MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
152<br />
correlated), spurious minima can also appear. This means that some patterns are associated with<br />
patterns that are not among the pattern vectors.<br />
Hopfield networks are sometimes called associative networks since they associate a class<br />
pattern to each input pattern. The importance of the different Hopfield networks in practical<br />
application is limited due to theoretical limitations of the network structure (capacity sensitive to<br />
correlated data) but, in situations where data can be decorrelated, they may form interesting<br />
models. Hopfield networks are typically used for classification problems with binary pattern<br />
vectors.<br />
6.9.1 Hopfield Network Structure<br />
xi<br />
w<br />
ij<br />
= w<br />
ji<br />
x<br />
j<br />
The Hopfield network is composed of K neurons and K2 connection weights w , i, j 1,.., K<br />
ij<br />
= . It<br />
is fully connected through symmetric, bi-directional weights, i.e. wij = wji<br />
. It has no selfconnections,<br />
i.e. w<br />
ii<br />
= 0<br />
∀ i .<br />
In the Hopfield network, learning takes one time step and retrieval takes several time steps!<br />
6.9.2 Learning Phase<br />
The learning rule is intended to make a set of desired patterns<br />
1<br />
stable states of the Hopfield network’s activity rule.<br />
Initialization:<br />
0 = 0 = 0 i,<br />
j<br />
At time t=0, set all the weights to w ( ) w ( )<br />
Update:<br />
ij<br />
ji<br />
∀ .<br />
r<br />
x n = { x n ,..., x n }, n=<br />
1,..., N<br />
The network weights are updated only once to represent the correlations across all bits from all<br />
patterns, following:<br />
w<br />
= η∑ x x<br />
(6.65)<br />
n n<br />
ji i j<br />
n<br />
x ∈−<br />
n<br />
j<br />
{ 1,1}<br />
K<br />
© A.G.Billard 2004 – Last Update March 2011