K Means

• Codebook vectors. No manifolds.
• Given: $(x_i{)}_{i=1,..,N}\in {\mathbb{R}}^n$
• Need : K clusters $C_1,\dots ,C_K$
• Randomly assign training points to K sets : $S_j(j=1,\dots ,K)$
• Repeat:
  • For each set $S_j$
    • Mean ${\mu }_j=|S_j{|}^{-1}{\Sigma }_{x\in S_j}x$
    • Create new sets by putting points into set where $||x_i-{\mu }_j||$ is minimal
    • If empty, dismiss and reduce K to K’
• Error quantity does not increase
  • Intra cluster variance.md
  • Clusters are bounded by line Decision Boundaries.md and forms a Voronoi Cell.md
• Does not work for curved boundaries

Codebook Vector

• Each cluster represented by it
• Vector pointing to the mean of all vectors in the cluster
• Center of Gravity.md