AI in the UK: ready, willing and able? Contents

Chapter 4: Developing artificial intelligence

130.The UK is currently one of the best countries in the world for researchers and businesses developing artificial intelligence. This cannot, however, be taken for granted. This chapter focuses on the challenges the UK may face in maintaining its position as a word leader, and offers solutions as to how we, as a nation, can ensure that the environment for businesses developing AI continues in good health, and that we have access to the best global talent.

Investment in AI development

131.A number of countries are currently investing significant sums in AI research and development, with varying degrees of state support and co-ordination. Estimating the size of these investments across the public and private sectors is difficult, but according to Goldman Sachs, between the first quarter of 2012 and the second quarter of 2016, the United States invested approximately $18.2 billion in AI, compared with $2.6 billion by China, and $850 million in the UK, the third highest investment by a country in this period.169 While the current United States administration appears to have no national-level AI strategy, individual government departments are continuing to invest in AI, with the Department of Defense, for example, spending approximately $2.5 billion on AI in 2017.170 Meanwhile, China has explicitly committed itself to becoming a world leader in AI by 2030, and aims to have grown its AI ecosystem to $150 billion by then.171

132.Given the disparities in available resources, the UK is unlikely to be able to rival the scale of investments made in the United States and China. Germany and Canada offer more similar comparisons. Of these two, Germany’s approach is strongly influenced by its flagship Industrie 4.0 strategy for smart manufacturing. This strategy seeks to use AI to improve manufacturing processes, and to produce ‘smart goods’ with integrated AI, such as fridges and cars.172 As Professor Wolfgang Wahlster, CEO and Scientific Director of the German Research Center for AI (DFKI), told us, “this is quite different from the US approach, which is based more on internet services”.173 Meanwhile, the Pan-Canadian Artificial Intelligence Strategy, which we discuss in Chapter 9, is less focused on developing AI in any particular sector, but is making C$125 million available to establish three new AI institutes across the country, and attract AI researchers to the country.174

133.Over the course of our inquiry, we were often told of the vitality of the UK AI development sector. The Hall-Pesenti Review recognised this, stating: “The UK has AI companies that are seen as some of the world’s most innovative, in an ecosystem that includes large corporate users of AI, providers large and small, business customers for AI services, and research experts”.175 Investment, as is the case with Canada and unlike in Germany, is not focused in any particular area, and the Government has, for the most part, declined to direct research priorities, with Lord Henley characterising this approach as “[letting] a hundred flowers bloom”.176

134.David Kelnar, Head of Research at MMC Ventures, told us “the number of AI start-ups founded annually in the UK has doubled since 2014, and since then, on average, a new AI start-up has been founded every five days in the UK”.177 The Hall-Pesenti Review estimated that there were more than 200 start-ups and SMEs developing AI products in the UK, as of October 2017.178 Dr Marko Balabanovic, Chief Technology Officer, Digital Catapult, told us that Digital Catapult had found that there were around 600 AI start-ups in the UK out of a total of 1,200 in Europe, putting the UK in a good position.179

135.The UK AI development sector has flourished largely without attempts by the Government to determine its shape or direction. This has resulted in a flexible and innovative grassroots start-up culture, which is well positioned to take advantage of the unpredictable opportunities that could be afforded by AI. The investment environment for AI businesses must be able to cope with this uncertainty, and be willing to take the risks required to seize the chances AI offers.

Box 4: Start-ups and SMEs

‘Start-up’ is a term usually used to refer to businesses which have recently established themselves. They are typically small, financed and operated by their founders, and usually attempt to offer innovative solutions to a shared problem. There is, however, no agreed definition of what a ‘start-up’ is, and companies as large as Uber and WhatsApp often self-define as start-ups.

Medium and small employers are often referred to as small and medium sized enterprises (or SMEs). A business is normally considered to be an SME if it employs between 10 and 249 staff.

