Deep learning approaches offer more stochasticity so converge to global solutions more readily (at the risk of overfitting).

Yeah, thats a fair point and another appealing reason for DL based methods