9 - Neural Networks

Question 1

What do neural networks attempt to imitate?

Answer 1

The type of non-linear learning that occurs in the networks of neurons found in nature, such as the human brain.

Question 2

What are the main components of a neuron?

Answer 2

Dendrites, cell body, and axon.

Question 3

What is the role of dendrites in a neuron?

Answer 3

To gather inputs from other neurons.

Question 4

What does an artificial neuron model typically consist of?

Answer 4

Inputs (xi), a combination function (such as summation), an activation function, and an output response (y).

Question 5

What is a key benefit of neural networks?

Answer 5

They are robust for noisy, complicated, or nonlinear data.

Question 6

What is a main drawback of neural networks?

Answer 6

They are relatively opaque to human interpretation.

Question 7

What are the three layers typically found in a neural network?

Answer 7

• Input layer
• Hidden layer
• Output layer

Question 8

What is meant by a ‘completely connected’ neural network?

Answer 8

Every node in a given layer is connected to every node in adjoining layers.

Question 9

What is the purpose of weights in a neural network?

Answer 9

To influence the strength of the connection between nodes.

Question 10

What happens if the hidden layer has too many nodes?

Answer 10

It can lead to overfitting.

Question 11

What is the combination function used in neural network nodes?

Answer 11

Usually summation (Σ) to produce a linear combination of inputs and weights.

Question 12

What does the term ‘net’ refer to in a node?

Answer 12

The single scalar value produced by the combination function.

Question 13

What is the most common activation function in neural networks?

Answer 13

The sigmoid function.

Question 14

What does the sigmoid function output range between?

Answer 14

0 and 1.

Question 15

What is the formula for the sigmoid function?

Answer 15

y = 1 / (1 + e^(-x))

Question 16

What is backpropagation in neural networks?

Answer 16

A method for adjusting weights based on prediction error.

Question 17

What does the sum of squared errors (SSE) measure?

Answer 17

How well the output predictions fit the actual target values.

Question 18

What is used to minimize the SSE in neural networks?

Answer 18

Optimization methods, specifically gradient-descent methods.

Question 19

What type of data does the Framingham Heart Study dataset contain?

Answer 19

Information on three variables for 7953 patients, including binary predictor ‘Sex’ and continuous predictor ‘Age’.

Question 20

What are the target variable values in the Framingham Heart Study dataset?

Answer 20

0 = survival and 1 = death.

Question 21

What does the activation function do within a node?

Answer 21

It produces an output value based on the net input.

Question 22

What does the weight in a neural network model represent?

Answer 22

What the model is trying to tell you, analogous to predictor coefficients in regression.

Question 23

Fill in the blank: The combination function produces a _______ of the node inputs and connection weights.

Answer 23

single scalar value

Question 24

True or False: Neural networks allow for looping or cycling within the flow of information.

Answer 24

False

Question 25

What do the predictor coefficients in a regression model indicate?

Answer 25

They indicate the strength and direction of the relationship between predictors and the response variable.

Question 26

What are the weights associated with the hidden layer node H1 and the output node O1?

Answer 26

WH1O1 = -5.8477

Question 27

What does a negative weight from H1 to O1 imply about the relationship with death probability?

Answer 27

It implies that a higher value of H1 lowers the probability of death.

Question 28

How does the Sex predictor influence the hidden layer node H1?

Answer 28

The weight WI1H = 0.6418 indicates that larger values of Sex (females) excite H1 to a higher value.

Question 29

What is the relationship between age and the hidden layer node H1?

Answer 29

The weight WI2H = -3.0784 indicates that higher values of Age lower the value of H1.

Question 30

What is the first step in using neural networks in R with the Framingham_training data set?

Answer 30

Read in the Framingham_training data set as fram_train.

Question 31

What command is used to convert the binary and ordinal variables Death and Sex to factors in R?

Answer 31

fram_train$Death <- as.factor(fram_train$Death) and fram_train$Sex <- as.factor(fram_train$Sex)

Question 32

What is the purpose of min-max standardization on the Age variable?

Answer 32

To scale the Age variable to a range between 0 and 1.

Question 33

What command installs the nnet and NeuralNetTools packages in R?

Answer 33

install.packages('nnet'); install.packages('NeuralNetTools')

Question 34

What is specified by the size parameter in the nnet function?

Answer 34

The number of units in the hidden layer.

Question 35

What is the command to plot the neural network after running the nnet algorithm?

Answer 35

plotnet(nnet01)

Question 36

What is the essential problem for the neural network?

Answer 36

To construct a set of weights that will minimize the error in predictions.

Question 37

What is backpropagation in the context of neural networks?

Answer 37

A process to update weights in the network by propagating the error backward.

Question 38

What does it mean when we say a neural network is completely connected?

Answer 38

Each node in one layer is connected to every node in the next layer.

Question 39

What are the benefits of having more nodes in the hidden layer?

Answer 39

Increased model capacity and ability to capture complex patterns.

Question 40

What is a drawback of using too many nodes in the hidden layer?

Answer 40

Risk of overfitting the model to the training data.

Question 41

Fill in the blank: The sigmoid function combines nearly linear behavior, curvilinear behavior, and nearly constant behavior in _______.

Answer 41

output transformation.

Question 42

What does the output of the nnet() function include?

Answer 42

Weights and structure of the trained neural network.

Question 43

Which baseline model do we compare the neural network model against?

Answer 43

A simple classification model, often logistic regression.

Question 44

True or False: The neural network model should outperform the baseline model in terms of accuracy.

Answer 44

True

Question 45

What is the purpose of constructing a contingency table in model evaluation?

Answer 45

To compare actual and predicted values.

Question 46

What are the criteria to compare different models in the exercises?

Answer 46

* Accuracy * Sensitivity * Specificity

Question 47

What is the first step in preparing the adult_ch6_training data set for neural network modeling?

Answer 47

Create a binary variable CapGainsLossesPositive.

Question 48

What should be done with the variables Marital.status, Income, and CapGainsLossesPositive before modeling?

Answer 48

Convert them to factors.

Question 49

What is the command to evaluate the NNM1 model using the test data set?

Answer 49

Construct a contingency table comparing actual and predicted values.

Question 50

What is the purpose of standardizing the variables in neural network modeling?

Answer 50

To ensure that all inputs contribute equally to the model.

Question 51

What is the command used to obtain the weights for the neural network?

Answer 51

nnet01$wts