Speech Processing
by Computer
LECTURE 6a
NEURAL NETWORKS
Objectives
By the end of the
session you should:
• be able to justify an interest in
neural networks
• be
able to contrast a real with an artificial neurone
• be
able to describe in general terms the computational processing that goes on
inside an artificial neurone
• appreciate
that networks of artificial neurones can store, process and learn information
• have
gained experience with perceptrons learning some simple tasks
• be
able to list some application areas for neural networks in speech
Outline
1. What is a neural
network?
q
Computation inside an artificial neurone
q
Units, inputs, weights, activation, output
2. Why are neural
networks interesting
q
The appeal of connectionist processing
3. Common
architectures for networks
q
Hopfield network
q
Interactive Activation and Competition network
q
Feed-forward (perceptron)
network
q
Dynamic (recurrent) network
q
Kohonen (self-organising) map
4. Example of
network learning
q
Perceptron learning
q
Self-organising map learning
Reading
J.L. McClelland,
D.E.Rumelhart, G.E. Hinton, "The Appeal of Parallel Distributed
Processing", in Parallel Distributed Processing Volume 1, ed
Rumelhart & McClelland, MIT Press, 1986.