‘AI for all’ and ‘the Democratisation of AI’ are current buzz phrases in artificial intelligence. But they can mean different things to different groups.
Generally, there is a two-way split. Commercially, they refer to the drive to develop and release new hardware and software tools that open development and innovation to groups beyond data scientists and algorithm authors. For civil society, they reflect concerns that AI is already being dominated and directed by a few very large companies – typically, the usual suspects of Google, Facebook, Amazon, Microsoft, and their Chinese counterparts like Tencent and Alibaba.
Although it is those same technology giants that are mostly delivering these new tools – along with key hardware players like Nvidia and new strongly backed ones like OpenAI – there is a commercial imperative.
As Kevin Scott, Microsoft’s executive vice president of technology and research, recently observed: “If you look at the early unfolding of the Industrial Revolution, the people who mostly benefitted from the technology were folks who had capital to build factories and businesses around the machines and people who had the expertise to design, build and operate them. Eventually, over very long periods of time, the technology democratised. What we, I think, need to do with AI right now is to try to dramatically contract that period of time where AI is so hard to do that only a handful of people can do it.”
Few would disagree with Scott’s comments. Creating good AI-based applications depends as much on having so-called domain specialists who can identify needs and have the expertise to validate the results as it does on having the engineers who can deliver on their specifications. Too many projects are thought to have failed because only one side of that equation has been fulfilled.
At the same time, concerns over the influence some companies already have seem justified or at least worthy of further exploration. The data here is piling up.
In October 2020, researchers Nur Ahmed, of the Ivey Business School at Western University, and Muntasir Wahed, of Virginia Tech, published an analysis of paper acceptances across 57 of the leading conferences covering AI. They found that a disproportionate amount were provided by the big players and the ‘elite’ universities with which they tend to partner. Mid- and lower-tier universities were, Ahmed and Wahed concluded, being “crowded out”.
Part of the problem, they argued, is that there the current trend in AI is towards a ‘big is beautiful’ approach to developing and delivering the models that underpin applications, which depends on massive processing resources: “The compute-intensive nature of modern AI has resulted in de-democratisation,” the pair said.
On this point, the quartet of Amazon, Facebook, Google, and Microsoft spent $37bn (£27bn) on data centres in the third quarter of 2020 alone, according to the Synergy Research Group. These numbers covered all their activities but AI is acknowledged as one of the primary drivers behind still increasing global investment in ‘hyperscale’ installations.
Meanwhile, there are also growing fears that the sector’s leaders are moving away from a broadly transparent, open-source approach, which has been a hallmark of AI’s more recent evolution. A frequently cited case concerns OpenAI.
It was originally set up as a not-for-profit venture by Elon Musk, former Y Combinator chairman Sam Altman and others, but moved to a commercial model in March 2019, citing the needs to attract staff with higher salaries and stock options and to access venture capital. However, since then, OpenAI has signed an exclusive deal with Microsoft for full source-code access to its GPT-3 language model and marketing of an API for access to it. Its previous GPT-2 model was publicly distributed.
There is a reasonable argument that the models are getting so huge that, for most potential small- to medium-sized users, only cloud-based access is practical – and that carries a cost which whoever provides it must recoup. GPT-3 has 175 billion parameters against 1.5 billion for GPT-2, and it has now been eclipsed by the staggering 530 billion in the recently announced MT-NLG (Megatron-Turing Natural Language Generation) model from Microsoft and Nvidia.
Quite apart from fears about predatory pricing and an AI oligarchy, the more restricted access to these huge models is also raising concerns about their reliability.
This has been openly highlighted as an issue by the MT-NLG team. “While giant language models are advancing the state of the art on language generation, they also suffer from issues such as bias and toxicity. Our observations with MT-NLG are that the model picks up stereotypes and biases from the data on which it is trained,” Microsoft and Nvidia said in a joint statement.
However, given the potential ubiquity of the use-cases to which such models can be put, there are calls for them to be opened to wider scrutiny – even if some restrictions on access may be necessary to prevent their misuse by bad actors.
Moreover, and returning to the wider point about Big Tech’s dominance, there is also the question of the bias not only in the models but also in their developers. While Silicon Valley has made some headway in diversifying the workforce, research shows that, for example, women made up just 10 per cent of Google’s and 15 per cent of Facebook’s AI R&D staff in 2020. Then, returning to the greater role played by ‘elite’ universities, the point is similarly made that their student bodies tend to be disproportionately white, male, and middle-class.
‘AI democratisation means making it possible for everyone to create artificial intelligence systems… potentially without requiring advanced mathematical and computing skills.’
With bias another hot button in AI, ‘Who guards the guards’ maths?’ has become as much of a first-order issue as commercial concerns.
Responding chiefly to the licensing deal between OpenAI and Microsoft, Mark Riedl, professor in the College of Computing at Georgia Tech, has come up with a now widely cited definition of how to bring together the tooling objectives of the big companies with wide concerns about access, transparency, and overly dominated players: “AI democratisation means making it possible for everyone to create artificial intelligence systems. This involves: 1. Having access to powerful AI models. 2. Having access to algorithms. 3. Having access to computing resources necessary to use algorithms and models. 4. Being able to use the algorithms and models, potentially without requiring advanced mathematical and computing skills.”
Riedl’s definition sounds good, but how are groups working towards delivering this in practice? One coalescing framework might be said to combine the technological, educational, and political. In June, the US government announced that its latest AI task force will look to advance the creation of a National Artificial Intelligence Research Resource. The aim is to establish “a system that provides researchers and students across scientific fields and disciplines with access to compute resources, co-located with publicly available artificial intelligence-ready government and non-government data sets”.
