Question 1

Axons, dendrites, and synapses are studied by what discipline, which is the inspiration for artificial neural networks? ___________________.

Accepted Answer

Neurobiology. 
In neurobiology, dendrites in neural networks collect signals which are fed to neurons. Neurons process a signal by sending a spike of electrical current along an axon, discharging at a synapse connected to other neurons, which in turn are excited or inhibited as a result.

Question 2

Neural networks may be represented by an equation or instruction set.

Accepted Answer

Where ordinary statistical procedures are based on instruction sets (programs), ANNs are not. Instead ANNs pass data through multiple intermediary digital processing entities (neurons) on the way to output entities (neurons in the terminal layer). Intermediary neurons "learn" through shifting patterns of weights based on input signals. Shifting weights reflect better or worse predictions or classifications at the terminal output layer. In ANNs, no "solution" or "answer" is ever stored at a particular computer memory address. Rather, the solution takes the form of weighted connections along linkages from the input layer through one or more processing layers to an output layer of neurons.

Question 3

Unlike linear regression, neural networks can handle nonlinearities and interaction effects even when these are not explicit input variables.

Accepted Answer

ANNs automatically handle nonlinearities and interaction effects in the data. Put another way, it is not necessary that causal dynamics in the field be well-enough understood that nonlinearity and interactions can be identified and explicitly incorporated in the model, as traditional statistical methods would require.

Question 4

It is impossible to assess the relative importance of input variables in neural networks, which are a "black box".

Accepted Answer

While this was a traditional view of neural networks and one finds the "black box" metaphor in descriptions of NN, this is no longer true. In Chapter 7, the Olden & Jackson (2002) method for assessing the relative importance of inputs was presented in worked examples.

Question 5

Cross-validation of a neural network (or other) model requires …
A) One or more test subsets
B) One or more hold-out subsets
C) One or more validation subsets
D) All of the above

Accepted Answer

Models are trained on the training subset, often labeled just "train". While "test" is often used to refer to the data subset on which train-based model is validated, there are synonyms for the test set. Synonyms include the hold-out, validation, or OOB (out-of-box) subsets.

Question 6

What is UNTRUE of hidden layers of neurons in a neural network model?
A) Depending on the software used, there may be more than one hidden layer.
B) It is possible to design a neural network with direct connections from the input to the output layer, skipping the hidden layer.
C) The more neurons in hidden layers, the better the resulting model fit.
D) Activation rules govern the firing of neurons.

Accepted Answer

Finding the optimal number of hidden layers and neurons is one of the empirical matters investigated when tuning a neural model. Depending on the problem, more neurons may actually diminish fit. A is true, though the nnet program only supports one. B is true, though not all programs support skip layers (nnet does, neuralnet does not). D is true. Common activation functions are sigmoid (often the default), tanh, and ReLU.

Question 7

Explain "gradient descent" and what it does in a neural network algorithm.

Accepted Answer

The answer of Explain "gradient descent" and what it does...

Question 8

What makes a neural network model "supervised"?
A) input data in the test set includes the correct output values
B) input data in the training set includes the correct output values
C) input data in the test set includes the correct input values
D) input data in the training set includes the correct input values

Accepted Answer

"Unsupervised learning", in contrast, takes input examples and classifies them inductively without reference to correct classifications.

Question 9

Which is possible with the "caret" package?
A) Running a nnet model
B) Running a neuralnet model
C) Both A and B
D) Neither A nor B

Accepted Answer

Both nnet and neuralnet may be run under caret, or run independently of caret. Caret also supports other neural network methods.

Question 10

Which package supports the Olden method of assessing the relative importance of predictors in a neural network model?
A) NeuralNetTools
B) neuralnet
C) nnet
D) caret

Accepted Answer

The olden command from NeuralNetTools supports model objects created by neuralnet, nnet, caret's train command, and others. However, B, C, and D do not themselves support the Olden method. Caret, however, supports the varImp command, which also supports a method of assessing variable importance. The Olden method is based on Olden & Jackson (2002). The caret method is based on Gevrey et al (2003).

Question 11

What was a "violin plot" used for in Chapter 7, for the "iris" example?
A) to estimate variable importance
B) to obtain graphical insight into how separable the three classes of Species are on the four predictor variables.
C) to visualize the results of a neuralnet model
D) to visualize the results on a nnet model

Accepted Answer

Violin plots were used prior to modeling. They anticipate the feasibility of a good neural network (or other) solution to a classification problem.

Question 12

Neural networks may achieve better results if data are normalized or scaled.

Accepted Answer

Normalization or scaling is often recommended. It can make a difference with regard to whether there is convergence on a solution at all. However, it is not a requirement of neural network analysis.

Question 13

If data are normalized or scaled, which model performance metric will NOT change?
A) MSE
B) RMSE
C) R-squared
D) All of the above will change
E) None of the above will change

Accepted Answer

Changing the scale will affect mean square error and root mean square error, but not R-squared or correlations.

Question 14

The plotnet command is from which R package?
A) nnet
B) neuralnet
C) caret
D) NeuralNetTools

Accepted Answer

Though the generic plot command may be used to visualize some neural models, arguably plotnet creates clearer and more aesthetic depictions of a neural model. This in turn may be more informative.

Question 15

What is NOT an advantage of running a neural model under caret?
A) optimization of all parameters is built in
B) tuning of some parameters is built in
C) cross-validation is built in
D) All of the above are advantages

Accepted Answer

Running a package such as nnet on a stand-alone basis rather than under caret allows optimization of any parameter. Caret just optimizes for selected parameters. B and C are advantages. D is untrue.

Question 16

Which package brings to R some of the functionality of the Python environment, which is often preferred for large-scale neural network analysis?
A) bigdata
B) mlr3keras
C) caret
D) neuralnet

Accepted Answer

The mlr3 and mlr3keras packages, supported by other packages associated with the mlr3 environment, bring to R many of the function associated with Python, Keras, and TensorFlow. This is the focus of a supplement to Chapter 7. There is no A. C and D do not bring the Python environment to R.

Quiz 7: Neural Network Models and Deep Learning

Axons, dendrites, and synapses are studied by what discipline, which is the inspiration for artificial neural networks? ___________________.

Neural networks may be represented by an equation or instruction set.

Unlike linear regression, neural networks can handle nonlinearities and interaction effects even when these are not explicit input variables.

It is impossible to assess the relative importance of input variables in neural networks, which are a "black box".

Cross-validation of a neural network (or other) model requires …

What is UNTRUE of hidden layers of neurons in a neural network model?

Explain "gradient descent" and what it does in a neural network algorithm.

What makes a neural network model "supervised"?

Which is possible with the "caret" package?

Which package supports the Olden method of assessing the relative importance of predictors in a neural network model?

What was a "violin plot" used for in Chapter 7, for the "iris" example?

Neural networks may achieve better results if data are normalized or scaled.

If data are normalized or scaled, which model performance metric will NOT change?

The plotnet command is from which R package?

What is NOT an advantage of running a neural model under caret?

Which package brings to R some of the functionality of the Python environment, which is often preferred for large-scale neural network analysis?