POSSESSION OF MOBILES IN EXAMS IS A UFM PRACTICE.
Name \lcwJ-a~ jhl Enrolment No. d-,{jal.l(QllQ /81.\.
Jaypee Institute of Information Technology, Noida
I
Tl Examination, Odd Semester 2023
Semester - 7th
Course Title: Introduction to Deep Learning Max. Hours: 1Hr
Course Code: 18812CS428 Max. Marks: 20
After pursuing this course, the students will be able to:
C434-3.1 Summarized tJ1e basic concept of machine learning and deep learning.
C434-3.2 ldentify the basic theory of ANN, probability theory, error minimization, and regulari:zation techniques
C434-3.3 Experiment with the Convolution Neural Networks for image recognition and Computer Vision.
C434-3.4 Apply Recurrent Neural Networks and LSTM for temporal data
C434-3.S Assess the various deep learning techniques on real-time problems.
~ T h e National Center for Health Statistics publishes data on heights and weights in Vital and Health
Statistics. A random sample of 11 males aged 18-24 years gave the following data, where x denotes
height, in inches, and y denotes weight, in pounds. (CO-l(Understanding), 6 Marks]
J/"Deteranine the regression equation f o v ~ a r k s ]
].:,. Apply the stochastic gradient descent to the above equation to minimize error up to 2 iterations
assu1ning learning rate a=0.1 fl Marks
X 65 67 71 71 66 75 67 70 71 69 69
y 175 133 185 163 126 198 153 163 159 151 155
p---A coin is flipped I 00 times. Given that there were 55 heads, find the maximum likelihood estimation
for the probability p of heads on a single toss. [CO-2(Understanding), 3 Marks)
· y Given in the table reduce the dimension from 2 to I using the principal component analysis (PCA)
~ algorithm. [C0-1 (Understandin2), 6 Marksl
Feature!·
XI 4 8 13 7
X2 II 4 5 14
4. ,,0:1assify the two-dimensional pattern shown in the figure below using the perceptron network. The
symbol '+' indicates the + I and the symbol 'x' indicates the -1. Assume initial weights and bias zero
and a=l. The target output for pattern "C" is+ 1 and for "A." is -I
[CO-2(Understanding), 3 Marks]
+ + +
+ X X
+ + + Pattern "C"
X + x·
+ + + .Pattern "A"
+ ·x +
~ you implement XNOR with MP neuron Justify your answer with a proper explanation.
[CO-2(Understanding), 2 Marks]
,( POS§JfSSION OF MO BIL ES IN EXA M IS
UFM PRACTICE.
Nam e ~ &\'oh cl,{ J b"
Enr ollm ent No. &..o\o ~ ~
Jay pee Ins titu te of Info rma tion Technology
, Noida
Test-2 Exa min atio n, Od d Semester 2023
B.Tech VII Semester
Course Title: Introduction to Deep Learnin
g Max. Time: 1 Hr
Course Code: 18B12CS428 •
Ma L Marks: 20
COi Understand the basic con t of machine leam
CO2
i .
Understand the basic theo of ANN, robabilit
C03 A the convolutional neural networks for i . iz.ation and re ulariz.ation techni ues.
C04 A recurrent neural networks and LSTM for d com uter vision.
COS Assess the various dee ues on rea - blems.
Note: Attempt all the questions.
~o ns ide r the convolutional neural network
defi
in the shape of output volume and number ~fpa ned by layer in the left column of table given below. Fill
rameters at each layer. You can write activation
fonnat (H,W,C), where H, W, C are height, widt shape in the
h and channel dimensions respectively. Unless
assume padding I, stride 1 where appropriate. specified,
Use notation as given belpw
• CONVx-N denotes the convolutional layer"wi
th N filters with height and width equal to x.
• POOL-n denotes an x n max pooling layer
with stride n and Opadding.
• FLA.TTEN flattens its inputs
• FC-N denotes a fully connected layer with
N neurons.
Layer Activation Volume Dimensions
j!l_Eut Number of parameters
32 X 32 X 3
CO~'V3-8 0
ReLU
POOL-2
CONV3-16
ReLU
POOL-2
FLATTEN
FC-10
~v en following input matrix, perform conv [C0-3(Apply); 6 mar u]
olution with stride 1 and 2.
1 1 1 0 0
0 1 1 1 0 1 0 1
0 0 1 1 1 0 1 0
0 0 1 1 0 1 0 1
0 1 1 0 0
Image input 3 x 3 kernel
I
Apply max pooling and average pooling to the
results from above convolutions. Use 2 x 2 filte
l for both the pooling operations. [C0-3(Apply r with stride
), 6 marks]
~ ~ u have a dataset Dl with 1 million labelled training exam
with l 00 labelled training examples. Your frien ples for classification and dataset D2
d trains a model from scratch on dataset D2.
train on D 1, and then apply transfer learning You decide to
to train on D2. State one problem your friend
with his approach. How does your approach is likely to find
address this problem? [ CO-l(Understandin
g), 2 Marks]
I
·, r
✓while the model does well on training set, it only achieves an accuracy of 85%
You conclude that model is over fitting and plan to use L on validation data set:
1 or L2 regularization to fix the issue. However,
before you can do so, you learn from lab mate that some
of the examples in the data may be incorrectly
labelled. Which form of regularization would you prefe
r to use and why? [ CO-2 (Und erstan ding) , 1
Marks]
~on side r the following network. Use backpropagation
to determine the updated weights of wS and w8
after first iteration, assum ing a learning rate of 0.5. the sigmo
id function is used as an activation function
in all layers. [C0-2 (lJnd entan ding) , S marks]
0-0~
!..l_
'1st POSSE SSION OF MOBIL ES IN EXAMS IS A UFM PRACT ICE.
