Can Federated Learning Save the Planet?

Add to your list(s) Download to your calendar using vCal

• Xinchi Qiu & Nic Lane | University of Cambridge
• Tuesday 27 April 2021, 11:00-12:30
• https://zoom.us/j/6708259482?pwd=Qk03U3hxZWNJZUZpT2pVZnFtU2RRUT09.

If you have a question about this talk, please contact mwlw3.

Machine learning (ML) underpins a wide range of systems that we all use on a daily basis. Every time we perform a web search, use our smartphone or interact with a home assistant, ML is required. However increasingly, the world is recognizing ML often has a considerable environmental cost due the energy used by powerful datacenters where it runs. For example, training a single state-of-the-art NLP model, that processes text, can result in CO2 emissions roughly equal to the lifetime carbon footprint of five cars. More alarmingly, for nearly a whole decade the resources required by ML have been approximately doubling every three months. Already datacenters account for 0.3% of the world’s carbon emissions; but given these existing trends—when combined with the rapidly expanding use of ML in industry, business and personal life—then the environmental consequences of ML must be addressed moving forward, with a level of seriousness and attention that it has never received before.

To this end, we will present our work that examines if Federated Learning (FL) may offer important opportunities in building future environmentally sustainable forms of ML. FL is an emerging technique that enables edge devices (e.g., phones and embedded devices) to collaboratively learn models. The key benefit of FL is traditionally grounded in user privacy, as it allows edge devices to keeping training data on the device rather than it being shared with a datacenter controlled by a third party. But by shifting resource usage from the datacenter to edge devices there are also potential benefits to carbon emissions, for instance, edge devices do not require datacenter-style cooling (that are often 60% of total datacenter energy needs). Our exploratory research considers the type of role FL can play in future low-carbon ML solutions, and seeks to overcome obstacles within FL itself (e.g., high communication overhead) that limit its current suitability. To the best of our knowledge, this is the first time FL has been studied from the perspective of its impact on environmental factors.

This talk is part of the AI4ER Seminar Series series.

This talk is included in these lists:

• AI4ER Seminar Series
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge talks
• Chris Davis' list
• Interested Talks
• Trust & Technology Initiative - interesting events
• bld31
• https://zoom.us/j/6708259482?pwd=Qk03U3hxZWNJZUZpT2pVZnFtU2RRUT09
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.