Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
DEPARTMENT OF ARTIFICIAL INTELLIGENCE&DATA SCIENCE
Sub Code / Sub Name : AD3501/DEEP LEARNING
Course Outcomes:
At the end of the course, the student should be able to:
1: Explain the basics in deep neural networks.
2: Apply Convolution Neural Network for image processing.
3: Apply Recurrent Neural Network and its variants for text analysis.
4: Apply model evaluation for various applications.
5: Apply autoencoders and generative models for suitable applications
Unit I: Deep Networks Basics
PART A
S.No Question CO BL
1. List any uses of Linear Algebra. CO1 BL2
2. What is meant by scalar? CO1 BL1
3. What is meant by Vector? CO1 BL1
4. Define Matrices CO1 BL1
5. What is meant by Tensors? CO1 BL1
6. What is meant by Probability Distribution? CO1 BL2
7. What is an optimizer in machine learning? CO1 BL2
8. Write any four applications of Gradient descent optimization. CO1 BL3
9. What is Stochastic Gradient Descent? CO1 BL2
10. What is meant by capacity in deep learning? CO1 BL2
11. What is meant by representational and effective capacity? CO1 BL2
12. What is a hyperparameter in machine learning? CO1 BL2
13. Explain the concept of cross – validation. CO1 BL4
14. Explain the concept of K-Fold cross – validation. CO1 BL2
15. Explain about Estimators, bias and variance. CO1 BL2
16. What is meant by curse of dimensionality? CO1 BL2
17. What is Feed Forward neural network? CO1 BL2
18. Analyze the layers present in a Feed Forward Neural Network and their
CO1 BL4
roles.
19. Identify applications of Feed Forward Neural Networks in real-world
CO1 BL3
problems.
20. Compare Machine Learning and Deep Learning with examples. CO1 BL3
Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
PART B
S.No Question CO BL
1. a) Discuss the bias variance trade off (7) CO1 BL1
b) Discuss overfitting and under fitting (6)
2. i. What is Deep Learning? (5) CO1 BL2
ii. Describe how different parts of an Artificial Intelligence system
relate to each other within different AI disciplines in detail with
diagram. (8)
3. i. Discussabout scalars.(7) CO1 BL2
ii. Give detail description of vectors. (6)
4. i. Give the Difference between deep learning and machine learning.(7) CO1 BL2
ii. Give the various concepts of probability. (6)
5. Analyze the following in details. CO1 BL4
i. Vectors. (6)
ii. Matrices. (7)
6. Assess the following. CO1 BL2
i. Expectation. (5)
ii. Variance. (4)
iii.Covariance . (4)
7. Analyze the Stochastic Gradient Descent (SGD) algorithm. Explain how it CO1 BL4
optimizes a model's parameters and discuss its advantages and disadvantages
compared to other optimization algorithms
8. i. Illustrate the importance of principal components analysis. (6) CO1 BL2
ii. Explain support vector machines in detail. (7)
9. Illustrate the importance of Principal Components Analysis (PCA) by CO1 BL3
applying its concepts to a problem of dimensionality reduction.
10. Illustrate the operational concepts of a Deep Feedforward Network using a CO1 BL3
diagram and a detailed step-by-step description.
Unit II: Convolutional Neural Networks
PART A
S.No Question CO BL
1. Define convolution. CO2 BL1
2. What is meant by sparse interaction in CNN? CO2 BL2
3. What is Meant by direct and indirect penalty? CO2 BL2
4. What is meant by parameter sharing? CO2 BL2
5. What are the uses of parameter sharing? CO2 BL2
6. What is meant by Translational equivariance? CO2 BL2
7. What is meant by pooling? CO2 BL2
8. What are the types of pooling? CO2 BL1
9. What are the disadvantages of pooling layer? CO2 BL2
10. What is tiled convolution? CO2 BL2
Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
11. Illustrate the process of transposed convolution to upscale an image. CO2 BL3
12. Analyze the trade-offs between using dilated convolution and standard CO2 BL4
convolution in a deep learning model for image segmentation.
13. State some advantages, disadvantages of dilated convolution. CO2 BL2
14. Analyze the impact of using non-linear activation functions on a neural CO2 BL4
network's ability to learn complex patterns.
15. Using a practical example, apply the concept of a Loss Function to train a CO2 BL3
neural network.
16. What are the types of Loss Function? CO2 BL1
17. Why do we use convolutions for images instead of using fully connected CO2 BL2
layers?