tnc
-nt
IC, Name~ ~:JR~ Enrollmen t Nodd ~tl 4n <'(,
l11
Jaypee Institute of Information Technology, Noida
End Semester Examination, Odd Semester 2023
-
=1J
Semeste r- 7th
Course Title: Introduction to Deep Learning
Max. Hours: lHr
rI Course Code: 18B12CS428
Max. Marks: 3S
r
After punuing this course, the students will be able to:
C434-3.1 Summarized the basic concept of machine learning and deep learning.
C434-3.2 Identify the basic theory of ANN, probability theory. error minimization, and regularization techniques
C434-3.3 Experiment with the Convolution Neural Networks for image recognition and Computer Vision.
C434-3.4 Apply Recurrent Neural Networks and LSTM for temporal data
C434-3.S Assess the various deep learning techniques on real-time problems.
Note: Attempt all questions.
A-:- Compute the principal components for the following two-dimensional dataset.
X {xl, x2 )-={(l,2),(3,3),(3,5),(5,4),(5,6),(6,5),(8, 7),(9,8))
[CO-I (Understanding), 4 Marks)
y Given the following data:
Item· xt x2 Class ✓Are the data linearly s~parab~e? State reasons for your answer.
A 1 2 yes=l ~We ~ill train a perceptron on the data. We add a bias xo = -1 to
B each of the data points. Suppose the current weights to be w = (0,
.I
2 1 yes=l
·c • -1, 1). .Assume a learning rate of f!.l. How should the weights be
1 1 no~O
update~ -if point. A is cons~dered? How would the weights nave been
.L_D_~ ll___l_o___
ln_o_=_O__ updated if the algorithm instead had consider ed point B?
[C0-2 (Unders tanding) , 1+2=3 Marks]
V
Using the back-pro pagation algorithm,
find the new weights (only wo, w1, w1)
for the net shown in the figure. It Ois
presente d with the input pattern [O, l]
and the target output is 1. Use a lea1ning
rate a=0.25 and binary -siginoidal
Vo2 = 0.5
activatio n function.
V22-= 0.4
~o Note: Calculations should
perform ed up to two decin,al places.
[C0-2 (Unders tand), 5 Marks]
be
0 1
~~xpla in what effect \\'ill the- f?llo~i ng operati~~s have on the bias and varian~
e of your model. Reply
~ as one of 'increa ses', 'decreases , or no change · 1n each of the
cells: .
~ Regularizing the weights in a linear/logistic regression model t •
Y. Using dropout to train a deep neural network l
f Increasing the number of hidden units in an artificial neural network •,
y{Wh y are convolutional layers more commonly used than fully-connected layers t;,. ..
for image
processing? r.
• [C0-3 (Apply), 1+1+1+2=5 Marks]
I Compare GoogLeNet and Residual networks (ResNets). What are the main archite
each, and how did they lead to improvements over previous design? Use diagram
ctural features of
s of the networks as
appropriate. [C0-3(Apply), 4 Marks)
/.
d?(Ske tch a simple recurrent network, with inputx , output y, and r~current
state h. '
ti,,(Giv e the updated equations for a simple RNN unit in terms of x, y, and/,.
• • Jinearity Assum e it uses Tanh non-
.
(,if What is the difference betwee n the c and I, recurrent states in an LSTM (Long
Short-T erm Memor y)
recurrent unit?
(CO-4 (Apply), 1+2+3 =5 Marks)
Y . ~ a t do you m~ ?Y self-attention? Explain with the help of an examp
le. What is the benefit of
~~he aded attenti on 1n a transfo rmer? • •
ryupp ose you have been given a sentenc e of two words "cat" and "sat".
•
(i) The word embed dings of these words are [0.85, 0.29, -0.61] and [-0.37, 0.72,
0.45] respect ively.
Final input word .e~bed qings ar~__(o~~y~,_,~~!~.!}~_g~pQ~i!i_~n ve_c!<;>r to these
word embed dings.
Th~ positio n vectors are [0.04, 0.05,. 0.06] and [0.07, 0.08, 0.09]. Find ~nal
,vord en1heddings.
(ii) Consid ering output of part (i) as input find the query vector, key vector,
and value vector if Wq
and \\'Tk and Wv are 3x3 matrix given as:
Wq = [[0.1, -0.2, 0.3], [-0.4, 0.5, -0.6], [0.7, -0.8, 0.9]]
Wk= [[-0.9, 0.8, -0.7], [0.6, -0.5, 0.4], [-0.3, 0.2, -0.1]]
Wv = [[0.2, 0.3, -0.4], [0.5, -0.6, 0.7], [-0.8, 0.9, 1.0]]
(iii) . Find the output vector Z ~f the self-att ention layer in the transfo rmer
of the word "cat" using
query vector, key vector, and value vector compu ted in part(ii) assum e two
words in the
senten ce "cat" and "sat".
(C0-4 (Apply), 2+1+3 +3=9 Marks]
