Given below is an implementation of the sigmoid activation function $($s$)$ during the backward pass:

Observe the input and output values for the function:

$($a$)$ Input

$30\mathrm{.}980\mathrm{.}960\mathrm{.}110\mathrm{.}127$

$1\mathrm{.}00\mathrm{.}110\mathrm{.}010\mathrm{.}92$

$($b$)$ Output

$30\mathrm{.}01960\mathrm{.}0\mathrm{.}3840\mathrm{.}09790\mathrm{.}10567$

$0\mathrm{.}000\mathrm{.}09790\mathrm{.}00990\mathrm{.}0736$

Note that the gradient values given as input to the function are all at the extremes $($near $0/1).$

Now explain why the values in $"$#$(\%)/\%$"are not good for learning and how that problem is overcome in ReLu.