Global Convergence of SGD On Two Layer Neural Nets