DEEP LEARNING
UNIT - III: CONVOLUTIONAL NETWORKS
PART – A
An essential feature of any convolutional network implementation is
1
the ability to implicitly zero-pad the input Justify
The output layer of convolutional network is usually relatively
2
inexpensive to learning layer. Justify.
3 What is convolutional networks?
4 Create a chart that demonstrates convolution with a stride.
5 How pooling handles inputs of varying size?
6 What is meant by convolution?
List three important ideas that help to improve a machine learning
7
system.
8 What is unshared convolution?
9 Define primary visual cortex.
10 How to reduce the cost of convolutional network training?
11 Simulate the idea behind reverse correlation.
12 Discuss about parameter sharing in neural network.
Give three properties of V1 that a convolutional network layer is
13
designed to capture.
14 Explain feature map.
Explain how a convolutional layer have a property called
15
equivariance to translation?
16 List three stages of a convolutional network
List out various formats of data that can be used with convolutional
17
networks.
18 Illustrate pooling stage in convolutional network.
Differentiate complex layer terminology and simple layer
19
terminology in convolutional network.
20 Write Short note on ResNet, AlexNet, DenseNet, PixelNet
Show three basic strategies for obtaining convolution kernels
21
without supervised learning
22 Give example for convolution
23 Illustrate reverse correlation
24 Explain complex layer terminology.
25 Examine equivariance to translation
26 Explain Transfer Learning
B
1 Write an example function for Convolution operation and explain in
detail.
Explain the following with suitable diagram. (i) Sparse interactions
2
(ii) Parameter Sharing
3 DescribePooling with suitable example.
Write an expression for Unshared convolution with explanation
4
and explain Tiled convolution
.
5 Discuss in detail the variants of the Basic Convolution Function.
Constructanarchitecturethat show complex layer terminology and
6
Simple layer terminology in convolutional neural network
.
Discuss local connections, convolution and full connections with
7
diagram?
Develop a table with examples of different formats of data that
8
can be used with convolutional networks.
9 Describe in detail about the following.
i. Parameter Sharing.
ii. Equivariant representation.
Differentiatelocally connected layers, tiled convolution and
10
standard convolution with suitable examples and diagram.
11 i. Write short notes Max Pooling.
ii. Explain Pooling with downsampling.
12 Explain random or Unsupervised Features.
13 Illustrateunshared convolutionwith suitable examples.
i. Show three properties of V1 that a convolutional network layer
14
is designed to capture.
ii. Provethe working learned invariances with necessary
example and diagram.
15 . Discuss parameter sharing
16 IllustrateEquivariant representation.
Evaluate the working learned invariances with necessary example and
17 .
diagram
18. Constructa graphical demonstration for sparse connectivity and
explain it in detail
.
19. Create a graphical demonstration for parameter sharing and explain
it in in detail
20 Evaluate variants of the basic convolution function.
21. Construct a convolutional network to demonstrate the effect of zero padding
on network size.
22 Explain Neuro scientific basis for Convolutional Networks
23. Create a table with examples of different formats of data that can be used
with convolutional networks.
UNIT - IV
.
PART –
A
1 What is Recurrent Neural Networks?
2 What is Encoder?
Give the blocks of decomposition of computation of most Recurrent
3
Neural
Networks.
4 What is Bidirectional Recurrent Neural Networks?
5 Give the advantage of recursive nets over recurrent nets.
6 What is decoder?
7 Describe Recursive Neural Networks.
8 Predict the concept of gated RNNs.
9 Compare echo state network and liquid state machines.
Distinguish content based addressing and location based addressing
10
in
memory networks.
11 Classify the different strategies for Multiple Time Scales.
12 Developa block diagram for LSTM.
13 Illustrateimportant design patterns for recurrent neural networks.
14 Summarize about echo state networks.
Point out the advantage of introducing depth in Deep recurrent
15
Networks.
16 Compare gradient descent with and without gradient clipping using
diagram.
17 Justify the major advantages of unfolding process in computational
graphs.
18 Illustrate block diagram of LSTM recurrent network “cell”.
19 What are leaky units?
20 Develop a schematic diagram of a network with an explicit memory.
21 Give a block diagram for Long Short Term Memory.
22 Illustrate echo state networks.
23 Explain liquid state machines.
24 Assess explicit memory.
PART – B
1 i. Describe Unfolding Computational Graphs.
ii. Explain Bidirectional RNNs.
2 Describe the following.
i. Teacher Forcing in Recurrent Neural Networks.
ii. Networks with Output Recurrence.
3 i. Describe Echo State Networks.
ii. Explain challenge of Long-Term Dependencies
4 Discuss Recurrent Neural Networks in detail.
5 Describe Deep Recurrent Networks in detail
6 Illustrate Encoder-Decoder sequence-to-sequence Architecture.
Explain Leaky Units and Other Strategies for Multiple Time
7
Scales.
8 Point out various features of Echo state networks.
9 Explain Optimization for Long-Term Dependencies.
10 Compute the gradient in a Recurrent Neural Network.
11 i. Illustrate Clipping Gradients.
ii. Illustrate Regularizing to Encourage Information Flow.
12 Describe the following.
i. Long Short-Term Memory.
ii. Other Gated RNNs.
Explain in detail about the
13
following.
i. Adding Skip Connections through Time.
ii. Leaky Units and a Spectrum of Different Time Scales .
14 Describe Explicit memory.
15 Discuss Echo State Networks.
16 Illustrate Bidirectional RNNs.
17 Explain challenge of Long-Term Dependencies.
PART – C
Develop an example for Unfolding Computational Graphs and
1
describe the
major advantages of unfolding process.
Explain how to compute the gradient in a Recurrent Neural
2
Network.
3 Explain a modeling sequences Conditioned on Context with RNNs.
Prepare an example of Encoder- Decoder or sequence-to-sequence
4
RNN
architecture.
5 Explain various Gated RNNs.