18. Why is it important to introduce non linearities in a neural network? CO2 BL2
19. What do you mean by Convolutional neural network? CO2 BL2
20. What are the different types of pooling? CO2 BL1
PART B
S.No Question CO BL
1. Write an example function for Convolution operation and explain in detail. CO2 BL2
2. Construct an architecture that show complex layer terminology and Simple CO2 BL3
layer terminology in convolutional neural network.
3. Discuss local connections, convolution and full connections with diagram? CO2 BL2
4. Develop a table with examples of different formats of data that can be used CO2 BL3
with convolutional networks.
5. Describe in detail about the following. CO2 BL2
Parameter Sharing. (7)
Equivariant representation. (6)
6. Analyze the locally connected layers, tiled convolution and standard CO2 BL4
convolution with suitable examples and diagram.
7. . Write short notes Max Pooling. (6) CO2 BL2
Explain Pooling with down sampling. (7)
8. Illustrate the concept of unshared convolution with a suitable example. CO2 BL3
9. I.Analyze the three key properties of the V1 visual cortex that a CO2 BL4
convolutional network layer is designed to capture. (6)
II.Prove the working learned invariances with necessary example and
diagram. (7)
10. Discuss parameter sharing. CO2 BL1
Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
Unit III: Recurrent Neural Networks
PART A
S.No Question CO BL
1. Recall unfolding in RNN. CO3 BL1
2. What is gradient descent in Deep learning CO3 BL2
3. What are the advantages of Gradient descent? CO3 BL2
4. What is the difference between LSTM and bidirectional? CO3 BL2
5. Illustrate the limitation of bidirectional RNN CO3 BL2
6. Define deep recurrent network. CO3 BL1
7. What is difference between CNN and RNN? CO3 BL2
8. Define Recursive neural network. CO3 BL1
9. Illustrate what happens when a learning rate that is too large is selected in a CO3 BL3
gradient descent algorithm.
10. What are leaky units? CO3 BL2
11. How many gates are in LSTM architecture? CO3 BL2
12. Demonstrate how skip connections can be used to improve the training of a CO3 BL3
deep neural network.
13. Analyze the vanishing gradient problem and explain why it is a significant CO3 BL4
issue for training deep RNNs.
14. What is the use of leaky ReLU function? CO3 BL2
15. Apply the concepts of an LSTM network to solve a long-term dependency CO3 BL3
problem in a time series.
16. Write the types of RNN. CO3 BL1
17. How many dimensions must the inputs of an RNN layer have? CO3 BL2
18. Compare Feed forward neural networks and Recurrent Neural Networks. CO3 BL2
19. What is the difference between stateful RNN vs Stateless RNN? CO3 BL2
20. Analyze the trade-offs in using an Encoder-Decoder RNN versus a plain CO3 BL4
sequence-to-sequence RNN for machine translation tasks.
PART B
S.No Question CO BL
1. Describe Unfolding CO3 BL3
Computational Graphs. (6)
Explain Bidirectional RNNs. (7)
2. Describe the following. CO3 BL3
Networks with Output Recurrence. (6)
Teacher Forcing in Recurrent Neural Networks. (7)
3. Illustrate Encoder-Decoder sequence-to-sequence Architecture. CO3 BL3
Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
4. Explain Leaky Units and Other Strategies for Multiple Time Scales. CO3 BL2
5. Compute the gradient in a Recurrent Neural Network. CO3 BL3
6. Analyze the different optimization strategies for dealing with long-term CO3 BL4
dependencies and explain why they are necessary.
7. Discuss Recurrent Neural Networks in detail. CO3 BL1
8. Illustrate Clipping Gradients. (7) CO3 BL3
Illustrate Regularizing to Encourage Information Flow. ( 6)
9. Describe the following. CO3 BL2
Long Short-Term Memory. (7)
Other Gated RNNs. (6)
10. Analyze the Echo State Networks and compare their advantages and CO3 BL4
disadvantages with other types of recurrent networks.
Unit IV: Model Evaluation
PART A
S.No Question CO BL
1. What is meant by performance metrics? CO4 BL1
2. Recall Confusion matrix. CO4 BL1
3. Illustrate how a Confusion Matrix can be used to calculate a model's CO4 BL3
precision and recall.
4. What are the performance metrics used in Regression? CO4 BL2
5. What is meant by meant absolute error? CO4 BL2
6. What is R Squared score? CO4 BL2
7. Define Base Line Model. CO4 BL1
8. Apply the concept of a baseline model to a specific classification problem. CO4 BL3
9. What are the benefits of baseline model? CO4 BL2
10. What are the different ways to create a baseline? CO4 BL2
11. What is meant by Manual hyperparameter tuning? CO4 BL2
12. Demonstrate how to use Grid Search to find the optimal hyperparameters for CO4 BL3
a machine learning model.
