[0. Recall to write a quick summary of the paper.]

1. Why does SecureML approximate the activation functions with polynomials?  In other words, why does it not compute the exact activation functions using 2PC (which is general enough to support them)?