Is it possible for fully connected layers to learn square/cube operations?

Is there any structure for fully connected layers that could learn square/cube operations?
As far as I known, fully connected layers could learn linear representation well. But for specific no-linear representation it remains a mystery.  2^2 = 4 2^3 = 8 

• This question is a bit of mystery to me... Why do you need neural networks to learn cubic or quadratic function? – Jan Kukacka Aug 22 '18 at 6:40
• @JanKukacka Because I want to interpret a model, I think this model may has implemented square operation. But this model is a bit complex. – Mark Aug 22 '18 at 6:50
• Are you asking if a neural network with linear activations can learn the square function, or something else? – Coolness Aug 22 '18 at 10:25