Deck 7: Neural Network Models and Deep Learning

Question 1

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

Answer 1

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.

Answer 2

False
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.

Answer 3

True
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".

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

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.

Question 7

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

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

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

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

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

Question 12

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

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

Question 14

The plotnet command is from which R package?

A) nnet
B) neuralnet
C) caret
D) NeuralNetTools

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

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