Source: Organisation for Economic Co-operation and Development (OECD), ‘Glossary of Statistical Terms’: [accessed 7 February 2018]

136.Eileen Burbidge MBE, Chair of Tech City UK, a Partner at Passion Capital, and a Treasury special envoy for fintech, told us there was a healthy appetite for investing in start-ups in the UK, and that “that there is no shortage of capital at every stage of a life cycle” in Britain.180

137.Other witnesses disagreed, and told us of the difficulties that UK investors face in competing with the largest American technology companies and investors based in Silicon Valley. Libby Kinsey, co-founder of Project Juno, told us: “European investors tend to have smaller funds and less appetite for risk than US investors”.181 Dr David Barber, Reader in Computational Statistics and Machine Learning, UCL, echoed this: “there are a lot of successful tech start-ups which came out of the UK which in the end became successful only when they went to Silicon Valley for investment”.182

138.It was clear from discussions with technology companies on our visit to Cambridge and in the course of our roundtable at techUK (see Appendices 6 and 8) that the appetite for investing in start-ups did exist in the UK, but that the lifecycle and scaling finance was less available. The Royal Society stated “the recent acquisitions of DeepMind, VocalIQ, Swiftkey, and Magic Pony, by Google, Apple, Microsoft, and Twitter respectively, point to the success of UK start-ups in this sector”.183 The Royal Society also, however, voiced disquiet at these acquisitions of UK-based start-ups by foreign-owned companies, stating “they reinforce the sense that the UK environment and investor expectations encourage the sale of technologies and technology companies before they have reached their full potential”.184 The reported cost of these acquisitions ranges from $50–100 million (for VocalIQ) up to £650 million for DeepMind. Professor Michael Wooldridge, Head of the Department of Computer Science, University of Oxford, acknowledged that “there is an incredibly vibrant start-up culture in London”.185 He also warned that “it is fragile and needs to be nurtured”.186

139.James Wise, a Partner at Balderton Capital (UK) LLP, highlighted the specific challenges for AI-focused companies:

“The most challenging area of finance in this field is for spin-outs from academic research between launching the company and getting to a first product. AI start-ups have a longer development period due to the complexity of the software involved, and the need for huge amounts of data, resulting in a ‘Valley of Death’ for start-ups due to the lack of funding before product launch”.187

140.Many of the problems are shared with the wider technology sector in the UK, but the potential of artificial intelligence for the UK’s economy means the challenge in scaling up is even more problematic. Recent Government announcements and the Green Investment Bank (GIB) model offer lessons as to how the Government can use its influence to improve the environment for AI start-ups, and enable them to scale up.

Figure 3: Investment rounds

Graphic showing the levels of funding

141.The GIB was established by the Coalition Government in October 2012 to finance the green economy—an economy that aims to support sustainable development without degrading the environment—and “to accelerate private sector investment, with an initial remit to focus on relatively high-risk projects that are otherwise likely to proceed slowly or not at all”.188 By March 2017 the GIB had invested in 100 projects, committing up to £3.4 billion of its own capital, and had attracted £8.6 billion of private capital. In August 2017, the GIB was sold to Macquarie, a global investment banking and diversified financial services group, for £1.6 billion.

142.The National Audit Office (NAO) reported in December 2017 on the effectiveness of the GIB, finding that “it quickly stimulated investment in the green economy”, in part because its structure as a public company gave it the freedom to pursue its objectives and intentionally constrained its investment activities. The NAO also said that the GIB had invested in, and attracted private capital to, each of its approved sectors. The NAO reported that the responsible Department (then the Department for Business, Innovation and Skills, now the Department for Business, Energy and Industrial Strategy) had not established clear criteria or evidence to judge whether the GIB was achieving its intended green impact. The NAO was also critical of the way in which the sale of the GIB was handled.189

143.This example shows that concerted policy interventions to incentivise private investment for the public good can work, if such policy interventions are committed to.

144.The Autumn Budget 2017 included a series of policy announcements to encourage the growth of innovative firms in the UK. Two of the most relevant were the establishment of a £2.5 billion Investment Fund within the British Business Bank, and proposed changes to the Enterprise Investment Scheme (EIS) and the Venture Capital Trust scheme (VCT). The Rt Hon Philip Hammond MP, Chancellor of the Exchequer, described these measures in the House of Commons as “an action plan to unlock over £20 billion of new investment in UK knowledge-intensive, scale-up businesses”.190

145.The British Business Bank is a Government-owned economic development bank, formed in 2014, with the aim of increasing the supply of credit to SMEs. The Government announced in its November 2017 Budget that the British Business Bank would be responsible for a new £2.5 billion Investment Fund. The British Business Bank said this fund could “unlock up to £13  billion of finance to support UK smaller businesses looking to scale-up and realise their growth potential”.191 Matt Hancock MP highlighted the value of the British Business Bank: “Approximately 40% of capital at some stages in the market is backed in some way by the British Business Bank, so we [the Government] should not underestimate the role that we are playing in this space”.192

146.The EIS is a series of tax reliefs designed to encourage investments in small companies operating in the UK, and has been in operation since 1994. The VCT is another long-standing scheme, designed to encourage people to invest indirectly in a range of unlisted, smaller, higher-risk trading companies, by investing through a VCT instead of directly. The proposed changes to the EIS include doubling the annual allowance for people investing in knowledge-intensive companies (from £1 million to £2 million), and increasing the annual investment such companies receive through the EIS and the VCT (from £5 million to £10 million).193 The EIS will also be altered to focus less on ‘low-risk’ businesses.

