Deep Learning Using SAS® Software

Question 1

What are the three Deep Learning model variants?

Answer 1

1. Deep fully connected neural networks (DNN)
2. Convolutional neural networks (CNN)
3. Recurrent neural networks (RNN)

Question 2

What other languages can read CASL?

Answer 2

Python, R, Java, and Lua

Question 3

What is Curriculum Learning?

Answer 3

slowly build up learning concepts aka shuffle action to randomize data

Question 4

What method of weight initialization is used in deep learning?

Answer 4

a normalized initialization in which the variance of the hidden weights is a function of the amount of incoming information and outgoing information

Question 5

What does RMSE stand for?

Answer 5

root means square error

Question 6

What does MAPE stand for?

Answer 6

mean absolute percentage error

Question 7

What is regularization?

Answer 7

the process of introducing or removing information to stabilize an algorithm’s understanding of the data

Question 8

What does the dropout regularization method do?

Answer 8

Dropout adds noise to the learning process so that the model is more generalizable

Question 9

How can you improve model generalization?

Answer 9

Dropout can improve model generalization

Question 10

What is a thinned network?

Answer 10

Each time that units are removed (USING DROPOUT), the resulting network is referred to as a thinned network

Question 11

What does the batch normalization regularization method do?

Answer 11

The batch normalization operation normalizes data being passed between layers in a neural network to prevent large input values in the combination function, which can lead to overfitting of the data.
batch normalization normalizes the information back to the linear region of the sigmoid, which is a safe output region;
batch normalization brings the weight values back to a familiar range

Question 12

Why do weight initializations have less impact on model performance if batch normalization is used?

Answer 12

batch normalization standardizes information that is passed between hidden layers

Question 13

When should you use GPUs instead of CPUs in Deep Learning?

Answer 13

The use of GPUs should be reserved for larger neural networks. The difference in performance between CPUs and GPUs is negligible in neural networks with a small number of parameters.

Question 14

Why are GPUs effective when modeling neural networks?

Answer 14

GPUs are designed to perform many operations in parallel

Question 15

What does Loss refer to in Neural Network output?

Answer 15

loss specifies the training error function value

Question 16

What does Validation Loss refer to in Neural Network output?

Answer 16

specifies the validation error function value

Question 17

What does Validation Error refer to in Neural Network output?

Answer 17

the validation misclassification rate

Question 18

What does Fit Error refer to in Neural Network output?

Answer 18

the misclassification rate for the training data

Question 19

What does a sudden increase in the L1 or L2 norm values indicate?

Answer 19

Overfitting (because the weights in the model must be getting really large to push those values up)

Question 20

What does ADAM optimization do if the signal to noise ratio is high?

Answer 20

adjusts the step size in order to take larger steps towards error minima

Question 21

What does ADAM optimization do if the signal to noise ratio is low?

Answer 21

adjusts the step size to move more slowly