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.
138<br />
There are two major advantages of using an ANN to process PCA as opposed to follow the<br />
algorithm given in Section 2.1.5. It allows on-line as well as long-life learning Classical PCA<br />
requires having at hand a sample of data point and runs in batch mode. ANNs can be run on-line.<br />
They will discover the PCA components as data are fed onto them. This provides more flexibility<br />
towards the data.<br />
Let us now show that the classical Hebbian rule converges to the principal components of the<br />
input vector:<br />
Recall that<br />
i i<br />
Δ w<br />
r =α yx<br />
r and that y = w rr T x. Thus, we have:<br />
i<br />
i<br />
Δ w r = α⋅y⋅ x r = α⋅ w r ⋅x r ⋅x<br />
r<br />
T<br />
( )<br />
(6.31)<br />
T<br />
Δ w r = α⋅ w r x r ⋅cos( θ)<br />
⋅x<br />
r<br />
where θ is the angle between the weight vector and the input vector.<br />
The weight vector moves proportionally to the distance across the weight vector and the input<br />
vector: the bigger the distance, the bigger the move. Hebbian learning will, thus, minimize the<br />
angle. If the data has zero mean, Hebbian learning will adjust the weight vector w r so as to<br />
maximize the variance in the output y . In other words, when projected onto this direction given<br />
by w r , the data will exhibit variance greater than in any other direction. This direction is, thus, the<br />
largest principal component of the data and corresponds to the eigenvector of the correlation<br />
matrix C with the largest corresponding eigenvalue.<br />
If we decompose the current w r in terms of the eigenvectors of C,<br />
expected weight update rule becomes:<br />
r r<br />
Δ w= α ⋅C⋅w<br />
= α ⋅ ⋅<br />
= α⋅<br />
C ae<br />
i i<br />
i<br />
∑<br />
i<br />
∑<br />
r<br />
aλe<br />
i i i<br />
r<br />
r<br />
w= ∑ae<br />
i i<br />
i<br />
, then the<br />
This will move the weight vector w r towards eigenvector e r i<br />
by a factor proportional to aiλ . Over<br />
i<br />
many updates, the eigenvector with the largest λi<br />
will drown out the contributions from the others.<br />
© A.G.Billard 2004 – Last Update March 2011