About this previous year paper
DEEP LEARNING (KCS-078)
B.Tech SEM VII – Complete Solved Question Paper (2022–23)
⏱ Time: 3 Hours | 📊 Marks: 100
SECTION A
Attempt all questions in brief (2 × 10 = 20 marks)
(a) Applications of Machine Learning
Machine Learning is used in image recognition, speech recognition, recommendation systems, fraud detection, medical diagnosis, self-driving cars, spam filtering, and predictive analytics.
(b) Boltzmann Machine
A Boltzmann Machine is a stochastic recurrent neural network that learns probability distributions over inputs. It consists of visible and hidden units connected symmetrically and is mainly used for feature learning and optimization problems.
(c) Can deep learning models be built using only linear regression? Explain
No. Linear regression can only model linear relationships. Deep learning requires non-linear transformations to learn complex patterns. Without non-linear activation functions, deep models collapse into a single linear transformation.
(d) Different layers of Convolutional Neural Network (CNN)
CNN consists of: Input layer
Convolutional layer Activation layer (ReLU)
Pooling layer Fully connected layer
Output layer
(e) Linear models
Linear models assume a linear relationship between input features and output. Examples include linear regression and logistic regression. They are simple, fast, but limited in modeling complex data.
(f) Importance of non-linearities in neural networks
Non-linearities allow neural networks to learn complex patterns and decision boundaries. Without non-linear activation functions, neural networks cannot solve non-linear problems.
(g) Limitations of perceptron
A perceptron can only solve linearly separable problems. It cannot solve non-linear problems such as XOR and lacks hidden layers.
(h) Why use convolutions instead of fully connected layers for images?
Convolutions reduce parameters, preserve spatial information, and exploit local patterns. Fully connected layers are computationally expensive and ignore spatial structure.
(i) Importance of GPUs in deep learning
GPUs enable parallel computation, faster matrix operations, and efficient training of large deep learning models compared to CPUs.
(j) Best algorithm for face detection
Convolutional Neural Networks (CNNs) are the best for face detection. Popular models include Haar-Cascade (traditional) and modern CNN-based models like MTCNN and YOLO.
SECTION B
Attempt any THREE (10 × 3 = 30 marks)
(a) Difference between Deep and Shallow Networks
| Deep Network | Shallow Network |
|---|---|
| Multiple hidden layers | Few or no hidden layers |
| Learns complex features | Learns simple features |
| High accuracy | Limited performance |
| Used in DL tasks | Used in basic ML |
(b) Architecture of Convolutional Neural Network
A CNN consists of stacked convolutional layers followed by activation functions, pooling layers for dimensionality reduction, fully connected layers for classification, and an output layer.
( In exam, draw neat CNN block diagram)
(c) Why CNN is preferred over ANN for image classification
CNNs preserve spatial relationships, require fewer parameters, and automatically extract features, whereas ANNs treat images as flat vectors and are inefficient for large images.
(d) LSTM (Long Short-Term Memory) and applications
LSTM is a special type of RNN designed to handle long-term dependencies using memory cells and gates (input, forget, output).
Applications: Speech recognition
Machine translation Time-series forecasting
Text generation
(e) Image Captioning in Deep Learning
Image captioning combines CNNs for feature extraction and RNN/LSTM for sequence generation. It generates natural language descriptions for images, widely used in accessibility tools.
SECTION C
Q3 (Attempt any one)
(a) Gradient Descent vs Stochastic Gradient Descent
| Gradient Descent | Stochastic Gradient Descent |
|---|---|
| Uses full dataset | Uses one sample |
| Slow for large data | Faster |
| Stable convergence | Noisy but efficient |
(b) GAN (Generative Adversarial Network) and its models
GAN consists of a Generator and Discriminator competing against each other.
Types of GANs: Vanilla GAN
DCGAN CycleGAN
Conditional GAN
Q4 (Attempt any one)
(a) Semi-Supervised Learning
Semi-supervised learning uses both labeled and unlabeled data. It reduces labeling cost and improves performance when labeled data is limited.
(b) Deep Learning: history, applications, and uses
Deep learning evolved from neural networks and gained popularity due to big data and GPUs. It is used in vision, speech, NLP, healthcare, and autonomous systems.
Q5 (Attempt any one)
(a) Backpropagation algorithm
Backpropagation updates weights by computing error gradients using the chain rule.
Steps: Forward pass
Error calculation Backward pass
Weight update
(b) Short notes
i) Deep Reinforcement Learning:
Combines deep learning with reinforcement learning to learn policies.
ii) Autoencoder Architecture:
Used for dimensionality reduction and feature learning.
iii) VGG:
Deep CNN architecture with small filters and high accuracy.
iv) SOA (State of the Art):
Refers to best-performing models at a given time.
'
Q6 (Attempt any one)
(a) PCA vs RNN
| PCA | RNN |
|---|---|
| Dimensionality reduction | Sequential modeling |
| Linear technique | Non-linear |
| No memory | Has memory |
(b) Batch GD vs Stochastic GD
Batch GD uses full dataset per update, while SGD updates weights per sample, making SGD faster and scalable.
Q7 (Attempt any one)
(a) Privacy issues in facial recognition
Facial recognition can lead to surveillance, misuse of personal data, identity theft, and lack of consent when used by private companies.
(b) How AI and Neuroscience drive each other
Neuroscience inspires AI architectures, while AI helps analyze brain data. Together, they advance understanding of intelligence.