147.Finally, one other significant incentive for companies looking to invest in AI R&D are the R&D tax credit schemes. There are two main types of tax relief which companies may be eligible: SME R&D relief, aimed at companies with less than 500 staff and a turnover of less than €100 million or a balance sheet total of less than €86 million, and Research and Development Expenditure Credit (RDEC), for larger companies.194 To be eligible for R&D relief, a company must show how a project:

Under the SME tax relief scheme, a company is allowed to deduct an extra 130% of their qualifying costs from their yearly profit, as well as the normal 100% deduction, to make a total 230% deduction. If the company is loss making, the tax credit can be worth up to 14.5% of the surrenderable loss. Under the RDEC and larger company R&D schemes, a tax credit worth 11% (soon to rise to 12%) of qualifying R&D expenditure may be claimed.195

148.Relatively few of our witnesses mentioned R&D tax credits to us, though at our roundtable event with techUK, some attendees believed the current system unduly benefited larger companies and was not assisting SMEs as much as it could be.196 In 2015–16, 21,865 of the 26,255 claims for R&D tax relief came from SMEs, and the number of SMEs claiming rose by 22% on the previous year.197 Within the Information and Communication sub-category, the single category most directly applicable to AI development,198 5,805 claims were made under the SME R&D scheme, worth £385 million, compared with £145 million, across 355 claims, under larger company schemes.199

149.In total, SMEs have claimed more in R&D tax relief than larger companies. However, 80% of the £22.9 billion in the total qualifying R&D expenditure used to claim R&D tax relief was by companies claiming under large company schemes. This suggests that a sizeable majority of the research and development being incentivised by tax relief schemes is actually being conducted by large companies rather than SMEs.200 Some business leaders have accordingly been critical of the decision to increase RDEC tax relief from 11 to 12% in the 2017 Autumn Budget, without any equivalent increase in the SME scheme. The Institute for Public Policy Research (IPPR) has argued that R&D tax credits and similar schemes for larger companies, which cost £1.8–1.9 billion annually, should be scrapped entirely and redirected towards SMEs.201 Additionally, many AI start-ups are likely to be excluded from claiming under existing rules, as they are experimenting with the application of established AI techniques and methods to new sectors, which is explicitly not covered by R&D tax relief.202

150.We welcome the changes announced in the Autumn Budget 2017 to the Enterprise Investment and Venture Capital Trust schemes which encourage innovative growth, and we believe they should help to boost investment in UK-based AI companies. The challenge for start-ups in the UK is the lack of investment available with which to scale up their business.

151.To ensure that AI start-ups in the United Kingdom have the opportunity to scale up, without having to look for off-shore investment, we recommend that a proportion of the £2.5 billion investment fund at the British Business Bank, announced in the Autumn Budget 2017, be reserved as an AI growth fund for SMEs with a substantive AI component, and be specifically targeted at enabling such companies to scale up. Further, the Government should consult on the need to improve access to funding within the UK for SMEs with a substantive AI component looking to scale their business.

152.To guarantee that companies developing AI can continue to thrive in the UK, we recommend that the Government review the existing incentives for businesses operating in the UK who are working on artificial intelligence products, and ensure that they are adequate, properly promoted to companies, and designed to assist SMEs wherever possible.

Turning academic research into commercial potential

153.The UK has a proven record of producing world-class academic research at globally renowned universities. It has, however, struggled to produce the businesses and commercial success which could flow from this. In addition to this widely recognised problem, artificial intelligence presents its own challenges, in particular the pace and intellectual property involved in AI research and development.

154.This was a problem identified in the Hall-Pesenti Review: “a key component that drives the creation (and success) of new businesses in AI is the ability and capacity for ideas and technologies to spin out of the university network, or be licensed from it, and be commercialised”.203 The Review identified the potential complexity of spin-out practices, and the differing approaches taken by universities make it all the harder for researchers to succeed in commercialising their research. Recommendation 11 of the Review was that “universities should use clear, accessible and where possible common policies and practices for licensing IP and forming spin-out companies”.204

Box 5: What is a spin-out?

