Question about ISAB

[zhiqihuang]

Not sure if I understand the Induced Set Attention Block correctly.

So basically SAB is a transformer without positional encoding (and dropout?). In the paper, you said that SAB is "too expensive for large sets". But set size here refers to the max sequence length in a transformer which is usually 512. Why not just use the SAB for SetTransformer? Is there any reason other than efficiency, to use ISAB for SetTransformer?