Learn more, but bother less: parameter efficient continual learning