3,PO5)
3. List the application of adaptive resonance theory based neural networks (C315.2, PO1,PO3)
4. What is the main purpose of Hop field networks/. (C315.2, PO1,PO3)
5. List the various types of learning rules. (C315.3, PO1,PO3)
6. Name the various features in computer vision. (C315.3, PO1,PO3)
7. What is the necessary for using early stopping and dropouts in deep learning? (C315.4,
PO1,PO3)
8. Identify the challenges in data augmentation? (C315.4, PO1,PO3)
9. List out the applications of auto encoders (C315.5, PO1,PO3)
10. How does RNN handles variable length inputs? (C315.5, PO1,PO3)
PART B: (5X13=65 Marks)

11 a) (i) Explain in detail about McCullah-Pitts Neuron architecture model with neat diagram
(C315.1, PO1, PO3, PO5) (7)
(ii) Compare and contrast Biological neural networks and artificial neural networks (C315.1,
PO1, PO3, PO5) (6)
(Or)
b) (i) What is Perceprons? Discuss in detail about perceptron learning algorithms and
convergence theorems. (C315.1, PO1,PO3,PO5) (8)

1
(ii) ) For the network shown in below calculate net input to the output neuron Y using activation
function binary sigmoid and bipoar sigmoid. (C315.1, PO1,PO3,PO5) (5)

12. a)(i) Discuss in detail Bidirectional associative memory (BAM) with a neat architectural
diagram.(C315.2, PO1,PO3) (7)
(ii) Evaluate the adaptability and learning capabilities of Adaptive Resonance Theory Network

(ART) in dynamic environments.(C315.2, PO1,PO3) (6)

(Or)
b) (i) A hetero associative network is trained by Hebb outer product rule for the input vectors
S=[x1, x2,x3,x4] to the output row vectors t=[t1,t2], find the weight matrix :. (C315.2, PO1,PO3)
s1=[1 1 0 0] t1=[1 0]
s2=[1 1 1 0] t2=[0 1]
s3=[0 0 1 1] t3=[1 0]
s4=[0 1 0 0] t4=[1 0]

(ii) with a neat diagram explain the kohonen self organizing maps and list its important features
(6)
13. a) (i) Define and explain Extreme Learning Machine Model and provide an example of its
application. (C315.3, PO1,PO3) (6)
(ii) A fully connected convolutional network or FCN became one of the major successful
architecture. Can you identify what are the advantages of FCN which make it a successful
architecture for semantic segmentation. (7)
(Or)
2
b) (i) Image compression is a type of data compression applied to digital image, to reduce their
cost for storage or transmission .Differentiate between lossy and lossless image compression in
terms the image quality, image size, and common usage. (C315.3, PO1,PO3) (6)
(ii) Discuss in detail about how do you differentiate between useful and non useful features in
computer vision. (C315.3, PO1,PO3) (7)
14.a) (i) perform one iteration of the training on the back propagation neural networks in below
figure. Assume the neurons use the sigmoid activation function for the forward and backward
pass. The target output is 0.5 and the learning rate is 1.(C315.4, PO1,PO3) (7)

(ii) Data augmentation is often used to increase the amount of of data you have. will you apply
the data augmentation to the test data? Justify. (C315.4, PO1,PO3) (6)
(Or)
b) (i) Feed forward neural network for a given application uses 10 neurons in its input layer and
5 neurons in its output layer. Determine the possible number of hidden neurons required in the
hidden layer. Assume that the network require only one hidden layer for the given application.
(C315.4, PO1,PO3) (6)
(ii) J(θ) =2θ2+2θ+2 is a given cost function ? find the correct weight updatae rule gradiate
descent optimization at step t=t+1. Consider ɑ=0.01 to be the learning rate .find a function (θ
0,θ1), if θ0 and θ1 are initialized at alocal minimum, then what should be the value of θ0 and θ1
after a single iteration of the gradient descentalgorithm? .(C315.4, PO1,PO3) (7)
15. a) (i) Explain in detail about vanishing gradient problem and how is it related to recurrent
neural networks?(C315.5, PO1,PO3) (6)
(ii) An auto encoder with 5 hidden layers has 10,004 number of parameters. If we use a dropout
in each layers with 50% drop rate , what will be the number of parameters of that auto encoder?
(C315.5, PO1,PO3) (7)
(Or)
b) (i) Elaborate in detail on how does Recurrent Neural Networks differ fron a feed forward
network in terms of its structure and operation. (C315.5, PO1,PO3) (6)
3
(ii) An autoencoder consists of 128 input neurons and 32 hidden neurons. If the network weights
are represented using single precision floating point numbers (size 4 bytes) then what will be the
size of the weight matrix? (C315.5, PO1,PO3) (7)
PART C: (1X15=15 Marks)
16.a) (i) An RGB input image has been converted into a matrix of size 257x257x3 and a kernel
of size 7x7x3 with a stride of 2 and padding =3 is used for 2D convolution. What will be the size
of output of convolution ? what changes can be made in the context of the size of the CNN
networks to improve the accuracy?. (C315.3, PO1,PO3) (8)
(ii) Design a NOR Gate with McCullah-pitts neurons. (7)
(Or)
b) (i) In deep CNN architecture, the feature map before applying a max pool layer with a (2 x 2)
kernel is given below: (C315.3, PO1,PO3) (5)
50 30 3 4
14 12 7 10
44 1 16 20
2 5 16 20
After a few successive convolution layers, the feature map is again up-sample using Max
unspooling. If the following feature map is present before the Max-unpooling layer, what will be
the output of the max-unpooling layer?
50 10

44 20
(ii) Given a convolution layer in a network with 5 filters, filter size of 7, a stride of 3, and a
padding of 1. For an input features map of 26 x 26 x26 , what is the output dimensionality after
applying convolution layer to the input? (C315.3, PO1,PO3) (5)
(ii) Given an input image of size 128x128 pixels and a CNN architecture with the following
specifications: .(C315.5, PO1,PO3) (5)
 Convolutional layer with 16 filters of size 7x7 , using stride of 2 and no padding.
 Max pooling layer with a pool size of 2 x2
 Fully connected layer with 512 neurons
 Output layer for binary classification
 Calculate the total number of parameters (weights + biases) in this CNN

PREPARED BY- VALIDATED BY- APPROVED BY-