|
m8 wtf are you doing
|
#
?
Jun 8, 2020 16:23
|
|
|
|
| # ? Apr 28, 2024 20:25 |
|
|
Arcteryx Anarchist posted:m8 wtf are you doing The gradient, aka the total loss, is clearly descending. Not sure what you want the guy to do beyond that.
|
|
#
?
Jun 8, 2020 16:44
|
|
|
Arcteryx Anarchist posted:m8 wtf are you doing trying to train an autoencoder-ish thing based on resnet-50 with ms coco. the total_loss is just the mse. i started training it on friday and basically it spent ~31k batches of images oscillating around learning nothing. time to start randomly changing various hyperparameters!
|
|
#
?
Jun 8, 2020 16:57
|
|
|
what's your learning rate strategy?
|
|
#
?
Jun 8, 2020 17:34
|
|
|
initial guess with 1e-5, detectron2 does some lr warmup which is why it starts out lower than that for the beginning batches. besides double checking all the code, lowering the learning rate was experiment #1 on the "randomly change hyperparams" list.
|
|
#
?
Jun 8, 2020 18:16
|
|
|
it won’t have actually learned anything until it’s able to explain what it has learned sorry, but someone had to say it
|
|
#
?
Jun 9, 2020 02:34
|
|
|
Max Facetime posted:it won’t have actually learned anything until it’s able to explain what it has learned this network has learned something but no animal ever has 
|
|
#
?
Jun 9, 2020 03:55
|
|
|
Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead
|
|
#
?
Jun 9, 2020 04:04
|
|
|
rchon posted:this network has learned something but no animal ever has ultrafilter posted:Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead idk seems good
|
|
#
?
Jun 9, 2020 05:13
|
|
|
also isn’t their counter example to COMPAS what COMPAS was? The only black box aspect of it is that it was proprietary and not publicly available and when it was looked into it was basically a spreadsheet doing exactly that?
|
|
#
?
Jun 9, 2020 05:15
|
|
|
https://twitter.com/cstross/status/1270273981283794948 https://twitter.com/jimwaterson/status/1270314033724481538 Microsoft’s AI editor’s job: be possibly racist Microsoft’s human editors’ job: cover up for the racist AI
|
|
#
?
Jun 10, 2020 10:40
|
|
|
their faces look like they were ai generated too tbh
|
|
#
?
Jun 10, 2020 12:31
|
|
|
thispopactdoesnotexist.com
|
|
#
?
Jun 10, 2020 13:47
|
|
|
echinopsis posted:what could you do with a lot of real i got some ideas, you got an address i can contact you on?
|
|
#
?
Jun 10, 2020 14:18
|
|
|
goonsource the project echi
|
|
#
?
Jun 10, 2020 15:15
|
|
|
ultrafilter posted:Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead  there's also some other work on how you can trick explanation methods. so like, if you're relying on visualizations of your deep neural network to determine if it's making bad decisions, an adversary can pretty easily trick the visualization tools as well as the network itself basically don't trust anything where you can't fully explain the algorithm it's using
|
|
#
?
Jun 10, 2020 16:43
|
|
|
animist posted:
|
|
#
?
Jun 10, 2020 16:46
|
|
|
There's some work showing that optimizing for accuracy leads to vulnerability to adversarial attacks (Robustness May Be at Odds with Accuracy, Theoretically Principled Trade-off between Robustness and Accuracy). A lot of this seems like it's really deep-learning specific, but maybe that's just because people are only looking there right now. Something else that may be interesting: Politics of Adversarial Machine Learning
|
|
#
?
Jun 10, 2020 17:05
|
|
|
I think DNNs come up a lot because its where you're really getting into interlayering more or less generic functions and trying to make the math work out to do what you want
|
|
#
?
Jun 10, 2020 18:15
|
|
|
Arcteryx Anarchist posted:I think DNNs come up a lot because its where you're really getting into interlayering more or less generic functions and trying to make the math work out to do what you want 
|
|
#
?
