B.E.

(Computer Engineering) | DEEP LEARNING (2019 Pattern) | (Semester - VIII) (410251) | May-June 2023 PYQ Solutions

Q1) a) Explain Pooling Layer with its need and different types
[6]
• Reduces spatial dimensions c) Explain Unfolding computational graphs with example [5] ∟ Generator creates fake images ∟ Evaluates and improves
• Helps manage complexity • Unfolding: Repeat neural network structure ∟ Discriminator identifies real images • Bellman Equations:
• Prevents overfitting ∟ Example: RNN with 3-time steps • GANs aim for realism ∟ Fundamental in DP
• Types: • Creates multiple time steps ∟ Learn to generate authentic images ∟ Express value function relationships
∟ Max Pooling • Each step linked to previous ∟ Discriminator improves over time • DP optimizes reward policies:
∟ Average Pooling • Unrolls recurrent connections Q6) a) Differentiate generative and discriminative models in GAN ∟ Balances exploration and exploitation
∟ Global Pooling • Facilitates understanding network behaviour (Generative Adversarial Network) [6] ∟ Solves complex RL problems efficiently
b) Draw and explain CNN (Convolution Neural Network) • Useful for training and analysis c) Explain Simple reinforcement learning for TicTacToe [5]
architecture in detail [6] Q4) a) What are types of RNN (Recurrent Neural Network)? How • Generative models create new data • Agent learns to play Tic-Tac-Toe:
• Input Layer: Initial data to train RNN explain in brief [6] ∟ Create from learned distribution ∟ Environment: Tic-Tac-Toe board
• Convolution Layer: Feature extraction • RNN Types: ∟ No direct classification involved ∟ Agent: Learns optimal moves
• Activation Layer: Adds nonlinearity ∟ Vanilla RNN: Basic recurrent network • Discriminative models classify data • States: Board configurations
• Pooling Layer: Reduces spatial dimensions ∟ LSTM: Long Short-Term Memory ∟ Distinguish between classes ∟ Actions: Placing X or O
• Fully Connected Layer: Makes predictions ∟ GRU: Gated Recurrent Unit ∟ Learn decision boundaries ∟ Rewards: Win, lose, draw
• Output Layer: Final prediction • Training RNN: b) What are applications of GAN (Generative Adversarial • Q-table stores state-action values:
∟ BPTT: Backpropagation Through Time Network)? Explain any four in detail [6] ∟ Agent updates Q-values
c) Explain ReLU Layer in detail What are the advantages of ReLU
∟ Gradient weight updates: Adjusts weights • Image generation: ∟ Uses exploration vs exploitation
over Sigmoid? [6]
∟ Creates realistic images • Learning through trial and error:
• Rectified Linear Unit (ReLU): Activation function b) Explain Encoder-Decoder Sequence to Sequence architecture
∟ Used in art, design ∟ Adjusts strategy based on outcomes
• ReLU sets negative values to zero with its application [6]
• Encoder: Converts input to context • Style transfer: ∟ Reinforces successful moves
• Advantages over Sigmoid:
∟ Captures information ∟ Apply styles to images Q8) a) Write Short Note on Q-Learning and Deep Q-Networks [6]
∟ Faster computation
∟ Avoids vanishing gradient problem • Context Vector: Condensed representation ∟ Alter art styles • Q Learning:
∟ Contains essential information • Super-resolution: ∟ Reinforcement learning algorithm
Q2) a) Explain all the features of pooling layer [6]
• Decoder: Generates output from context ∟ Enhance image resolution ∟ Learns value for actions
• Reduces spatial dimensions
∟ Reconstructs input from context ∟ Improve image quality • Uses value iteration
• Helps in feature selection
• Application: Machine translation • Data augmentation: ∟ Updates Q-values iteratively
• Controls overfitting
∟ Translates text between languages ∟ Generate synthetic data ∟ Finds optimal policy
• Increases computational efficiency
∟ Enhance training datasets • Deep Q-Networks (DQN):
• Types: c) Differentiate between Recurrent and Recursive Neural Network
c) Write Short Note on Deep generative model and Deep Belief ∟ Combines Q-learning with NN
∟ Max Pooling [5]
Networks[6] ∟ Learns Q-values
∟ Average Pooling • Recurrent Neural Network (RNN):
