125 points by mathwhiz19 4 months ago flag hide 10 comments
deepmathguy 4 months ago next
This is so cool! I've been hoping for something that cracks the nut of efficient training for so long.
deepmathguy 4 months ago next
@qwerty: From my understanding, the scale of the network shouldn't matter.
deepmathguy 4 months ago next
@bwh345: I'm seeing some pretty impressive results, haven't run any detailed comparison analyses yet.
ai_enthusiast 4 months ago next
@bwh345: I hear you, I have a few more experiments I need to run for that, hopefully soon.
qwerty 4 months ago prev next
Does this work for any sized neural network?
qwerty 4 months ago next
I guess it does work with sufficiently large NN's at least.
bwh345 4 months ago next
We definitely need to see more real-world performance data, especially for deeper networks.
bwh345 4 months ago prev next
How's the performance compared to traditional methods?
ai_enthusiast 4 months ago next
From my experience, this method is miles ahead in terms of GPU time and solution convergence.
quant_learner 4 months ago next
Fascinating—I’m doing some review to better understand the assumptions and limits.