Daily Management Review

Study: Machine learning is five times more harmful for the environment than a car


Scientists at the Massachusetts Institute of Technology (MIT) decided to find out how much energy is required by advanced machine learning models and how much carbon dioxide they emit in the process of improving their skills. Comparing the data of four models, they concluded that AI training is much more harmful to the environment than production and operation of a car.

Thousands of scientists from different countries and hundreds of companies are working to create a digital counterpart of the human brain - artificial intelligence (AI). Such an idea appeared back in 1956, but no one has yet succeeded in creating a fully-fledged AI that would be able to independently learn and perform a large list of complex tasks - even despite significant technological progress, which allows machines to analyze today terabytes of data at high speeds.

To accomplish the task, scientists will obviously need even more power. However, scientists at the University of Massachusetts in Amherst are convinced that already now the process of creating AI is fraught with serious damage to the environment.

In their work, employees of the College of Computer Science studied work of the four most advanced machine learning programs on neural networks that have appeared in recent years and are focused on recognition of natural languages. These programs have achieved good results in machine translation, tests for the completion of sentences and other typical tests for this field. The selection included Transformer, ELMo, BERT and GPT-2.

Scientists from the University of Massachusetts launched each of these programs on a single graphics processor to find out how much power they consume. Then, from the technical documentation of the models, the number of hours each individual program needed for training was taken into account, which made it possible to calculate the total amount of energy consumed. Based on this, scientists approximately calculated how much carbon dioxide each of the studied AI models produces.

The result was disappointing. Energy consumption and emissions increased in proportion to the power and complexity of the programs. So, the basic transformer model emits only 11.8 kg of carbon dioxide equivalent (calculated on the basis of the average energy consumption structure in the USA). Education of the same program using neural networks and with a large set of parameters emits five times more CO2 equivalent than the average one car in the United States for its entire service life.

Moreover, in the case of a car, the emissions generated by an automobile plant were also taken into account. Thus, the equivalent amount of carbon dioxide emissions from a car is 57.15 thousand kg, and the number is 284 thousand kg for the Transformer neural network model.

These are only initial indicators, the report’s authors note. “To teach one model is the minimum amount of work you need to do,” explained the lead author of the study. Full operation will require a lot of additional training sessions and settings. ” So, the real indicators of harmful emissions into the atmosphere from a working model of AI will be several times greater.