Interpreting vision transformers via residual replacement model