Consider the following two multilayer perceptron's, where all of the layers use linear

activation functions.

$[3$ Marks$]$

Network A

Network B

$($a$)$ State the advantage of Network A over Network B$.$

$($b$)$ State the advantage of Network B over Network A

$($c$)$ Suppose you design a multilayer perceptron for classification with a single hidden

layer with the hard threshold activation function. The output layer uses the softmax

activation function with cross$-$entropy loss.

What will go wrong if you try to train this network using gradient descent?

Justify your answer in terms of the backpropagation ruies.