Referential communication in heterogeneous communities of pre-trained visual deep networks