Perceptron Learning: Good
Some nice features...
No local minima on the error surface (if the target output values are within the range of
g
(
x
))
Given a sigmoidal activation function of the form:
the derivative is easy to compute:
g
'(
x
) =
g
(
x
)(1-
g
(
x
)).
If the learning rate is ``small enough,'' learning will converge on the correct weights in finite time
Next:
Perceptron Learning: Bad
Up:
NEURAL NETWORKS
Previous:
Gradient Descent on Squared
