Two-argument activation functions learn soft XOR operations like cortical neurons