Jun 10, 2020 20:35
|
|
|
uuuh tensors thanks 
|
|
#
?
Jun 10, 2020 22:51
|
|
|
Arcteryx Anarchist posted:uuuh tensors thanks “tenser” said the tensor.
|
|
#
?
Jun 10, 2020 22:55
|
|
|
vodkat posted:i got some ideas, you got an address i can contact you on? mandruku at gmail but idk if you’re serious  Max Facetime posted:https://twitter.com/cstross/status/1270273981283794948 tell me is MSN microsoft? and and MSN and NBC combine to make MSNBC?
|
|
#
?
Jun 12, 2020 20:24
|
|
|
echinopsis posted:tell me is MSN microsoft? and and MSN and NBC combine to make MSNBC? they did but microsoft sold their stake in the network years ago.
|
|
#
?
Jun 12, 2020 20:29
|
|
|
a neurotic ai posted:gradient descent means literally into hell. post / username 
|
|
#
?
Jun 13, 2020 16:33
|
|
|
Schadenboner posted:thispopactdoesnotexist.com if you could both generate the look and the lyrics and the music and put it all together this could be awesome
|
|
#
?
Jun 13, 2020 18:54
|
|
|
https://twitter.com/MaartenvSmeden/status/1272613128304578561
|
|
#
?
Jun 15, 2020 20:40
|
|
|
lol https://twitter.com/MaartenvSmeden/status/1272613790593548289?s=20
|
|
#
?
Jun 15, 2020 20:48
|
|
|
wtf
|
|
#
?
Jun 16, 2020 09:31
|
|
|
I had to read that like 5 times before I realised he meant they used the full date and not like day of week or something
|
|
#
?
Jun 16, 2020 13:00
|
|
|
Patients admitted on a weekend are more likely to cheat on their spouses 
|
|
#
?
Jun 16, 2020 22:08
|
|
|
Statistics can get real weird man. We get it beaten into us in formal logic that correlation does not imply causation, but then stats basically turns around and says ‘yeah but if this R is high enough then we gonna build a model that assumes it is anyway’.
|
|
#
?
Jun 16, 2020 22:39
|
|
|
i suspect that was supposed to be Mortality but who knows
|
|
#
?
Jun 16, 2020 23:18
|
|
|
a neurotic ai posted:Statistics can get real weird man. We get it beaten into us in formal logic that correlation does not imply causation, but then stats basically turns around and says ‘yeah but if this R is high enough then we gonna build a model that assumes it is anyway’. that doesn't sound like a good use of stats
|
|
#
?
Jun 16, 2020 23:26
|
|
|
really its just that stats are a tool and people can ask dumb questions and then those tools have no defense against them
|
|
#
?
Jun 16, 2020 23:29
|
|
|
a neurotic ai posted:Statistics can get real weird man. We get it beaten into us in formal logic that correlation does not imply causation, but then stats basically turns around and says ‘yeah but if this R is high enough then we gonna build a model that assumes it is anyway’. No one in statistics is saying that.
|
|
#
?
Jun 16, 2020 23:34
|
|
|
ultrafilter posted:No one in statistics is saying that. coulda fooled me
|
|
#
?
Jun 17, 2020 00:48
|
|
|
Carthag Tuek posted:coulda fooled me That's because you've never talked to anyone who honestly works with stats. Yeah, there are a whole bunch of liars who set out to use the formalisms of stats to justify whatever agenda they're trying to push, and also a bunch of idiots that throw stats machinery at problems and blindly pick things with a strong correlation. A good statistician doesn't do either of those things though.
|
|
#
?
Jun 17, 2020 01:09
|
|
|
no true statistician lol
|
|
#
?
Jun 17, 2020 04:18
|
|
|
|
| # ? Apr 28, 2024 20:25 |
|
|
Jabor posted:That's because you've never talked to anyone who honestly works with stats. Statistically, most statisticians are the idiots and liars
|
|
#
?
Jun 17, 2020 04:50
|
|
