Question 1

a) Explain the THREE motivations given in the subject guide for studying ANNs.

b) Define the terms network, weight, activation function, and bias as they relate to ANNs.

c) Describe the features of a neural network that we need to specify when giving an architecture.

d) Compare and contrast supervised and unsupervised learning. Give examples of where each is appropriate and the types of networks to which each type of learning can be applied to.

Question 2

a) Using examples and a diagram, explain the limitations of a single Perceptron.

b) Despite the limitations hinted at in part a) above, Networks of Perceptrons are in some sense ‘universal’. Explain what this means and illustrate your answer by means of the XOR function as an example.

c) Design a two-input network of Perceptrons (threshold units) that produces an output of 0 if and only if both of its inputs are between -0.7 and 0.7. Explain how it achieves its design goal.

Question 3

a) Compare and contrast THREE activation functions that we have met during this course, giving a diagram showing their form and an expression for each.
b) Briefly describe network paralysis and overfilling and the steps that can be taken to avoid or overcome these problems.

c) A Backpropagation network has weights as shown in Figure B3. Calculate the weights after training with the examples shown.

 

Question 4

a) Explain how Kohonen Networks differ from Perceptron Networks.

b) Why is normalization often necessary in the training of Kohonen Networks? What may go wrong if this is not done?

c) Give the algorithm for training the Kohonen layer of such a network.

d) What is the function and form of the Grossberg layer of a Kohonen-Grossberg Network?