The Biden administration has moved in response to an open letter, drafted by the Stanford Institute for Human-Centred AI and signed by multiple institutions, calling for the establishment of a National Research Cloud that would give academia some independence from the growing compute dominance of Big Tech.
In the UK, the government has meanwhile said that it is to “publish a review into the UK’s compute capacity needs to support AI innovation, commercialisation and deployment” by September 2022 as part of the recently released National AI Strategy.
Both moves reflect concern within primarily but not exclusively academia that with the compute demands on AI R&D doubling, according to one measure, every 3-4 months, Big Tech is racing to set the agenda and forcing what is still a relatively immature sector too far down the road of applied research.
At a recent debate of the US Association for the Advancement of Artificial Intelligence, Professor Subbarao Kambhampati, a specialist in AI at Arizona State University, summarised these concerns (albeit with a rhetorical flourish that the format demanded). “We no longer have Bell Labs with Monopoly money pushing very long-term research,” he said. “Unless you think there are ‘deep minds’ at work in industry, let me remind you that they’re very much dependent on the whims and fancies and share prices of the mothership.”
A greater balancing role in compute for government is also inherent in policies that justify a more interventionist role in AI on the grounds that it is as much an issue of national security as potential economic growth.
Beyond the Stanford-led call for a National Research Cloud, we see similar concerns in the recommendations of the US National Security Commission on AI, chaired by former Google CEO Eric Schmidt, and the inclusion of AI cooperation within the new Aukus defence alliance between the US, UK and Australia.
The problem, of course, is whether governments will be able to even come close to matching the billions Big Tech is already investing in compute – they almost certainly will not. But there is hope that the political perception of AI is now such that the cheese will not be pared too much.
However, another strategy is based on making AI less computationally intensive. The not-for-profit research group fast.ai was founded by sector pioneers Jeremy Howard and Rachel Thomas in 2016 with the goal of “making neural nets uncool again”. To do that, Thomas has explained that it seeks to offer wider access for AI to an open-source library built on PyTorch, research into different strategies, and education (though its current training course does require knowledge of Python and a recommended year-worth of coding experience). It also aims to dispel four ‘myths’:
- That you need a PhD in computer science.
- That you need really big data.
- That you need hugely expensive compute resources.
- That you need to understand advanced math.
“Innovation comes from doing things differently, not doing things bigger,” Thomas has said.
Students of the fast.ai approach have achieved some interesting results based on a single GPU (graphical processing unit) including DeOldify, a colourising program for stills and video developed by Jason Antic (deoldify.ai), and a system for diagnosing mastitis udder infections by Canadian goat farmer Cory Spencer.
Notwithstanding the recommended coding skills, there is a focus on allowing those domain specialists to harness the technology’s potential. Thomas says most of fast.ai’s students are also “working professionals learning part-time”.
The founders’ reputations mean that fast.ai has the highest profile in the drive for AI democratisation-through-efficiency. Howard was one of the entrepreneurs behind FastMail and the former president of the Kaggle data science community; Thomas set up the Centre for Applied Data Ethics at the University of San Francisco. But it is by no means the only entry.
The concept of Transfer Learning is, in simple terms, based on using subsets of large datasets to deliver more compute-efficient implementations, and there is also a drive within research to reduce the number of ‘shots’ – or samples – needed to train an AI to recognise something. Research at the University of Waterloo is under way to explore a promising technique that the team there describes as ‘less-than-one-shot-learning’, whereby a system can infer an object based not on examples of it, but from ‘soft labels’ of similar ones. Data-set condensation again promises much greater efficiency and wider access.
Examples such as these are a reminder that AI is still in its infancy. It is only nine years since GPUs emerged as a viable platform for the deployment of neural networks – a concept that may have been originally outlined in the 19th century but which had been beyond the practical processing power available until that recently. As such it is likely to be open to disruption.
If technology may be able to mitigate the emergence of an oligarchy, there is also a social dimension. Teemu Roos is a professor of computer science at the University of Helsinki and the lead instructor for an online course, Elements of AI, that has already been accessed by 1 per cent of the Finnish population – one of those cases where a small-looking number is actually a big one – and 730,000 people worldwide. Roos and his colleagues want to get 1 per cent of the world’s population signed up eventually.
In late 2019, Roos explained to the Fantastic Futures conference at Stanford that Elements of AI is already aimed at everyone, with the goal of combatting “the societal cost of ignorance”.
The course was beta-tested with high school students and takes a maximum of 30 hours to complete, with modules digested at the student’s pace. It chiefly aims to inform the societal debate around AI, explaining what it promises, counteracting myths and then informing how the public views use-cases, political and commercial decision-making, and regulation.
“Among the general public, the awareness of AI is really nowhere where we’d like it to be,” Roos said.
In this case, Elements of AI, jointly developed by Roos’ institution and online teaching company Reaktor Education, does stand out with, so far, little else being done to take AI’s implications for all and not just access out to civil society. Even for those with some background in the field, it is 30 hours well spent – and to broaden debate around the topic, an online community for the course has been added.
Concerns about an ongoing AI land grab should not devalue the work big companies are doing to make AI more accessible, with a necessary initial focus on the enterprise market. At this early stage, AI needs as many contributors as possible. Moreover, even the developers of some of the most advanced models acknowledge that they still need plenty of further work.
It clearly also needs to find a balance – and soon. While there are obvious tensions, as shown by the various paths being taken by those who want to democratise the technology in the way Mark Riedl proposes, there is not yet direct confrontation. Finding ways to accommodate both the needs of Big Tech – and tech of all sizes will always need companies prepared to bet big and provide tools for enablement – and those of the wider research community, government, SMEs, and civil society is thankfully climbing the agenda. *