We believe that the use of subspace-restricted error norms (e.g., eigenspace matching) for inferring can result in a highly robust and fast recognition system.