In June 2019, some researchers at the University of Massachusetts had estimated that training and searching of the architecture of a certain neural network required power that would lead to emission of approximately 6.26 lakh pounds of carbon dioxide.
Artificial intelligence: Despite the technological marvel that the artificial intelligence is, it has been at the centre of criticism by many over its sustainability and the carbon footprints. Now, a team of researchers at the Massachusetts Institute of Technology (MIT) have come up with a system to reduce the carbon footprint during the training and running of some neural networks, the institute has said in a press release.
AI: Old system
In June 2019, some researchers at the University of Massachusetts had published a report that estimated that training and searching of the architecture of a certain neural network required power that would lead to emission of approximately 6.26 lakh pounds of carbon dioxide. To put this figure in context, this amount of carbon dioxide is nearly five times the emissions caused by an average car in the US in its lifetime, including the carbon dioxide released during its manufacturing.
Moreover, the situation worsens during the deployment of the model because it needs to deploy deep neural networks on various different hardware platforms, all of whom have different properties and different computational resources.
Now, researchers at MIT have come up with an automated system for AI to train and run certain neural networks. As per their statement, they have found that the system can cut down the emission of carbon, even bringing the figure to low triple digits in some cases, by incorporating some ways in which they have improved the computational efficiency of the system.
How does the system work?
The system, which is called a once-for-all network, undertakes the training of a large neural network, which comprises of many pretrained subnetworks, all of different sizes. These subnetworks can be modified to the different hardware platforms without any need for retraining, the statement said.
This would lead to significant reduction in the energy that would otherwise be needed to train each special neural network for new platforms. Each such platform has the capability to include billions of internet of things (IoT) devices.
If this system is used, the MIT said, for training a computer-vision model, then according to their estimates, the process would require around 1/1,300 the carbon emissions as compared to the presently available state-of-the-art neural architecture state approaches.
The press release also quoted Department of Electrical Engineering and Computer Science Assistant Professor Song Han as saying that the team aims to get neural networks which are smaller and greener. With the new methods devised by this team of researchers, they have been able to reduce carbon footprint by huge amounts, Han added.