Why are activation functions essential in neural networks? What would happen if a neural network used only linear activation functions?
