Insights on representational similarity in neural networks with canonical correlation