• Deep Generative Model: • Uses experience replay
b) Explain Dropout Layer in Convolutional Neural Network [6] ∟ Loops input through time
∟ Generates complex data ∟ Stores experiences in memory
• Randomly drops neurons ∟ Handles sequential data
∟ Uses deep learning techniques ∟ Enhances learning efficiency
• Prevents overfitting ∟ Example: Time series prediction
• Recursive Neural Network (RecNN): • Deep Belief Networks (DBNs): b) What are the challenges of reinforcement learning? Explain any
• Improves generalization four in detail [6]
∟ Processes data recursively ∟ Stack of Restricted Boltzmann Machines
• During training phase ∟ Unsupervised learning • Exploration vs exploitation dilemma
∟ Handles hierarchical data
• Retains model's flexibility • DBNs model hierarchical data: ∟ Balance exploration and exploitation
∟ Example: Parsing sentences
• Enhances model's performance ∟ Learn hierarchical representations • Credit assignment problem
• Key Differences:
c) Explain working of Convolution Layer with its features [6] ∟ RNN loops over time ∟ Useful for feature learning ∟ Assign credit accurately
• Applies filters to input data ∟ RecNN processes recursively • Deep generative models: • Non-stationarity of environments
• Feature extraction ∟ RNN for sequences, RecNN for hierarchies ∟ Generate realistic data ∟ Environments change constantly
• Stride moves filter Q5) a) Explain Boltzmann machine in details [6] ∟ Capture data distribution accurately • Reward design complexities
• Padding adds zeros Q7) a) Explain Markov Decision Process with Markov property [6] ∟ Design effective reward functions
• Nonlinear activation • Markov Decision Process (MDP): • Sample inefficiency
• Boltzmann machine:
• Preserves spatial information ∟ Models’ sequential decision-making ∟ Needs many training samples
∟ Stochastic neural network
Q3) a) What is RNN? What is need of RNN? Explain in brief about ∟ Comprises states, actions, rewards • High computational requirements
∟ Model’s joint probability distribution
working of RNN (Recurrent Neural Network) [6] ∟ Markov Property: ∟ Demands substantial computing power
∟ Uses Gibbs sampling for learning
• RNN: Recurrent Neural Network ∟ Consists of visible and hidden units ▪ Future depends on present c) What is deep reinforcement learning? Explain in detail [5]
• Handles sequential data ∟ Connections have weights ▪ History is not needed • Combines RL with deep learning
• Maintains memory of past ∟ Learning via contrastive divergence • States: Represent situations ∟ Learns complex tasks
• Addresses time dependencies b) Explain GAN (Generative Adversarial Network) architecture • Actions: Available choices ∟ Uses neural networks for decision-making
• Loops information over time with an example [6] • Rewards: Outcomes of actions • Deep RL learns representations
• Suitable for time series analysis b) Explain in detail Dynamic programming algorithms for ∟ Extracts features automatically
• GAN Architecture:
b) How LSTM and Bidirectional LSTM works [6] reinforcement learning [6] ∟ Handles high-dimensional data efficiently
∟ Generator: Creates fake data
• Dynamic Programming (DP) algorithms: • Benefits:
• LSTM: Long Short-Term Memory ∟ Discriminator: Distinguishes real from fake
∟ Value Iteration ∟ Solves complex tasks
∟ Handles long-term dependencies • Example:
∟ Policy Iteration ∟ Learns from raw data
∟ Input, forget, output gates ∟ Generating realistic faces
∟ Cell state maintains memory ∟ Generator learns data distribution • Value Iteration: • Challenges:
∟ Updates value function iteratively ∟ Requires substantial computational resources
• Bidirectional LSTM: ∟ Discriminator distinguishes real vs fake
∟ Converges to optimal policy ∟ May suffer from instability during training
∟ Processes input forward and backward c) Do GANs (Generative Adversarial Network) find real or fake
∟ Captures past and future context images? If yes, explain it in detail [6] • Policy Iteration:
∟ Combines outputs for prediction • GANs generate images from noise ∟ Iteratively improves policy