Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by sxp on Tuesday July 13, 2021 @03:29PM
(#27825176)
> By doing so, in addition to reducing the environmental impact of our experiments...
How does reducing the cost of individual experiments reduce the total environmental impact? Won't more efficient experiments (measured as experiments/Watt-hour or experiments/$) just trigger https://en.wikipedia.org/wiki/Jevons_paradox as more people take advantage of the lower cost?
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by monocasa on Tuesday July 13, 2021 @04:05PM
(#27825631)
It reads to me as a response to Gebru's paper that led to her getting fired. A paper on the externalities of training large models, like the environmental impacts due to the large amount of computation required: "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?"
That would explain why such a mundane subject is in a blog post, it's focused at tech news to be all "we really do care about this thing" rather than academia.
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by visarga on Tuesday July 13, 2021 @05:59PM
(#27826802)
> Can Language Models Be Too Big?
That's interesting, GPT-n with 100T parameters would be bored by just reading the whole internet. Too little information, repetitive and on average, junk.
You can try the question in reverse too: Can evolution use too much energy? How much energy has it already consumed?
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by pas on Wednesday July 14, 2021 @02:06AM
(#27830155)
It was not the paper that get her fired. (Probably there isn't just one ultimate cause that led to her firing, but maybe we can say that big G has a tendency to fire those who rock the boat too loudly.)
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by TaylorAlexander on Tuesday July 13, 2021 @05:56PM
(#27826770)
Amazing that they fired Timnit Gebru [1][2] after she pushed back against the removal of this very subject from one of her research papers, [3] only to publish their own work on it without mentioning her.
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by TaylorAlexander on Wednesday July 14, 2021 @01:49AM
(#27830070)
See this article [1] on the retracted paper. The article says:
"Gebru’s draft paper points out that the sheer resources required to build and sustain such large AI models means they tend to benefit wealthy organizations, while climate change hits marginalized communities hardest. “It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources,” they write."
Meanwhile TFA concludes:
"We are simply urging researchers to consider smaller-scale environments as a valuable tool in their investigations, and reviewers to avoid dismissing empirical work that focuses on smaller-scale environments. By doing so, in addition to reducing the environmental impact of our experiments, we will get both a clearer picture of the research landscape and reduce the barriers for researchers from diverse and often underresourced communities, which can only help make our community and scientific advances stronger."
Looking more closely:
Gebru:
"It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources"
TFA:
"We are simply urging researchers to consider smaller-scale environments... and ... to avoid dismissing empirical work that focuses on smaller-scale environments. ... in addition to reducing the environmental impact of our experiments, we will ...reduce the barriers for researchers from diverse and often underresourced communities"
So yes the similarity is striking. In fact its remarkable that google fired Gebru for questioning the retraction of the paper only to realize that these factors must be considered and then to publish nearly the same message six months later.
Having worked at Google, I think they wanted Gebru to sugar coat the message more. That’s what they’ve done in TFA. But in doing so they’ve also taken away a publishing opportunity from a talented black woman (who they just proved right) and I think that’s a shame.
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by ReadEvalPost on Wednesday July 14, 2021 @02:57AM
(#27830433)
This is not the first Google paper on the question of energy efficiency. The ask was for Gebru to acknowledge existing Google work on the subject in her opinion paper. Her response to this ask is why she was fired, not the contents of the paper.
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by pas on Wednesday July 14, 2021 @02:10AM
(#27830180)
> google fired ____ for questioning ....
This is the Google way apparently. They fired Damore too for pointing out that either the diversity effort is an intended sham, or it's so unscientific that it's useless.
Re:Reducing the Computational Cost of Deep Reinforcement Lea...
by monocasa on Tuesday July 13, 2021 @06:18PM
(#27826961)
She said that they could discuss a possible resignation when she returned from her preplanned vacation. They said don't bother you don't work here anymore.
Threatening to resign is no more resigning than threatening to fire is actually firing.
How does reducing the cost of individual experiments reduce the total environmental impact? Won't more efficient experiments (measured as experiments/Watt-hour or experiments/$) just trigger https://en.wikipedia.org/wiki/Jevons_paradox as more people take advantage of the lower cost?
use p100s hosted in data centers powered by renewables, stop flying to conferences, don’t do drl just because it’s sexy
https://dl.acm.org/doi/10.1145/3442188.3445922
That would explain why such a mundane subject is in a blog post, it's focused at tech news to be all "we really do care about this thing" rather than academia.
That's interesting, GPT-n with 100T parameters would be bored by just reading the whole internet. Too little information, repetitive and on average, junk.
You can try the question in reverse too: Can evolution use too much energy? How much energy has it already consumed?
I'm guessing that's exactly what will happen.
The windmills and solar are still on the way. Looks like the new goal in the US is 2035:
https://www.scientificamerican.com/article/bidens-infrastruc...
[1] https://www.nytimes.com/2017/12/31/technology/google-images-...
[2] https://www.economist.com/science-and-technology/2017/03/02/...
[3] https://www.theverge.com/22309962/timnit-gebru-google-harass...
"Gebru’s draft paper points out that the sheer resources required to build and sustain such large AI models means they tend to benefit wealthy organizations, while climate change hits marginalized communities hardest. “It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources,” they write."
Meanwhile TFA concludes:
"We are simply urging researchers to consider smaller-scale environments as a valuable tool in their investigations, and reviewers to avoid dismissing empirical work that focuses on smaller-scale environments. By doing so, in addition to reducing the environmental impact of our experiments, we will get both a clearer picture of the research landscape and reduce the barriers for researchers from diverse and often underresourced communities, which can only help make our community and scientific advances stronger."
Looking more closely:
Gebru: "It is past time for researchers to prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources"
TFA: "We are simply urging researchers to consider smaller-scale environments... and ... to avoid dismissing empirical work that focuses on smaller-scale environments. ... in addition to reducing the environmental impact of our experiments, we will ...reduce the barriers for researchers from diverse and often underresourced communities"
So yes the similarity is striking. In fact its remarkable that google fired Gebru for questioning the retraction of the paper only to realize that these factors must be considered and then to publish nearly the same message six months later.
Having worked at Google, I think they wanted Gebru to sugar coat the message more. That’s what they’ve done in TFA. But in doing so they’ve also taken away a publishing opportunity from a talented black woman (who they just proved right) and I think that’s a shame.
[1] https://www.technologyreview.com/2020/12/04/1013294/google-a...
This is the Google way apparently. They fired Damore too for pointing out that either the diversity effort is an intended sham, or it's so unscientific that it's useless.
Threatening to resign is no more resigning than threatening to fire is actually firing.