The design of an adaptive learning rate, which tracks the
input signals in an on-line manner
and which modifies the synaptic weights of
a neural net in case of changes in signal statistics.
We have proposed a learning rule for the adaptation
of learning rate (in matrix and vector form)
in case of feed-forward and recurrent networks.
See the paper presented at ISCAS 1996.
The validity of proposed learning rate adaptation
for feed-forward neural networks performing blind separation
is demonstrated below.
Example:
Fig. 1 Three image sources - one natural and two synthetic images.
Fig. 2 It is switched twice on the input between above two mixtures. The top mixture
appears during epochs 1-2 and 5-6, whereas the bottom mixture appears during epochs 3-4.
Fig. 3 The behavior of synaptic weights (3x3-matrix) during learning.
Fig. 4 The behavior of the learning rate vector Eta and the combined
separation error index PI.
Learning rate for nonstationary signals:
W. Kasprzak, June 1996