Another initialization approach is to use a specific method that learns
pattern distribution from the data and uses it as the filter’s
initialization. The principal component analysis (PCA) was used to
initialize a self-organizing map as a filter, and faster convergence was
observed. K-means-based pre-initialization was experimented with and
compared with random initialization. The K-means initialization uses an
orthonormal matrix to rotate centroids symmetrically for an optimal
solution through iteration. The generated filters are shown in Figure 2.
Some filters learned with random initialization are prone to be noisy,
as highlighted in red. K-means-based initialization would reduce noisy
filters and result in faster convergence.
It
was also noted that any clustering method could produce a similar
outcome.
FIGURE 2 filters learned
after random initialization (Top) and after using K-means-based
initialization (Bottom)