Multiple Models Book: Local learning in Local Model Nets
This chapter deals with the local/global trade-off in learning. It points out problems with global learning
methods in Local Model Networks when the network is locally over-parameterised
or poorly structured. The bias-variance trade-offs for local and
global learning are examined, illustrating the regularising effect of
local learning, which can be advantageous compared to direct maximisation
of a global cost function.
Smoothing analysis is used to explain the results, and graphical
interpretation of the theoretical smoothing results allows more intuitive explanations of differences in learning algorithms.
The methods are then related to Expectation Maximisation (EM) techniques,
and the smooth-based estimates of the degrees of freedom are used to develop
a new learning algorithm which minimises the generalised cross-validation
(GCV) error statistic.
R Murray-Smith
Last modified: Tue Mar 25 16:27:59 GMT