CS 5541 Study Questions and Homework 5
Homework (15 points) due Wednesday, December 10, 2003.
For the final exam, also see the first four sets of study questions (Review
1, Review 2, Review 3,
Review 4). The final exam will be cumulative.
For this Homework, turn in your answers for the following study questions:
Q1 [2 pts]
Q6 [1 pts]
Q12 [2 pts]
Q14 [2 pts]
Q15 [2 pts]
Q18 [2 pts]
Q19 [2 pts]
Q20 [2 pts]
- For the following data set, give a single layer neural net architecture
suited to attempting approximation of the data set:
|0 0 0 0
|1 0 0 0
|0 1 1 0
|0 0 0 1
- In the last question, what function do you think is given in the table?
- Use a spreadsheet like that used in class to iterate the delta rule through
20 epochs (cycles through the data) to attempt to find weights that solve
this problem for your ANN in question #1. Use the sigmoidal activation function.
- Were you able to find weights that approximate the function? Explain your
- Give four additional inputs (binary valued vectors of length 4) for the
network, and run them with the weights you found in #3. Give the outputs of
your network. Explain your results.
- What mathematical idea does the delta rule rely on? (You don't have to give
the derivation, just state the idea).
- Show that without an activation function, a neural network always outputs
0 when the inputs are 0.
- What problem does backpropagation solve that is not solved by the delta
- Is the delta rule a supervised training method? Explain.
- Is backpropagation a supervised training method? Explain.
- What would an unsupervised method that implements association learning in
neural networks look like?
- Is a self-organizing map (a) supervised, or (b) unsupervised? Explain.
- Draw a diagram in terms of neural units, and connection weights for the
general self-organizing map (SOM) architecture.
- A given SOM has 3 inputs and two outputs. Draw and label the architecture
of the SOM.
- How many categories can the SOM in question #14 represent? Explain your
- How do we test for generalization in a SOM?
- In the algorithm for training (establishing) the weights in a SOM, before
the weight update rule can be applied, what is done?
- Give the weight update rule for the SOM algorithm. Make sure to explain
- For a SOM with three inputs and two outputs, one input data sample is: (1,
1, 0). Assuming the weight vectors are (.5 .2 .3) for unit 1 in the output,
and (.8 .9 0) for unit 2 in the output, compute the weight update for that
- For a SOM with four inputs and two outputs, and weight vector (0 0 .5 1.0)
for output unit 1, and (1.0 .5 0 0) for output unit 2, what category does
the input vector: (0 1 0 0) fall into?
- On what basis do SOMs achieve categorization?