CMSC 491N/691N Introduction to Neural Networks Spring 2001

CMSC 491N/691N Introduction to Neural Networks Spring 2001

Makeup Exam 1

1.  Short questions (10 points each)

a)  What are the major differences between artificial neural networks and the traditional Von Neumann architectures?

b)  What is the “cross-talk” term in an associative memory? How does it affect the associative recalls for such memories? And under what conditions the cross-talk term becomes zero?

c)  What is a feedforward network and what is a recurrent network? Indicate which of the following are a feedforward and which are recurrent:

i.  Perceptron

ii. BP networks

iii.  Hopfield associative memory model

iv.  Bi-directional associative memories

2.  (40 points) This question concerns backpropagation (BP) learning

a)  What is the network paralysis problem in BP learning? Can you think of a way to ease this problem?

b)  In what ways does BP learning generalize the delta rule used by Adaline?

c)  What is the approach that BP model takes to deriving the learning rule? What is the major limitation of this approach?

d)  Give the weight update rule for hidden units in BP network. Please indicate the meaning of each term in the rule.

e)  List three major disadvantages of BP networks.

3.  (30 points) This question concerns discrete Hopfield model for associative memory

a)  Construct the weight matrix for a discrete Hopfield associative memory of four nodes to store a single bipolar pattern of s(1) = (1, -1, -1, 1).

b)  Describe the recall procedure for Hopfield associative memory model.

c)  Show that the stored pattern can be recalled by the input pattern (1, 0, 0, 0), which misses three of the four elements of s(1).

d)  What will be the recall result with the same input pattern (1, 0, 0, 0) if the memory in (a) is extended to store the second pattern s(2) = (-1, 1, -1, 1)?