Skip to content
Hackerly

Comment on Energy consumption comparison in machine learning platforms

Comments

sva_
TF [...] The final mean squared error is 0.0003.
Neural Designer [...] reaches a mean squared error of 0.023.
The following table summarizes the the[sic] most important metrics that the two machine learning platforms yielded .

[omits MSE]

They should train both to the same loss and then compare.

BobbyJo

I'd like to know why there is a difference in the final loss at all. If the two networks had the same architecture, used the same loss function, and had random uniform initialization, then 1000 epochs should have them converging on very similar final loss values. Especially if one was able to converge to 3e-4.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.