Hacker News Reader

Comment by ALittleLight | Hacker News Reader

The article points out that the energy to train BERT is comparable to a flight across America. There are hundreds of thousands of flights every day - why should we be concerned about the equivalent of one extra? Especially given that BERT improves ~10% of English language Google search results, it seems like we're getting a lot in return for relatively small energy use. On top of that, Google buys 100% of their energy usage from Green sources.

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.

View on HN

BERT is much smaller than GPT-3 (500x fewer parameters), not sure why the article doesn't point that out more explicitly.

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.

We’re already at soon. Another order of magnitude increase (which at our pace is all but guaranteed in the near term) will make large scale training a significant R&D line item even for the FAANGs. It will also push efforts to be commercially viable, as you can longer do proof-of-concept nets at $500m

> 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.

I think it is reasonable to assume that any advances in efficiency will be subsumed by increased use, much in the same way increased energy efficiency doesn't decrease absolute overall use.

ML hardware is no doubt progressing. Whether we can expect a rapid rate is a different matter. The various Moore's Laws have been breaking down (chip speed stopped doubling a while back). GPU are the sort of chip that could most benefit from just transistors doubling (last Mooore's law standing) but even this Moore's Law is breaking down.

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

So much for the "democratization of AI", I suppose.

> There are hundreds of thousands of flights every day

2019 was peak flight with 38.9 million flights over the year. That averages to 106k flights per day.

My brain does nlp better than any system out there. I’m also able to ride bicycles and do motor control better than Boston Dynamics. I can also construct and prove math, do physics and code all of this in Matlab and C. My brain can handle all these wide range of tasks almost seamlessly with just 15 watts of power, that Silicon Valley’s super computers can barely do 1% of.

To be a fair comparison wouldn’t you have to include all the energy used to train “your” brain through generations of evolutionary training? Your latest model is like taking an already trained BERT model and adding a few tweaks

I see BERT as non sensical. You need to be scientific and have a mathematical theory on how humans learn language, which is a multi disciplinary task requiring physicists, mathematicians, neuroscientists, cognitive psychologists and linguists. Benchmarks are useless, theories, models, experiments and testable predictions is how science progresses. You’re making a comment on cognitive science, and trying to imply that language learning in humans isn’t learned, but pre baked. The psychological, linguistic, evolutionary biology and neuroscience evidence doesn’t seem to corroborate. The evidence points stronger to humans having general learning and problem solving abilities. For instance, there was no evolutionary pressure for humans to be good at math or programming. I was not born knowing english or calculus or probability theory, these were learned abilities. Evolution favoured brain mechanisms that lead to behaviour for success in a rapidly changing world. Had I been born in ancient Rome as a farmer, I would learn to speak Latin, and learn how to be a successful farmer, instead of the physics, math, probability, computer, driving, reading skills that I learned in my life time.

>You’re making a comment on cognitive science, and trying to imply that language learning in humans isn’t learned, but pre baked.

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!’”

> evolution selected for a language skill

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.

It’s a little hard for me to follow your last paragraph but it sounds like you are confusing AI and AGI. I don’t think anybody using BERT is claiming the latter.

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

Only if you're going to include all the time & energy spent creating BERT's precursors as well when calculating its cost.

Good point. I guess from that perspective it’s impossible to quantify either side

To be fair, it takes years of energy to train our brains to the point where it can do all those things, and our brains aren't extensible in the same way hardware/software is. I guess there's also a lot more variation in yields. ;)

Well, that’s not entirely true. Alpha go self played 29 million games of Go, which is in feasible for humans. Assuming a game of Go is 5mins, it would take a human 286 years to achieve the same, assuming this person doesn’t sleep, eat or do anything else other than just play Go. CPU time is an order of magnitude faster than real time, especially on GPU clusters.

Yeah but your clone() function is very wasteful, and we can't stick 1 million of you in a dark room that just looks at people's google searches and figures out what they're really going for.

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)

Not Google search, but many companies are exploiting cheap labor over seas for different industrial use cases because our algos suck. Not just in manufacturing, there are tech companies outsourcing labeled data for tasks like object detection (Hive.ai as anvexample). Whether it makes you depressed or not, humans are cheaper than algorithms, and I don’t see that changing unless we abandon the current paradigm for ai/ml.

Comparing your pre-trained brain (that has a lot of structure (=training) through evolution) with the training costs of a new algorithm isn't really fair.

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.

> Comparing your pre-trained brain

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?

Your ancestors might not have driven cars, but they used the motoric skills required for controlling one, and they had a vision system optimized for detecting and avoiding moving objects.

Human brains are also the culmination of billions of years of iterative development. Computers were only invented in the last century. I would not be surprised if computers could catch up given a few tens or hundred more years.

Yeah, I'd give 30(realistically 15) years tops for AGI. And then we'll call it the end of history.

That is if we don't kill ourselves by making stupid mistakes.

> I'd give 30(realistically 15) years tops for AGI.

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”.

I think this is false. My general opinion of computer scientists and engineers (especially in silicon valley) lack the scientific training that one gets in other disciplines like in physics or other hard sciences. Psychology, linguistics, cognitive science and neuroscience has generated a rich diverse experimental data in the past 50 years, and now is the prime time for a theory to emerge to connect everything. To make what I’m saying clearer, you needed Newtonian mechanics, Maxwell and Plank and Coulomb to develop and discover the science of electricity and atoms in order for us to engineer the silicon transistor. Without the original scientific knowledge, building such devices would have been impossible. This machine learning, ai, deep learning and this obsession with benchmarks are, in my opinion, a hinderance to AGI

I think the error in your thinking is the assumption that scientific discovery is predicated on previously defined theory.

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

> There are hundreds of thousands of flights every day - why should we be concerned about the equivalent of one extra