Exploring disentangled feature representation beyond face identification