By: Moritz (email@example.com), May 22, 2020 1:20 am
Room: Moderated Discussions
I think the divide of training, inference and 'regular' compute (iterative solving of large linear equation system) will break down in the coming years and that all will be done in the same place at different times as an iterative process of "-> use model to 'guess' direction -> compute/simulate/explore -> update/refine model and progress simulation or optimization system wide to the best known state -> repeat". Today computers walk curvy valleys in high dimensional spaces in small steps by brute force, oblivious to the nature of the numerical task they are performing, soon they might tunnel by using "intuition" incorporating experience with that type of problem.
|Coming compute style?||Moritz||2020/05/22 01:20 AM|