The article is from the public number: big data digest (ID: BigDataDigest) , author: Zhao Bilecik, cattle Wan Yang, from the title figure: Oriental IC .

Rain forestIt is burning, the glaciers are melting, and even the Antarctic penguins are experiencing a heat never before.

A while ago, BBC photographers said in an interview during Antarctica shooting that “Here it is very hot.” They saw the impact of global warming on this place, and wept about the future that these animals will face.

The topic of “global warming” has been hotly debated. Recently, the World Meteorological Organization (WMO) released a report saying that the 10-year period 2010-2019 can be said to be “the warmest on record. 10 years”.

The development of artificial intelligence is in full swing, model training time is getting longer and longer, data sets are getting larger and larger, and machine learning continues to be blamed— “has become one of the important sources of ‘carbon emissions’.

A paper in June this year pointed out that machine learning is bringing a huge amount of carbon debt, and the data on carbon dioxide emissions from model training is surprisingly large. According to the relevant calculations in the paper, training and optimization of an advanced conversion model called neural structure search requires approximately 284 tons of carbon dioxide emissions.

Paper link: https://arxiv.org/pdf/1906.02243.pdf

284 tons of carbon dioxide is indeed a huge amount, but should artificial intelligence really be a blessing for “global warming”?

Michael Barnard, chief strategy officer and clean energy expert at TFIE Strategy Inc., made a new assessment of the carbon emissions calculations for this paper and came up with new conclusions. He believes that I, of course, should not ignore the carbon debt brought by machine learning, but considering the reuse value of the model, the carbon liability of machine learning is far worse than estimated in the paper.

(Michael Article Collection: https://cleantechnica.com/author/mikebarnard/)

Let’s take a look back at the June study and the assumptions based on it before making a contrary opinion. This paper was published by Strubell, Ganesh, and McCallum at the University of Massachusetts Amherst in June this year. The paper studied the energy and policy issues related to deep learning in NLP. Next, it was reported by many technology media as headlines.

Although this paper can make the industry reflect on the carbon emissions of artificial intelligence, it ignores an important fact. Neural network models are trained less and used more.

Tesla’s machine learning model is used as an example. Although training the model requires a certain amount of consumption, once it is completed, more than 500,000 cars equipped with this neural network chip currently use this model.

So when considering the “carbon debt” generated by training a neural network, the number of actual executions and the ultimate goal must be considered. If we compare each Tesla car with an oil car, equipped with a neural network chip will definitely improve the operating efficiency of the car, then this type of machine learning application is very worthwhile.

Can carbon liability models reduce carbon emissions?

Michael Barnard believes that good machine learning models can fundamentally change carbon emissions. Previously, Michael Barnard wrote an article exploring the machine learning model CoastalDEM, which is used to determine the risk of sea level rise.

Before CoastalDEM, the traditional method of judging sea level rise has been the SRTM data jointly measured by NASA and the National Institute of Surveying and Mapping (NIMA) of the Department of Defense. The altitude data obtained by this method A positive vertical deviation of 5% will greatly underestimate the risk and exposure of coastal floods.

In October of this year, Scott A. Kulp and Benjamin H. Strauss published a paper in the natural science journal Nature communications titled CoastalDEM: A Global Coastal Digital Elevation Model Improved from SRTM Using Neural Networks. In predicting the risk of sea level rise, although CoastalDEM will bring relatively high carbon liabilities, the prediction results obtained are much more accurate than SRTM.

By 2050, a traditional model predicts the risk of sea level rise in southern Florida by

Updated sea level rise risk chart using CoastalDEM

In this case, CoastalDEM acquired North American satellite radar coastal elevation data, trained it with ground truth from lidar, verified it with Australian lidar, and then expanded its application to the world.

The model has only been executed a few times, but the end result is an adjusted coastal altitude static dataset that has been used globally for policy and climate action plans. In this case, Michael believes that C oastalDEM’s accurate prediction of climate change and the results that can be used repeatedly can bring more value than the carbon liability itself.

Of course, there are many useless models.

Another Michael article evaluates the improvement of start-up Heliogen Concentrated Solar Power (CSP) .

(Article link: https://cleantechnica.com/2019/11/22/heliogen-is-bill-gates-latest-venture-that- is-only-good-for-oil-gas-part-1 /)

According to foreign media CNET, cities around the world continue to need to build more buildings, but the creation of materials such as cement and steel is a huge factor in carbon emissions. Bill Gates-backed startup Heliogen believes that the solution to this problem lies in combining solar energy with artificial intelligence.

