Learning overcomplete representations