Instead of a single massive data centre powered by expensive Nvidia chips, Flower Labs' "federated learning" approach can be used to train an AI by using spare chip capacity on small systems distributed around the world, which could include home and office computers, and even smartphones. Image / Getty Creative
Instead of a single massive data centre powered by expensive Nvidia chips, Flower Labs' "federated learning" approach can be used to train an AI by using spare chip capacity on small systems distributed around the world, which could include home and office computers, and even smartphones. Image / Getty Creative
A lot of AI startups are so-called “ChatGPT wrappers” that add a dash of chatbot smarts to existing software. But a startup co-founded by a New Zealander addresses one of the most fundamental elements of the super-hot new technology.
It’s fresh out of the ground: In March last year, justas most of us were becoming aware of ChatGPT for the first time, Nic Lane co-founded a start-up, Flower Labs, that is aiming for a radical shift in the way we “train” artificial intelligence software, a huge and controversial field.
“We believe that artificial intelligence, and the world in general, will be much better off if AI becomes a collaborative activity between many organisations and people,” Lane told the Herald from the UK - where the former pupil of Western Heights High School in Rotorua and Waikato University graduate is now a full professor at Cambridge, leading the university’s Machine Learning Systems Lab. Alongside his academic role, he serves as Flower Labs’ chief scientific officer.
“Right now, it is heading in the opposite direction. AI is dominated by a rapidly decreasing number of companies,” he explained.
“This is because AI training is done in large, centralised data centers that require the data to be all in one place, copied in, and processed by many co-located GPUs.”
GPUs, or graphical processing units, are the super-powerful, super-expensive and super-power-hungry graphics chips - mostly made by Nvidia - used in AI training, or using a collection of data (today, typically the likes of books, or Wikipedia or Reddit content) to school-up an AI and teach it how to speak or write like a human.
(From left): Flower Lab co-founders Nic Lane, Daniel Beutel and Taner Topal. After graduating from Waikato in 2001, machine-learning specialist Lane did a Masters' degree at Cornell then a PhD at Dartmouth. He was lead researcher for an AI lab that Microsoft opened in China before becoming director of the Samsung AI Centre at Cambridge. The Rotorua-raised Lane returns to NZ a couple of times a year "to hang with family and friends".
Lane said a centralised approach “requires a huge amount of money. This is shutting out important stakeholders like NGOs [non-government organisations], government departments and small companies”.
“People think this is the only way to do it. But there are decentralised AI alternatives, like federated learning, that could be used - [though] they have been unstudied, under-utilised and not sufficiently invested in; but the methods do exist to support a future AI landscape where everyone can participate.”
And it’s just that kind of decentralised approach that Lane and co-founders (and Cambridge colleagues) Daniel Beutel and Taner Topal sought to champion.
Flower is a bid to decentralise the AI training process through a platform that allows developers to train models on data spread across thousands of devices and locations. Its “federated learning” approach provides only indirect access to the data to allay privacy and security concerns.
Instead of one pool of data being used to train an AI at a colossal data centre, Flower Labs’ “training at the edge” decentralised, federated approach could utilise, for example, “A handful of GPUs in a hospital IT dept, a cluster of CPUs available in an office environment, or even the increasing amount of computation available in a smartphone.”
Flower pushes an open-source approach - or software that anyone can access, freely, and contribute to. Lane calls it community-driven.
Big-name backers come on board - fast
That sounds like a quixotic or even dare we say it an ivory tower, idealised approach.
But just a couple of months after its launch, Flower landed US$3.6 million ($5.8m) in seed funding in a round led by Spark Ventures, which numbers Eric Schmidt as a partner (Schmidt is famed for being Google’s CEO from 2001 to 2011 - the designated-adult who kept the firm’s wunderkind founders Larry Page and Sergey Brin in line).
And earlier this month, Flower raised US$20m ($32m) at a US$100m ($161m) valuation.
The round was led by Felicis Ventures, a Silicon Valley venture capital firm that was an early backer of Canva, Shopify and Twitch.
While the concept of federated learning has been around for some time, it’s historically been a lot thicker than centralised systems. For three years before Flower’s launch, its founders pursued an academic research project to make it fundamentally easier. Community collaboration was essential to the project.
“To make the barrier to entry to these methods radically lower, we created an open-source ecosystem of tools that were as easy to use as existing ML [machine-learning] tooling - and where there was a global community building prototypes, deploying real systems, experimenting and inventing when that what was needed, and doing this all by working together transparently and in the open.”
NHS, banks, universities among early adopters
There are already more than 3000 developers using Flower’s technology for more than 1100 AI projects.
They come from organisations that span major universities (including MIT, Oxford and Harvard) to the UK’s National Health Service (where a trial saw anonymised data from 130,000 patients used to create better Covid screening) to corporates like Nokia, Porsche, Bosch, Siemens and Samsung (the latter has symmetry; Lane was formerly the director of Cambridge’s Samsung AI Centre at Cambridge). A major bank is using Flower to help detect money laundering. There are also smaller outfits on board like Brave, the creator of a privacy-focused web browser.
“In the end, we think all AI models will need to be trained in this way, because distributed data, scattered in various computing systems in offices, companies, homes etc, is much larger and representative of the real world than the web-scrapped data used to train our AI models today. Models trained in this decentralised manner will be less biased, better able to generalise and overall superior to centralised data centre built AI.”
I can by myself Flower
Lane remains a true believer in free, collaborative software.
“The community and open-source nature of Flower is essential. Users need to completely understand how the software works. Vulnerabilities can be identified much faster and the required level of robustness can be achieved,” he said.
“Progress also happens faster in open-source [contexts] - Flower is not limited to the ideas and abilities of Flower Labs employees but community members working at various companies or in their free time can build key parts to the ecosystem. We expect a lot of progress in how decentralised methods work, and because researchers already use it for their experiments, new discoveries and practices can flow much faster into the wider community.”
But with Series A money in the bag, you also have to start thinking about revenue.
“We believe there are many forms of monetisation models possible that can support the growth and maintenance of the Flower eco-system. For example, commercial usage of Flower-managed services,” Lane said.
Chris Keall is an Auckland-based member of the Herald’s business team. He joined the Herald in 2018 and is the technology editor and a senior business writer.
