ai

Question 1

Given a convolution layer with input channels 3, output channel 64, kernel size 4x4 and stride 2, dilation 3, padding 1, what are the parameter size of this convolution layer?

Answer 1

3x64x4x4

Question 2

In Pytorch (import torch.nn as nn), which of the following layer downsamples the input size into half?

Answer 2

nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=2, padding=1, dilation=1))

Question 3

Which following statement is True about convolution layer?

Answer 3

Convolution layer is linear and it is often used along with activation function.

Question 4

In the design of auto-encoder, the encoder and decoder should follow the exact same structure.

Answer 4

FALSE

Question 5

All regularizations (e.g., L1 norm, L2 norm) penalize larger parameters.

Answer 5

TRUE

Question 6

When updating parameters using gradient descent, which way of calculating loss works better (ie. a better trade-off between efficiency and robustness)?

Answer 6

calculate loss for a mini-batch of data examples in every iteration

Question 7

MaxPooling preserves detected features, and downsamples feature map (image)

Answer 7

TRUE

Question 8

What is the size of receptive field for two stacking dilated convolution layers with kernel size 3x3, stride 1, and dilation 2?

Answer 8

9x9

Question 9

In CNN, two conv layers cannot be connected directly, we must use a pooling layer in the middle.

Answer 9

FALSE

Question 10

In the design of CNN, fully connected layer usually contains much more parameters than conv layers.

Answer 10

TRUE

Question 11

What is the purpose of the ReLU activation function in a CNN?

Answer 11

To introduce non-linearity

Question 12

What is the main advantage of using dropout in a CNN?

Answer 12

Preventing overfitting

Question 13

In the mini-batch SGD training, an important practice is to shuffle the training data before every epoch. Why?

Answer 13

It helps the training converge fast and prevents bias.

Question 14

Logistic Regression is widely used to solve a classification problem with predicting probabilities of discrete (or categorical) values.

Answer 14

TRUE

Question 15

Which of the following statements is true about activation functions in the context of neural networks and backpropagation?

Answer 15

Activation functions like ReLU (Rectified Linear Unit) introduce non-linear properties to the model, allowing it to learn complex patterns.

Question 16

Which case is overfitting?

Answer 16

Training error is low, but testing error is high

Question 17

What approach could be used to handle overfitting?

Answer 17

Use regularization

Question 18

Besides penalizing larger parameters, which regularization makes parameters more sparse?

Answer 18

L1 norm

Question 19

In Backpropagation, which claim is true?

Answer 19

Backward-pass uses the information preserved in forward-pass to calculate gradients

Question 20

As an activation function, tanh avoids the vanishing gradient problem.

Answer 20

FALSE

Question 21

As an activation function, ReLU solves the vanishing gradient problem.

Answer 21

TRUE

Question 22

About SGD optimization, which is not correct?

Answer 22

Randomly initialize the parameters will affect the performance.

Question 23

In reinforcement learning, what is the benefit of using network instead of lookup table?

Answer 23

Generalization

Question 24

Which way do we usually use to train an autoencoder model?

Answer 24

We usually train the encoder model and the decoder model together

Question 25

Which claim is true about attention and self-attention? In the sequence-to-sequence model, at different steps, "attention" lets the model "focus" on different parts of the input. Self-attention is usually used to model dependencies between different parts of one sequence (e.g., words in one sentence). Both the above claims. None of the above claims.

Answer 25

Both the above claims.

Question 26

What's the major purpose of multi-head attention?

Answer 26

Catching multiple relationships between words of the input sequence

Question 27

In the transformer neural network architecture, the encoder blocks usually use the identical neural network structure.

Answer 27

TRUE

Question 28

In the transformer neural network architecture, the output of the final encoder block will go to _____.

Answer 28

Every decoder block

Question 29

In the autoregressive model, the output variable at the current step depends on only the hidden states at all previous steps.

Answer 29

FALSE

Question 30

In Transformer, how does the decoder use the information (features) from the encoder?

Answer 30

Cross-attention (mixed attentions from multiple inputs)

Question 31

In the policy gradient approach for reinforcement learning, the reward R(?^n ) is considered based on

Answer 31

The cumulative reward in every entire trajectory

Question 32

In the two major approaches of reinforcement learning, i.e., policy gradient and Q-learning, which is usually more sample-efficient?

Answer 32

Q-learning

Question 33

In Q-Learning, we can use either Q-table or a neural network to predict a Q-value for a pair of (state, action). Which one is more scalable?

Answer 33

Neural network

Question 34

In the two major approaches of reinforcement learning, i.e., policy gradient and Q-learning, which one uses randomly sampled transitions instead of the entire trajectories?

Answer 34

Q-learning

Question 35

In the two major approaches of reinforcement learning, i.e., policy gradient and Q-learning, which one is on-policy training?

Answer 35

Policy gradient

Question 36

For Discrete-event modeling, what approach do we often use?

Answer 36

Q-learning

Question 37

What approach do we often use for Discrete-event modeling?

Answer 37

Q-learning

Question 38

In the information theory, which event includes more information?

Answer 38

Low probability event occurs

Question 39

What is the correct statement about activation functions in neural networks and backpropagation?

Answer 39

Activation functions introduce non-linear properties to the model, allowing it to learn complex patterns.

Question 40

What is a common sign of overfitting?

Answer 40

Low training error, high testing error

Question 41

What technique can help to mitigate overfitting?

Answer 41

Regularization

Question 42

Do all regularizations penalize larger parameters?

Answer 42

TRUE

Question 43

What is the primary purpose of using dropout in a CNN?

Answer 43

Preventing overfitting

Question 44

What is the purpose of the ReLU activation function in a CNN?

