SRM Institute of Science and Technology Mode of Exam
College of Engineering and Technology
OFFLINE
School of Computing
DEPARTMENT OF COMPUTATIONAL INTELLIGENCE
SRM Nagar, Kattankulathur – 603203, Chengalpattu District, Tamilnadu
Academic Year: 2024-2025 (ODD/EVEN)
D
Test: CLAT-2 Date: 06-11-2024
Course Code & Title: 21CSE326T - Artificial Neural Network Duration: 100 Minutes
th
Year & Sem: III & 5 Max. Marks: 50
Part - A
(10 x 1 = 10 Marks)
Instructions: Answer all
Q. No Question Marks BL CO PO *PI Code
1 In the Delta learning rule, how is the error calculated? 1 2 3 1 1.6.1
A) As the difference between the predicted output and actual output
B) As the sum of all input weights
C) As the product of input and output
D) As the sum of input activations
2 Which regularization technique adds a penalty proportional to the absolute value of the weights 1 2 3 1 1.6.1
to prevent overfitting?
A) L2 regularization
B) L1 regularization
C) Early stopping
D) Batch normalization
3 If a dataset has distinct, complex patterns across multiple dimensions, how could Multi-SOM 1 1 4 1 2.5.3
assist in organizing it?
A) By creating linear boundaries for classification
B) By combining multiple SOMs to capture different feature clusters
C) By applying supervised learning to guide the clustering
D) By ignoring outliers and random patterns
4 How does the interaction between multiple SOM layers in Multi-SOM networks en- 1 2 4 1 2.5.1
hance clustering?
A) By averaging output across layers for stable results
B) By allowing each layer to specialize in different features or data sections
C) By ignoring data from earlier layers
D) By using supervised labels in each layer
5 Which unsupervised learning algorithm is commonly used to reduce data dimensionality? 1 2 4 1 1.6.1
A) Convolutional Neural Networks (CNNs)
B) Support Vector Machines (SVMs)
C) Principal Component Analysis (PCA)
D) K-Nearest Neighbors (KNN)
6 What is the role of data selection in pretraining a neural network? 1 2 5 1 1.6.1
A) To determine the activation function
B) To choose a loss function
C) To improve generalization
D) To identify features of the input data
7 Which of the following methods helps initialize neural network weights? 1 1 5 1 1.7.1
A) Backpropagation
B) Xavier Initialization
C) Stochastic Gradient Descent
D) Momentum
8 Which stopping criteria indicates that the network has converged during training? 1 2 5 1 1.6.1
A) Loss function reaches zero
B) No change in weights after several iterations
C) Validation error continues decreasing
D) Increase in loss after every epoch
9 What does time delay in a neural network refer to? 1 1 5 1 1.7.1
A) Adding a delay to gradient updates
B) Propagating inputs with a temporal dependency
C) Limiting the number of hidden layers
D) Regularizing the model to prevent overfitting
10 Consider a neural network with a softmax output layer, and the network is 1 1 5 1 1.7.1
classifying between 3 classes. The output from the network before applying the
softmax is [2.0,1.0,0.1]. What is the softmax probability for the first class? (Use
e2.0≈7.39, e 1.0 ≈ 2.72 and e0.1≈1.10)
A) 0.67
B) 0.50
C) 0.76
D) 0.89
SRM Institute of Science and Technology Mode of Exam
College of Engineering and Technology
OFFLINE
School of Computing
DEPARTMENT OF COMPUTATIONAL INTELLIGENCE
SRM Nagar, Kattankulathur – 603203, Chengalpattu District, Tamilnadu
Academic Year: 2024-2025 (ODD/EVEN) D
Test: CLAT-1 Date: 27-11-2024
Course Code & Title: 21CSE326T - Artificial Neural Network Duration: 100 Minutes
th
Year & Sem: III & 5 Max. Marks: 50
Part – B
(2 x 5 = 10 Marks)
Instructions: Answer ALL
11 Describe the process by which GNG adapts its network structure over time. How does the 5 2 3 1 1.6.1
GNG algorithm manage the addition and removal of nodes, and why are these mechanisms
important for accurate data representation?
12 Explain the process of selecting a network architecture for a specific problem. 5 2 4 1 1.2.1
Part – C
(3 x 10 = 30 Marks)
Instructions: Answer ANY Three
13 Analyze the steps taken for Stochastic Gradient Descent Algorithm and provide the need for 10 2 3 1 1.2.1
Batch Gradient Descent and Compare Stochastic Gradient Descent with Batch Gradient
Descent in with various parameters in real world application.
14 A neural network model is being used to classify plant species based on leaf images. During 10 3 4 1 1.2.1
training, you notice that certain layers in the network begin to show neuron specialization,
where some neurons become highly responsive to specific leaf textures, while others barely
activate. This uneven activation is causing a drop in generalization performance across
various plant species. Explain how the backpropagation process and weight adjustments
may lead to this specialization. Propose modifications to the architecture or training strategy
that could reduce this imbalance and promote more generalized feature extraction across all
neurons.
15 A neural network is being trained to predict stock prices based on historical price data, 10 2 5 1 1.2.1
social media sentiment, and market indicators. The model performs well initially, but as
new market trends emerge (e.g., driven by unexpected economic events), its accuracy
deteriorates. Describe how the network could be updated in real time to improve its
adaptability without losing learned patterns. Discuss the benefits and potential challenges of
implementing concepts like transfer learning or fine-tuning to handle new, evolving data in
a time-sensitive environment.
16 A neural network model is implemented to detect early signs of diabetes based on medical 10 3 5 1 1.2.1
images and patient records. As the model is deployed in various hospitals, it encounters data
from different imaging devices and varying patient demographics, leading to
misclassifications due to these discrepancies. Analyze how the network’s robustness could
be affected by these variations in data and propose strategies, such as data augmentation or
domain adaptation techniques, to improve the model’s accuracy across diverse data sources
and patient profiles.
*Performance Indicators are available separately for Computer Science and Engineering in AICTE
examination reforms policy.
Course Outcome (CO) and Bloom’s level (BL) Coverage in Questions
CO COVERAGE (%)
35
30
25
20
15
10
5
0
CO1 CO2 CO3 CO4 CO5
Approved by the Audit Professor/Course Coordinator
