Machine learning (ML) often mimics how human brains operate by attaching virtual neurons with virtual synapses. Deep learning (DL) is a subset of ML putting ‘steroids’ into the virtual brain and growing in magnitude.
The neuron count of virtual brains has skyrocketed together with the advances in computational power. Most headlines about ML relate to it solving hard problems like self-driving cars or facial recognition using DL, but these ‘steroids’ come with an environmental cost.
Increasing the model size increases computational cost and uses more energy, which often leads to a larger carbon footprint. Various sources have highlighted the dizzying carbon emissions of training a bleeding-edge machine learning model.
Recommended AI News: Cyient to Acquire Portugal-Based Celfinet to Strengthen its Wireless Communications Offerings
For example, scholars at the University of Massachusetts Amherst studied the cost of common DL models, which led MIT scientists to the following conclusion:
Training such a model would cost US $100 billion and would produce as much carbon emissions as New York City does in a month.
This simply isn’t true. While these headlines are very important in elevating the discussion around this important topic, the paper overestimated the cost by orders of magnitude. The authors assumed legacy hardware, an average US data centre efficiency, and misunderstood how neural architecture search (NAS) was used.
In reality, a Google investigation later revealed that the model was trained with the latest tensor processing units (TPUv2), used a highly efficient hyperscale data centre, and weaponised NAS smartly by using ‘miniature brains’ as a proxy. The magnitude of error in the original paper was out by 88 times.
Recommended AI News: Getty Images Revolutionizes Visual Communications with VisualGPS Insights Launch
While ML and DL models keep growing, the future is brighter and greener than what some headlines suggest. Advancements in hardware and data science are offsetting the increased costs, and big cloud providers like Google are committed to operating on 100% carbon-free energy by 2030.
While there have been some overestimations in the impact that ML and DL have on the environment that doesn’t mean that companies should not do all they can to reduce their
carbon footprint. Valohai suggests three ways in which companies involved in ML can reduce their carbon footprint:
- Measure early and often. “Companies involved in ML face pressure to reduce their environmental impact, but optimising is impossible if your measurements are blatantly inaccurate or non existent. Everyone should worry about their carbon footprint, but you must make sure that you are measuring it correctly first. By doing so, you’ll start accumulating data insights to make well-informed decisions,” said Eero Laaksonen CEO of Valohai.
- Don’t lock yourself into a single cloud provider. “The big cloud providers are constantly competing for the top spot so it’s important that you remain flexible. It is hard to react if your infrastructure is locked into a single provider. Cloud-agnosticism brings you leverage as it allows organisations to spread the computation cost across multiple providers and choose the ‘greenest’ one depending on the task at hand,” said Laaksonen.
- Invest in compatible technology. “Select technologies that allow you to switch between cloud providers and regions without re-engineering. “Valohai, for example, allows to switch between cloud providers with a single click, enabling ML pioneers to deliver products using the most affordable and ecological options,” said Laaksonen.
Recommended AI News: Huobi Global Unveils Blockchain Ecosystem Investment Arm Ivy Blocks to Spur Industry Development
[To share your insights with us, please write to sghosh@martechseries.com]
Comments are closed.