The paper does not establish why EGD and vanilla SGD converge to different regions on the modular tasks, or whether the trajectory observations extend to randomized EGD or other tasks; these questions are left for future work. · CiteArk
The paper does not establish why EGD and vanilla SGD converge to different regions on the modular tasks, or whether the trajectory observations extend to randomized EGD or other tasks; these questions are left for future work.
