Learning invariant features using the Transformed Indian Buffet Process