I shall describe past and current projects on non-convex optimization arising in signal/image recovery and classification. The non-convexity comes from either sparse constraint or objective function constructed from probability or information theory. We shall explore techniques beyond convex relaxations.