Identifying Interpretable Representations of RNN weights