There is unprecedented growth in global CO2 emissions which is profoundly affecting all aspects of our lives. The culprit is not a single factor, rather there is a wide range of factors which contribute to the increase in emissions such as technological advancements and elevated energy consumption.
Unfortunately, the development of technology aimed at facilitating advancements also has the potential to adversely impact the environment.
Digitalization has been on the rise for the past two decades thanks to the development of high-performance computing technologies. Although these technologies have helped accomplish numerous scientific goals, it has also resulted in an extensive increase in carbon emissions through electricity usage and energy consumption. It is estimated that the energy consumption of digital technologies has increased by around 10% a year.
One technology that is prevalent in all sectors and disciplines is Machine Learning (ML). ML involves the use of data and algorithms to make classifications and predictions. The output of ML is a trained model that helps in decision-making.
The heavy reliance of almost all sectors on ML stems from the fact that it has numerous real-life applications, which has resulted in industrial advancements (hardware and software).
This has encouraged the development of complex and more powerful algorithms. However, all these algorithms require extensive large-scale computations which require substantial energy consumption that causes an increase in CO2 emissions.
A study by the University of Massachusetts around energy consumption of computationally-complex ML algorithms estimated that CO2 emissions from training a single model (Natural Language Processing – a particular application) were larger than those of a vehicle spanning its lifetime. It should be mentioned here that numerous algorithms have to be trained and tested several times before an optimized and ready-to-use model is achieved. This revelation is startling and should persuade researchers to think about “green” computational techniques.
A good start for a paradigm shift to Green ML would be to explore the root cause for developing complex and resource-hungry techniques. One of the reasons for developing complex algorithms is the drive for ever-increasing accuracy. This emphasis on accuracy has overtaken our pursuit of energy-efficient ML algorithms.
Secondly, it is important to investigate the segregated cost and complexity of developing (i.e. design), training (i.e. learning from data), and running (i.e. implementation) the models. It is important to understand that several factors can increase the computational complexity of ML algorithms but all these factors also provide opportunities to develop efficient solutions.
Can We Move Towards Greener Approaches?
Embedding energy efficiency considerations within academic undertakings can be the first step toward bringing change. There are numerous academic programs offering specializations in ML.
As academics, it should be our responsibility to make environmentally-conscious decisions while developing algorithms and applications. While the industry, academics, and researchers push for technological advancements, it is important to observe our energy consumption throughout that process and be aware of the carbon footprint.
All stakeholders engaging with ML should make their best efforts to use an already trained model if possible, understandably with a few tweaks, for another application. Undoubtedly, not every available model will suit every application or scenario. However, there can be many potential overlaps. The reproducibility and transfer of learning can reduce the number of computations needed.
ML complexity influences the choice of hardware as well. It is vital to invest more in the development of efficient hardware as resources (power, energy, electricity usage, etc) can be utilized more responsibly. High-performance and energy-efficient hardware can speed up the learning process and complete the task in a shorter time with less energy and electricity usage thus reducing carbon footprint.
The awareness for developing Green ML protocols is rising, but there is still a lot more work to do. To tackle the persistent challenges of climate change, all stakeholders would need to come together to propose green, efficient, and sustainable solutions.
Extensive large-scale computations might not be the biggest contributor to CO2 emissions but every step counts towards achieving Sustainable Development Goals.
In conclusion, it must be stated that “Green Machine Learning is not just a phrase rather it embodies a significant research goal.
By Dr. Saad Aslam, Senior Lecturer, Department of Computing and Information Systems, School of Engineering & Technology, Sunway University
