7/20/22, 11:52 PM Module 3 Quiz: Attempt review

Second Summer CSC 7333 for Jianhua Chen / Module 3: Neural Networks and Deep Learning / Module 3 Quiz

Started on Wednesday, July 20, 2022, 11:31 PM

State Finished
Completed on Wednesday, July 20, 2022, 11:51 PM
Time taken 20 mins 17 secs
Grade 75 out of 75 (100%)

Question 1
Correct 10 points out of 10

Which statement is true for feed forward neural networks?

Select one:
A. The graph of the network’s topology has directed cycle.
B. Any bounded continuous function f can be approximated arbitrarily close with a feedforward network with one 
hidden layer.
C. The only possible activation for any feedforward network is the sigmoid function.
D. Any Boolean function can be represented by a single perceptron.

Question 2
Correct 10 points out of 10

Which update rule is correct (for stochastic gradient descent) for a single node neural network with
sigmoid activation function? Assume the learning rate is η > 0 and d= (x, t) is an input training data
point and the network produces o as the output for this data point. The error function is
⎯ 1 ⎯
E d " (W ⃗ ) " = (t−o) 2, Here W = (w 0,w 1,…,w n )T , and x = (x 0,x 1,…,x n )T ,with x 0 = 1.
2

Select one:
A. w i ← w i +η (t−o) *x i f or 0 ≤ i ≤n

B. w i ← w i −η (t−o) *x i f or 0 ≤ i ≤n

C. w i ← w i +η (t−o) *x i f or 0 ≤ i ≤n
D. w i ← w i +η (t−o) *o * (1−o) *x i f or 0 ≤ i ≤n 

https://lsuonline.moodle.lsu.edu/mod/quiz/review.php?attempt=427550&cmid=264536 1/4
7/20/22, 11:52 PM Module 3 Quiz: Attempt review

Question 3
Correct 10 points out of 10

Consider the following statements about backpropagation:

1. Backpropagation is useful even if the network has NO hidden layer.
2. Training neural network using Backpropagation is guaranteed to converge if the learning rate is
sufficiently small.
3. Backpropagation algorithm is based on the gradient descent idea.
4. When backpropagation training converges, it always finds the optimal weights with minimal
error.
5. One nice feature of backpropagation is to compute the “error” at a hidden node from
combination of errors in the down-stream nodes.

Which choice below identifies ALL true statements above?

Select one:
A. 1, 3, 5
B. 2, 3, 4
C. 2, 3, 5 
D. 1, 4, 5
E. 1, 3, 4

Question 4
Correct 10 points out of 10

Identify ALL true statements below that suggest possible ways to handle local minimal and avoid
overfitting in Backpropagation (BP) training:
1. Consider doing multiple BP runs each with randomly selected initial weights to enhance the
chance of finding the global minima
2. The learning rate should always be very small
3. Consider adding a regularization term to the error function to tackle overfitting
4. Consider adding momentum in BP training to help avoid local minima
5. When designing the network topology, we should use many hidden layer nodes to fit the data
well

Which choice below is correct?

Select one:
A. 1, 2, 4
B. 1, 3, 4 
C. 3, 4, 5
D. 1, 4, 5
E. 2, 3

https://lsuonline.moodle.lsu.edu/mod/quiz/review.php?attempt=427550&cmid=264536 2/4
7/20/22, 11:52 PM Module 3 Quiz: Attempt review

Question 5
Correct 5 points out of 5

The most important characteristics of deep learning (DL) that differentiates DL from traditional
machine learning approach is that DL uses more layers in neural networks.

Select one:
True
False 

Question 6
Correct 5 points out of 5

One advantage of deep learning is that features automatically extracted by deep networks can be
adapted easily to other domains

Select one:
True 
False

Question 7
Correct 5 points out of 5

Learning in autoencoders is unsupervised learning.

Select one:
True 
False

Question 8
Correct 5 points out of 5

Deep autoencoder networks can be obtained by stacking autoencoders in a layer-by-layer fashion.

Select one:
True 
False

https://lsuonline.moodle.lsu.edu/mod/quiz/review.php?attempt=427550&cmid=264536 3/4
7/20/22, 11:52 PM Module 3 Quiz: Attempt review

Question 9
Correct 10 points out of 10

Consider the following statements about deep learning:

1. Deep learning can automatically learn useful features from data and thus avoid manual feature
engineering in traditional machine learning.
2. Deep learning is nothing but just neural networks which are deep in layers.
3. Convolutional neural networks (CNN) learn to support learning hierarchical features at different
level of abstraction.
4. Recurrent neural networks (RNN) are feedforward networks
5. RNN has strength in handling sequential data b/c it provides a method to “remember” history.

Which choice below identify ALL true statements above?

Select one:
A. 1, 2, 4, 5
B. 1, 3, 4
C. 1, 3, 5 
D. 2, 3, 5
E. 2, 4

Question 10
Correct 5 points out of 5

The training the denoising autoencoder, the loss function checks the error with respect to the
corrupted data.

Select one:
True
False 

Jump to...

https://lsuonline.moodle.lsu.edu/mod/quiz/review.php?attempt=427550&cmid=264536 4/4