Ask HN: What is your take on carbon emissions of Machine Learning | Hacker News Reader

Ask HN: What is your take on carbon emissions of Machine Learning

11 pointsdanschmitt10 comments

At the latest since the widely cited article on Green AI [1], there has been increased discussion about the energy consumption of ML. Whenever I talk to colleagues about this, we always agree that there is a dangerous potential lying dormant here, but somehow it cannot be properly grasped. There is simply a lack of really good data. Calculators for emission like from Victor Schmidt and the whole Yoshua Bengio Team at Mila [2] [3] are very good tools for your own use, but can tell you little about the overall situation. So here are two questions for the Hackernews community: 1) How do you assess the topic of ML energy consumption in general and what are factors you consider important? 2) Is it possible to make the energy consumption of ML transparent via legal regulation (such as the currently negotiated AI Act of the EU) and if so, how can something like this look exactly? Or would this be a disaster in the making, because in reality ML is only a small building block of a larger architecture and thus far too laborious to measure for small stakeholder?

[1] https://arxiv.org/abs/1907.10597 [2] https://arxiv.org/abs/1910.09700 [3] https://github.com/mlco2/impact

To cut carbon emissions significantly, we need to move away from fossil fuels.

Anything else (other sources of emission) is simply a distraction from this simple & obvious solution.

As much as I agree, I also wonder if we can switch to sustainable energy sources fast enough. Efficiency of training and inference may be able to make at least a small saving here and thus gain valuable time.

A carbon tax would increase the price of electricity. People could decide if the improved accuracy of a more complicated model is worth more than the extra power costs.

Interesting how the majority of use, if economics is to help us judge, of machine learning is to manipulate people into wasting time or buying certain products and the discussion of energy so far, has no one up in arms against it, and I doubt there are many brave enough here to risk being called a Luddite to do so. Crypto currencies certainly don't get the same nuanced treatment - but perhaps discussion is too filled with loud minorities.

(Context: I work on recommendation systems at a large co - opinions my own)

I think there is a real blind spot in terms of the carbon emissions of ML, somewhat on the research side, but particularly on the production side. Let's take YouTube for example (note: I don't work for YouTube/Google, but I am speculating based on experience with similar complex systems).

A typical YouTube user will interact with a variety of machine learning powered systems:

- Newly uploaded videos to watch (recommendations)

- Recommendations defined by category

- Auto-play next video

- Search

- Ads

- Automated video captions

Each of these is likely a separate team or multiple teams. Each of these will be powered by one or more machine learning models as part of a complex pipeline. These ML models may analyze video content (computationally expensive), audio, user history, etc. Some of these ML models will run in real time, some of them will recompute huge datasets on a daily basis.

What's my point? Very few people have any insight into the carbon impact of all this compute. Perhaps it's negligible, perhaps it's huge, we just don't know. In the last two years, there has been a huge focus on the carbon emissions of PoW cryptocurrencies (e.g. Bitcoin). The impact is real, but I also believe crypto is an easy target. The data is public, hash rate is known and total power use is easy to estimate. The same can't be said for ML, there are a lot of hidden variables and few people are aware of the vast amount of compute that occurs to serve their favourite apps.

Final note: I'm only using YouTube as a common relatable example here. Almost every complex modern service will look similar in terms of number of ML models and downstream carbon impact.

The emissions of ML are not a special category. It's like caring about the emissions of addition vs subtraction, but ignoring every other math operation.

Training is infrequent, and everyone involved cares about making training more efficient and faster.

If you care about emissions, then tax, penalize, and regulate emissions. Everyone doing ML will then look at what's costly, and adjust how they use ML models to fit their budget.

Frankly, ML is the best and most efficient way to solve many important problems. For some problems, it's the only technique that works. Trying to limit ML as though it's a special source of emissions is one of the most counterproductive lines of thinking I've ever come across.

The rate of carbon emission growth is a concern in places like India and China, not from bitcoin or machine learning. Someone should check me on this but I don’t even think getting the entire West to reduce carbon emissions will mean anything unless you can slow down the parabolic rate of India and China.

So, no, machine learning is like at the bottom of the list in terms of concerns.

carbon emissions from deep learning and machine learning will drop once we switch to the new analog hybrid circuitry thats being developed.

Could you explain this in more detail? I am not very familiar with analog hybrid circuits for ML.

basicly, we can do things like matrix operations and differntial equations using specialized hardware that performs computation over analog signals rather than buidlign discrete systems. the end result is that we dont' need anywhere near the level of compute to do the job as its done using an analog of the physical process.

here's two youtube videos that explain is much better than I could. https://www.youtube.com/watch?v=IgF3OX8nT0w&ab_channel=Verit... https://www.youtube.com/watch?v=GVsUOuSjvcg&ab_channel=Verit...

suffice to say, your average house fly exhibits behavior that absolutely dominates our best drones and with a fraction of the circuitry.

Reply on news.ycombinator.com