The mentioned paper itself notes a 300,000x increase in compute used for training language models over the last 6 years.
I think the point is, if nothing changes, the costs will be very significant very soon.
The energy efficiency of machine learning hardware is progressing at a rapid rate. It's not accurate to assume that the energy costs will stay the same. Just look at how much it would have taken to train something like GPT-3 five years ago.
ML model size has had it's own Moore's Law, with standard model growing exponentially in size [1]. And this implies model are going butt up against the limits even more than they have already. Whether the "bitter lesson"[2] of ML is true inherently is up for a question. That current researchers have accepted it seems given.
[1] https://openai.com/blog/ai-and-compute/
[2] http://www.incompleteideas.net/IncIdeas/BitterLesson.html
2019 was peak flight with 38.9 million flights over the year. That averages to 106k flights per day.
You make good points, but this one isn’t quite what I meant. I didn’t mean we are trained by evolution for a particular language, but that evolution selected for a language skill. In other words, we are “pre-baked” with the ability to learn a language. It would be analogous to a DL model being trained for generalized regression but not a particular problem. To that extent, I think we’re saying the same thing. Some of the theories related to our generalized learning abilities postulate they stem from this base ability (our aptitude for music, for example, being a consequence of our language learning ability)
>I see BERT as non sensical. You need to be scientific and have a mathematical theory
This is a matter of contention. A lot of science progresses by starting with an empirical result that drives a change in theory rather than the other way around. I can’t remember who to attribute it to, but there’s a quote to the effect that “‘Thats odd’ is the most productive phrase in science, rather than ‘Eureka!’”
Yes it did, and we do know a lot about general principles behind, from different disciplines. This is still an early science. NLP research still hasn’t considered many important aspects of language learning that we have discovered in such a short period of time.
> starting with an empirical result
What makes you think these benchmarks are empirical? They were hand constructed to fit some objective, assuming that being good at the said objective is required for NLP tasks. Where’s the empirical experiments to validate the notion that said objectives lead to language? Science hasn’t worked this way, datasets aren’t constructed, you do an experiment and MEASURE it. Then you make models, try to explain the phenomenon, and test new ideas and validate your models. Can your model extrapolate new information and suggest new experiments to validate? I used the word extrapolate over predict intentionally.
I assume the empirical results are the validation sets run. I.e., the tests that show it provides better results than the base rate. Again, it’s important not to conflate verifiable results with understanding the underpinnings of why it works. If you’re walking and I race you with my car over and over again, I can conclude my car is a faster mode off transportation without understanding anything other than “push right peddle to go faster”. My ignorance doesn’t invalidate the results
And it's not the whole picture to say the brain only uses 15 watts - when there's all sorts of necessary support systems that it couldn't run without. So it's closer to 100 watts (2000kcal/day)
All in you require about 100W (2000kcal/day), maybe 3 times that when doing a lot of physical activity. Boston Dynamic's Spot uses about 400W. I can probably outperform it in some disciplines while it would beat me in some others. That would be a fair comparison.
You’re making sweeping assertions that require domain experts in several different disciplines. The scientific evidence points to the contrary. It is demonstrated that humans have a general ability to learn wide range of things, without being genetically programmed to. None of us evolved to drive cars. Yet nearly everyone in my grandparents generation were able to learn this totally new skill despite being a new invention, where you couldn’t possibly have had time for evolution to act. They weren’t genetically evolved to drive, it was learned within their lifetime and generation. There are tribal humans in different parts of the world that haven’t developed written language. Yet you can teach them written language. Where’s your “pre trained brain” theory there?
That is if we don't kill ourselves by making stupid mistakes.
I think 30 years is probably reasonable, if for “30 years” one reads “twice as long from now as commercially viable fusion generation actually was when first widely hailed as 15 years away”.
Theory is required for understanding, not for discovery. They are two different things. What gets a lot of scientists and statisticians worked up is how ML can outperform traditional models without always being interpretable. It’s like claiming the Wright brothers couldn’t build a flying machine without having a thorough understanding of the mechanics of flight. I can improve performance of my car by remapping the fuel without necessarily understanding the nuances of optimizing enthalpy. There’s levels of understanding, and they may not correlate completely to effectiveness. To that extent, it’s more engineering than science
See also: Sorites fallacy.
There is a limited desire for flights, which benefit people as far as it allows them to get things where they need to be, and no more than that; there is presumably an unlimited desire for machine learning.
https://blog.google/inside-google/infrastructure/data-center...
The more folks who elect to pick where to compute based on low electrical carbon intensity, the faster the grid turns over to clean generation. You must vote with your fiat. I encourage technologists to include this consideration in their workload scheduling requirements. Renewables are almost always cheaper than fossil generation as well.
Meanwhile, the aforementioned transatlantic flight cannot get greener, whether Bert is trained or not.
TFA:
"What does this mean for the future of AI research? Things may not be as bleak as they look. The cost of training might come down as more efficient training methods are invented. Similarly, while data center energy use was predicted to explode in recent years, this has not happened due to improvements in data center efficiency, more efficient hardware and cooling."
Also. One flight in and of itself isn't an environemntal problem; thousands of flights are.
Training this one instance of a model isn't an environmental problem; I would be curious to see some educated guesses about the number of models being trained over the next ten years. Not an expert but I know that use of ML is exploding - lots of new use-cases and thus lots of new models.
So it makes sense to me think about this stuff.
To
There are hundreds of thousands of flights per day. Adding an additional flight, or even an additional thousand flights, to substantially improve Google doesn't seem like a big cost. Consider: Would your life be worse if one random trans-American flight was cancelled today, or if Google searches became 10% worse?
Another way of thinking about this same point, is, why is the author writing about language models if they are so concerned about the environment. Surely the airline industry is a better subject as, again, they fly hundreds of thousands of flights each day. It's hard to take someone seriously when they are focusing on an infinitesimal part of a huge problem.
It's also misleading because there is a difference between energy used by an airline flight, where the energy comes from burning jet fuel, and energy used in a data center. In Google's case (the people training BERT) the energy used was 100% renewable - Google reached that goal in 2017. Perhaps Open AI didn't use renewable energy to train GPT-3, but I wager they didn't power their machines by burning jet fuel either.
Maybe electricity used to power to train language models will become a meaningful issue at some point in the future. I don't think that future is close at hand though.
That's a good point about renewable energy sources.
I'd still argue it's reasonable to assume that ML will consume more power as use-cases for it grow (i.e. more and more models are trained), and, therefore, I don't think it's unreasonable to consider the energy usage.
Honestly - I think what you may have heard is "ML is bad/evil." And yeah some folks probably feel/vibe that.
That's different from "Like any other technology the full context of ML deserves to be considered, and since it's so hyped by billion-dollar companies right now, maybe a bit of pushback on all the hype isn't a terrible idea."
And yeah we should definitely look at the airline industry; pretty sure Mother Jones does, and there's no reason we can't look at the energy impact of multiple industries.
Likewise, yes, language models use energy, but objectively, it's not that much energy and the sources for the energy are sometimes green and always efficient (i.e. Coming from the grid and not by burning jet fuel).
You can't really solve the problem of lung cancer by addressing scented candles. Reducing energy used by language models likewise won't have an effect.
I think it's great to talk about other methods of training or architectures that don't require so many parameters. The point about how BERT consumes vastly more text than humans do when they learn to read is interesting. But trying to phrase this like an environmental issue just seems disingenuous and misleading.