Re:Autopsy of a deep learning paper
by throwaway080383 on Saturday July 21, 2018 @09:54AM
(#17581973)
Yeah, I'm not a deep learning researcher, but I didn't understand the incredulity at 100 GPU's. Surely Uber has that many lying around, so why not overkill on the hardware when testing? This leaves no doubt that the feature is worthless for the given task, and does not leave open the question of whether more hardware could produce better results.
[ Parent
| on HN ]
If training requires N operations, using more GPUs (that otherwise would idle) simply means you finish (and iterate) faster.
Re:Autopsy of a deep learning paper
by HelloNurse on Saturday July 21, 2018 @02:15PM
(#17583045)
It's a genuine scientific question: adding a useless feature slows down NN training, but how much? Is the impact as negligible as can be expected?
There's no reason to spend lots of brain cycles on low-value theory and possibly wrong estimates instead of lots of GPU cycles on a conclusive experiment.