13. What is Random search? CO4 BL2
14. Differentiate between Grid and Random search. CO4 BL2
15. State any five debugging strategies. CO4 BL2
16. What is meant by Automatic Hyperparameter tuning? CO4 BL2
17. Compare manual hyperparameter tuning with automatic hyperparameter CO4 BL2
tuning in machine learning.
Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
18. Analyze the trade-offs between Grid Search and Random Search for CO4 BL4
hyperparameter optimization, considering efficiency and effectiveness
19. Analyze why the F1-Score is often a more informative performance metric CO4 BL4
than accuracy, especially in the context of imbalanced datasets.
20. Define the Hyperparameter to tune. CO4 BL1
PART B
S.No Question CO BL
1. Discuss the various performance metrics to evaluate a Deep learning Model CO4 BL2
it’s types
2. Using a practical example, illustrate how to use a Naive Classifier as a CO4 BL3
baseline model.
3. Compare the need of Base line Models and benefits of Base line Model CO4 BL2
4. What are hyper parameters? Discuss the steps to perform a hyperparameter CO4 BL1
tuning
5. Analyze the trade-offs between Grid Search and Random Search for CO4 BL4
hyperparameter optimization
6. Illustrate the process of Grid Search as an optimization technique for a CO4 BL3
machine learning model.
7. What are all the Debugging Techniques in Deep learning and Machine CO4 BL2
Learning
8. Illustrate the Applications of Debugging Strategies CO4 BL3
9. Discuss about some Debugging Strategies when a Machine learning systems CO4
performs poorly
10. Analyze why a model's performance on a test set might be different from its CO4 BL4
performance in a real-world scenario, and suggest debugging techniques to
address this issue.
Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
Unit V: Autoencoders and Generative Models
PART A
S.No Question CO BL
1. Define Autoencoders. CO5 BL1
2. Comment on Regularized autoencoders. CO5 BL2
3. What is Stochastic encoder? CO5 BL2
4. List the advantages of Stochastic encoders CO5 BL2
5. Discuss variational autoencoders CO5 BL2
6. Comment on Generative adversarial networks. CO5 BL2
7. Why GAN is called adversarial Networks. CO5 BL2
8. Apply the concept of a Variational Autoencoder to a data generation task. CO5 BL3
9. Analyze the trade-offs between a Variational Autoencoder and a Generative CO5 BL4
Adversarial Network for generating new data.
10. Recall regularized neural network. CO5 BL1
11. What do undercomplete auto encoders have? CO5 BL2
12. What are the applications of Contractive autoencoder? CO5 BL1
13. Demonstrate how a denoising autoencoder can be used to remove noise from CO5 BL3
an image.
14. Analyze the Difference between autoencoder and sparse autoencoder? CO5 BL4
15. How many types of autoencoders are there? CO5 BL2
16. List the types of autoencoder. CO5 BL1
17. What are deep autoencoders? CO5 BL2
18. Why are generative adversarial networks (GANs) so popular? CO5 BL2
19. Where are auto encoders used? CO5 BL2
20. Illustrate the architecture of a GAN and describe how its components interact CO5 BL3
during training.
Approved by AICTE, Affiliated to Anna University
Accredited by NAAC | Recognized by UGC with 2(F)
PART B
S.No Question CO BL
1. Explain Sparse Coding CO5 BL2
2. Discuss Autoencoders. CO5 BL1
3. Illustrate on the following: CO5 BL3
Denoising Autoencoders. (6)
Sparse Autoencoders. (7)
4. Analyze the differences between a Sparse Autoencoder and a Predictive Sparse CO5 BL4
Decomposition model.
5. Compare Structured Probabilistic Model and Unstructured Modeling CO5 BL2
6. Summarize usage of various Graphs to describe Model Structure. CO5 BL3
7. Develop an example distribution equation for energy-based model and explain CO5 BL3
in detail.
8. Explain Generative Adversarial Networks (GAN) CO5 BL2
9. Analyze the advantages and disadvantages of transfer learning and domain CO5 BL4
adaptation.
10. Discuss Manifold interpretation of PCA. (13) CO5 BL2
Bloom’s Taxonomy Level (BL)
1. Remember 2. Understand 3. Apply 4. Analyze 5. Evaluate 6. Create
Summary
Percentage of questions from different Cognitive levels
Unit No Total %
1 2 3 4 5 6
1 17% 58% 16% 9% - - 100
2 17% 58% 16% 9% - - 100
3 17% 58% 16% 9% - - 100
4 17% 58% 16% 9% - - 100
5 17% 58% 16% 9% - - 100
Average 100
17% 58% 16% 9% - -
%
