Has any of you retards encountered this problem before, and how did you resolve it?
>>109940598You're disrespectful. They have feelings too. Bigot
did you read the paper because they literally provide an alternative algo to help fix this
>>109940598>person tries one thing first>ends up with that being a lens they see everything else throughI don't think humans have solved this problem themselves either.
>>109940598I call it "baby duck syndrome".
>>109941960only thing i got from it was to use LayerNorm, testing it right now actually, the rest of their ideas are unproven, i will try a few of their solutions, some of the architecture they use i cannot do since most of my VRAM is already occupied by weights
>>109942123do you agree with 50:50 offline:online distribution? because i do 90/10, maybe that is my mistake
>>109940598Protip: It is an open secret that Aaron Courville steals credits. In particular, he organizes his students in two main groups: one group just lazes around doing nothing in general, and the other student(s) are isolated and not allowed to know who else is in the group or where they are. He makes the isolated students do work, and when they produce something useful, he tells them to drop it because it won't be publishable. He then hands the work over to the "student group". This has been happening since at least 2013 and Yoshua Bengio has known that the entire time. He is in on it, rather than merely "letting it happen for political reasons", because it is good for his reputation as a lab due to the setup being reputatively successful.
>>109942472Not him and I haven't done RL in a hell of a long time, but back in my days it was understood that offline scales better but goes off-track more than online. The ideal was always to do 100% online for that reason, but it was not practically doable. Of course we would prefer to be able to do 100% offline if there was any chance that could go anywhere. tl;dr: the more online you can tolerate, the better. This was pre-trpo and pre-ppo though.
>>109940598Lolol just solve it with claude pro maxx
>>109944465how do u know this
>>109944465i mean i am pretty sure my piece of shit research professor at NYU did the same shit with my research, he literally misdirected me, shut me down when i showed him data, then when i was done with my thesis he goes "oh we should have gone with that other path" and he did it because he really disliked me by the end because i would not come at 9am to do his fucking research for free, but there was no point if i could do everything in 4 hours of my own choosing.
>>109944472thank you so much this is helpful, yeah online is such a pain in the ass it slows everything down by 5 minutes per iteration, making experiments super painfuli noticed this instability only at the start when you use scratched weights, if weights are stable and you run experiments, you can get away with offline, there is turbulence but its palatable, and as i understand after validation you can rerun with online or just cure it in the end
>>109944480can't its too stupid for this, so is astra
>>109944504Trade secret (TM)>>109944524There's a lot of that kind of shit in academia in general, but this very specific form is not that common. The more common variant is to bully someone who did the whole work into putting everyone's name on the paper under the premise that the next paper will carry their name. The more advanced version of this is to change the first author under the same guise. Of course, the promise is never realized. Many scummy profs will keep PhD students around for 10 years by preventing them from graduating on purpose because they're cheap slave labor using these tricks.
>>109944568that sounds horrible like a very very deep spite and envy, i cannot believe there is no system to root people like that out, on the contrary they will be protected and defended
>>109944590Yes, tenure is necessary but also a great evil. The worst part is that there is an oversight committee for these kinds of things, but if you lodge a complaint you will ruin your entire career even if something gets done. In practice, because the oversight committee is allowed to mingle, the profs most likely to do this shit will cozy up to them to prevent any action to happen against them.
