On Spectral Windows in Supervised Learning From Data