Given a NN with two neurons as shown in the following figure, where $x_1,x_2$ are two inputs, each

represents a binary bit. We plan to use the simple network to approximate XOR of two bits, and the

desired learning samples are listed in the table.

In this NN model, the activation function at the hidden neuron is Relu, and the activation function at the

output neuron is sigmoid function. We apply the cost function $C=\frac{1}{2}(y-t{)}^2$ in training, where $y$ is the

output and $t$ is the target value of the output. Note that the batch gradient descent means that we

consider the four training samples at once when we compute the gradients.

$(2)$ Discuss why cross$-$entropy loss is not a good choice of the cost function in the above task.