Answer 44

To introduce non-linearity

Question 45

Which statement about convolution layers is true?

Answer 45

Convolution layers are linear and often used with activation functions.

Question 46

What is the parameter size of a convolution layer with input channels 3, output channels 64, kernel size 4x4, stride 2, dilation 3, and padding 1?

Answer 46

3x64x4x4

Question 47

Which layer in PyTorch downsamples the input size into half?

Answer 47

nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=2, padding=1, dilation=1))

Question 48

Does MaxPooling preserve detected features and downsample the feature map?

Answer 48

TRUE

Question 49

Do two convolution layers in a CNN always need a pooling layer in between them?

Answer 49

FALSE

Question 50

Does a fully connected layer in CNN usually contain more parameters than convolution layers?

Answer 50

TRUE

Question 51

What is the primary purpose of using dropout in a CNN?

Answer 51

To prevent overfitting

Question 52

What is the main advantage of using dropout in a CNN?

Answer 52

Preventing overfitting

Question 53

What is the purpose of the ReLU activation function in a CNN?

Answer 53

To introduce non-linearity

Question 54

Which statement about convolution layers is true?

Answer 54

Convolution layers are linear and often used with activation functions.

Question 55

What is the parameter size of a convolution layer with input channels 3, output channels 64, kernel size 4x4, stride 2, dilation 3, and padding 1?

Answer 55

3x64x4x4

Question 56

Which layer in PyTorch downsamples the input size into half?

Answer 56

nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=2, padding=1, dilation=1))

Question 57

Does MaxPooling preserve detected features and downsample the feature map?

Answer 57

TRUE

Question 58

Do two convolution layers in a CNN always need a pooling layer in between them?

Answer 58

FALSE

Question 59

Does a fully connected layer in CNN usually contain more parameters than convolution layers?

Answer 59

TRUE

Question 60

What is the primary purpose of using dropout in a CNN?

Answer 60

To prevent overfitting

Question 61

What is the main advantage of using dropout in a CNN?

Answer 61

Preventing overfitting

Question 62

What is the purpose of the ReLU activation function in a CNN?

Answer 62

To introduce non-linearity

Question 63

Which statement about convolution layers is true?

Answer 63

Convolution layers are linear and often used with activation functions.

Question 64

What is the parameter size of a convolution layer with input channels 3, output channels 64, kernel size 4x4, stride 2, dilation 3, and padding 1?

Answer 64

3x64x4x4

Question 65

Which layer in PyTorch downsamples the input size into half?

Answer 65

nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=2, padding=1, dilation=1))

Question 66

Does MaxPooling preserve detected features and downsample the feature map?

Answer 66

TRUE

Question 67

Do two convolution layers in a CNN always need a pooling layer in between them?

Answer 67

FALSE

Question 68

Does a fully connected layer in CNN usually contain more parameters than convolution layers?

Answer 68

TRUE

Question 69

What is the primary purpose of using dropout in a CNN?

Answer 69

To prevent overfitting

Question 70

What is the main advantage of using dropout in a CNN?

Answer 70

Preventing overfitting

Question 71

What is the purpose of the ReLU activation function in a CNN?

Answer 71

To introduce non-linearity

Question 72

Which statement about convolution layers is true?

Answer 72

Convolution layers are linear and often used with activation functions.

Question 73

What is the parameter size of a convolution layer with input channels 3, output channels 64, kernel size 4x4, stride 2, dilation 3, and padding 1?

Answer 73

3x64x4x4

Question 74

Which layer in PyTorch downsamples the input size into half?

Answer 74

nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=2, padding=1, dilation=1))

Question 75

Does MaxPooling preserve detected features and downsample the feature map?

Answer 75

TRUE

Question 76

Do two convolution layers in a CNN always need a pooling layer in between them?

Answer 76

FALSE

Question 77

Does a fully connected layer in CNN usually contain more parameters than convolution layers?

Answer 77

TRUE

Question 78

What is the primary purpose of using dropout in a CNN?

Answer 78

To prevent overfitting

Question 79

What is the main advantage of using dropout in a CNN?

Answer 79

Preventing overfitting

Question 80

What is the purpose of the ReLU activation function in a CNN?

Answer 80

To introduce non-linearity

Question 81

Which statement about convolution layers is true?

Answer 81

Convolution layers are linear and often used with activation functions.

Question 82

What is the parameter size of a convolution layer with input channels 3, output channels 64, kernel size 4x4, stride 2, dilation 3, and padding 1?

Answer 82

3x64x4x4

Question 83

Which layer in PyTorch downsamples the input size into half?

Answer 83

nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=2, padding=1, dilation=1))

Question 84

Does MaxPooling preserve detected features and downsample the feature map?

Answer 84

TRUE

Question 85

Do two convolution layers in a CNN always need a pooling layer in between them?

Answer 85

FALSE

Question 86

Does a fully connected layer in CNN usually contain more parameters than convolution layers?

Answer 86

TRUE

Question 87

What is the primary purpose of using dropout in a CNN?

Answer 87

To prevent overfitting

Question 88

What is the main advantage of using dropout in a CNN?

Answer 88

Preventing overfitting

Question 89

What is the purpose of the ReLU activation function in a CNN?

Answer 89

To introduce non-linearity

Question 90

Which statement about convolution layers is true?

Answer 90

Convolution layers are linear and often used with activation functions.

Question 91

What is the parameter size of a convolution layer with input channels 3, output channels 64, kernel size 4x4, stride 2, dilation 3, and padding 1?

Answer 91

3x64x4x4

Question 92

Which layer in PyTorch downsamples the input size into half?

Answer 92

nn.Conv2d(in_channels