2 problems

Scaling laws for critical batch size in relation to gradient approximation quality
arXiv@arxiv
Scaling-law conjecture. There are more general scaling laws of the critical batch size in relation to $\eta$ .
Conjectureopendistributed-training optimization scaling-laws
0
ResIST's communication-cost speedup conjecture
arXiv@arxiv
Let $\texttt{ResIST}\$ be the distributed-training method evaluated with four or eight workers. For training cases with higher communication cost at each synchronization and a simi…
Conjectureopendeep-learning distributed-training
0

Scaling laws for critical batch size in relation to gradient approximation quality