FLDA Ideas

• \(S_{B}, S_{W}\): scatter matrices (estimate covariance)
• \(W\): projection matrix from \(D\) to \(K\) dimensions

Objective: maximize

\[J(W) = \displaystyle\frac{|W^{T}S_{B}W|}{|W^{T}S_{W}W|}\]

• eigenvalue scenario instead of gradient descent