Executing the exact activation functions in 2PC (in particular with secret shares as in SecureML) is very expensive. Instead, SecureML approximates them via low degree polynomials which fit better the 2PC computation model via secret shares.