Heliogen uses an “advanced computer vision software” to calibrate a “big array” mirror to reflect sunlight onto a single target. The company added that it will eventually be able to generate 1500 degrees Celsius of solar energy, which could produce completely clean hydrogen.

Michael found that while the part of machine learning is interesting and can be reused in other areas, the end result is of little value. Moreover, the case where machine learning improves the heating effect of CSP has not been proven.


The carbon liability of machine learning is not as serious as the paper describes

Let ’s start with the calculations and take a closer look at the assumptions made in the research mentioned at the beginning. (The following sections involve a lot of data discussion, readers who are not interested can skip it as appropriate.)

This paper assumes that the model training assumes CO2e = 0.954pt (0.954 pounds of carbon dioxide equivalent per kilowatt-hour) . This is the average for the United States, and when Michael saw it, he thought it might be overstated.

To this end, Michael first aggregated the current data on the amount of carbon dioxide per kWh.

It can be seen from the figure that the US average masks a huge difference in computing power. A model trained in Washington has a carbon debt that is only one-tenth that of the model trained in Wyoming if it uses a direct grid.

Michael’s FakeSuppose that many of the models in the paper are based in California, and 0.47 pounds of carbon dioxide equivalent per kilowatt-hour from the California power grid is only 50% of the US average.

With that in mind, Michael went a step further. He looked at each of the major carbon liability calculation models in the paper to understand where they were actually trained, assuming that at least one or two would be trained in a Google data center, and Google had 100% renewable commitments and compensation. And the results were far beyond Michael’s expectations.

Carbon liabilities from the models mentioned in different papers and their related training

When you look into the computing resources used in the research, you can find that, except for one case, they are all trained with Google or Microsoft Azure. Columns 3 to 6 are the variance calculations between the estimates in this paper and the possible exact values. To be clear, the NAS Evolved Transformer model can still see 10 tons of CO2e, which is quite considerable, but only accounts for a small part of the research conclusion.

Earlier this year, Michael conducted a rough assessment based on public data. Which cloud computing vendors have carbon debt and how much are they? The evaluation found that among the largest cloud providers, Google and Microsoft Azure have the lowest carbon debt to date. They are committed to 100% carbon neutral power and have purchased high-quality carbon offset products.

The operations of these companies are based entirely on wind, solar and hydropower, which has reduced carbon dioxide emissions per kilowatt-hour to about 0.033 pounds. AWS’s renewable resource rate is actually not good enough, but in 2018, the renewable energy share of its data center still reached 50%, which means that its operating emissions are far below the average level in the United States.

The author of the paper uses another method to assess data center load-Greenpeace report data on the subject in 2017. Although the source is reliable, the data used by the report itself is the electricity actually purchasedThe combined percentage of force sources. Such data will cause all major cloud providers to buy low-carbon electricity much higher than the grid average, but at the same time they still have to buy electricity generated from coal and natural gas.

You don’t need to question Greenpeace’s investigation methods, but Michael finds that there is a big difference between the fact that Google and Microsoft are buying large amounts of renewable electricity and claiming that their data centers use large amounts of natural gas and coal to generate electricity.

Michael and his team believe that Google and Microsoft are buying enough electricity from renewable sources to run their businesses, but Greenpeace did not explicitly point out in the report.

But the biggest problem the paper assumes is not here. The assumption is that because Amazon’s AWS is the most popular cloud computing platform, and its segmented energy consumption calculated by Greenpeace is roughly the same as the overall US segment, the US average is appropriate. However, it can be seen from the evaluation results in the table above that the evaluation model does not use Amazon’s cloud computing platform, so the reliability of one of the results is questionable.

Of course, Michael also pointed out in the end, Although the relevant data should be reconsidered, we should still pay attention to the questions raised by this study.

The dangers caused by global warming are only the tip of the iceberg. Although no one can organize the advancement of science and technology, the development of science and technology and environmental protection must be equal.


Note: Michael states that the study’s lead author has been contacted for comment. If they get a reply, the article will be updated and the digestive bacteria will continue to follow up.

Related reports:

https://cleantechnica.com/2019/11/30/no-machine-learning-does-not-have-a-huge-carbon-debt/

The article is from the public number: big data digest (ID: BigDataDigest) , author: Zhao Bilecik, cattle Wan Yang