GREEN AI : Time for AI to Become Greener

The last few years have seen many new advancements in the field of artificial intelligence(AI), one of the most prominent ones were in areas like object recognition, machine translation, and most importantly the red AI. 

The massive carbon footprint left by the training of today’s resource-hungry machine learning models has captured the main concerns of AI critics.

According to certain reports, it is believed that it has the potential to transform environmental sustainability for good by bringing an algorithmic approach to that work. And so now is the high time for AI to become greener.

Oren Etzioni, CEO of Allen Institute for AI says  "..... The computations required for deep learning research have been doubling every few months, resulting in an estimated 300,000x increase from 2012 to 2018, and these computations have a surprisingly large carbon footprint.

If we don’t want AI research to become electricity-hungry as bitcoin mining, we need to do things differently. ..." A recent study from the Allen Institute for AI proposed a "Green AI” initiative that would focus on the energy efficiency of these future systems.

Here's the link for downloading the research paper by Allen Institute  :

https://arxiv.org/pdf/1907.10597.pdf

“Optimizing efficiency rather than accuracy”. The AI2 researchers classify models as either environmentally friendly “Green AI” or as the (dominant) “Red AI” research, where Red AI seeks to improve accuracy through the use of massive computational power while disregarding the cost —this accuracy, at the cost of measures of efficiency such as speed or energy cost eventuates for the need of a way towards the greener approach.

Although these red projects are leading science towards a major step in natural processing, computer vision, and other computer areas of AI, unfortunately lately these projects are not sustainable. 

So a simple, easy-to-compute efficiency metric that could help make such research greener and more inclusive.

According to the AI2 researchers, green AI refers to research that yields results without increasing computational cost, and ideally reducing it. 

One of the major challenges in the process to achieve greener AI, yielding efficient models with minimal computational cost(hence making “efficiency” as an important evaluation metric), is that there are multiple potential measures of efficiency, each limited in various ways. 

For instance, the amount of carbon emitted by developing a given AI system is hard to measure accurately, and also it largely depends on the local electricity infrastructure. 

The red AI isn't all bad;  the carbon costs may look significant today but this is not an end to this research as researchers believe that it has valuable contributions to the field of AI and hence is on the rise (despite the well-known diminishing returns of increased cost). Of course, both Green AI and Red AI have their contributions to make.

The vision of Green AI is to reduce the computational expense with a minimal reduction in performance or even improve performance with the help of more efficient methods; thereby overcoming the past challenges in Red AI. 

Increasing research activity in Green AI, which is more environmentally friendly and inclusive, aims to shift the balance towards the Greener side.