Effect of Activation Functions on the Training of Overparametrized Neural Nets