A spin-out company is not dissimilar to a start-up, but with the crucial difference that it will often have a minority shareholder, such as a higher education institute, alongside the founders as owners. Spin-outs offer a mechanism for universities to benefit from the research of its staff when they look to apply it commercially.

155.Our witnesses shared these concerns. They told us of the challenges faced by artificial intelligence researchers seeking to commercialise their work. The Royal Society told us:

“ … this standard model [for typical university spin-out companies] may fit less well for machine learning spin-outs. There may not be any IP [intellectual property] per se to be licensed or transferred into a machine learning spinout but rather know-how on the part of the academic founders that is central to the new business”.205

156.David Kelnar agreed that the existing model for spin-out companies was a challenge for AI researchers, and said:

“Universities, typically, seek quite substantial ownership stakes in spin-outs in return for assets, such as patents, the substantial support they offer and the expectation of significant dilution of ownership that will occur over time due to the spinout’s large capital requirements. In the era of AI, though, researchers’ primary assets are more likely to be a little different—it is more a case of their expertise and capability rather than those existing assets”.206

157.Kelnar also identified that limited access to commercial experience and advice was also an issue.207 He told us that this was because there were “not very many commercially experienced AI leaders”.208

158.Witnesses highlighted the work of Imperial Innovations, a subsidiary of IP Group plc (a developer of intellectual property-based businesses, which works exclusively with Imperial College London).209 Imperial Innovations launched Founder’s Choice as a pilot programme for 18 months in August 2017. The aim of the programme was to help reduce the equity share requirements of Imperial Innovations, based on the changing nature of support required for, and by, spin-outs. While the outcome of the programme awaits to be seen, the initiative seems promising.

159.The UK has an excellent track record of academic research in the field of artificial intelligence, but there is a long-standing issue with converting such research into commercially viable products.

160.To address this we welcome, and strongly endorse, the recommendation of the Hall-Pesenti Review, which stated “universities should use clear, accessible and where possible common policies and practices for licensing IP and forming spin-out companies”. We recommend that the Alan Turing Institute, as the National Centre for AI Research, should develop this concept into concrete policy advice for universities in the UK, looking to examples from other fields and from other nations, to help start to address this long-standing problem.

Improving access to skilled AI developers

161.One of the most pressing roadblocks we heard about was the substantial shortfall in skilled workers available to the AI development sector in the UK. Almost all the companies and organisations active in AI development from whom we received evidence complained that developers with advanced knowledge of machine learning, particularly at the PhD and master’s degree levels, were difficult to find, and expensive to hire. Balderton Capital told us that “the skills required to build competitive AI start-ups today are relatively rare, and as a result the costs for starting a company in this space are higher than other areas of technology”.210 The Royal Society also argued that “additional resources to increase this talent pool are critically needed”, with particular emphasis on increased provision for the training of PhD students in machine learning.211

162.At our roundtable event with SMEs, hosted by techUK, we also heard that a number of companies had taken it upon themselves to fund PhD students in machine learning. Drawing in PhD funding from the private sector is indeed encouraged by the research councils, as with the EPSRC’s Doctoral Training Partnership (DTP) scheme, whereby the cost of funding a PhD is split between the research council and a private sponsor.212

163.We were told how the high private sector demand for machine learning expertise risked eroding training pipelines, as the academics needed to train the next generation of talent were being attracted away from universities into private companies. The Future of Humanity Institute at the University of Oxford warned that high salaries, “as well as other benefits of working in industry (such as proximity to other talented researchers and access to large amounts of data and computing power) present a formidable obstacle to the UK Government (and academia) in recruiting AI experts”.213 They suggested that lessons might be learnt from “other domains, such as finance and law, where competition for talent with the private sector has been fierce”, and universities could “consider novel initiatives such as special authority for a department to pay higher than usual salaries”.214

164.It was pointed out that while PhDs in machine learning and AI more widely were important, there was also a need for shorter postgraduate qualifications, such as master’s degrees. The development of new machine learning platforms and tools, such as the open source TensorFlow from Google, means that the level of skill needed to deploy AI in a variety of circumstances, and to facilitate the adoption of AI-enabled services by companies outside the AI development sector, is steadily decreasing.

165.The Royal Academy of Engineering highlighted that there was a “skills gap for people who can work with an AI system but are not AI experts. These people understand the potential of the technology and its limitations and can see how it might be used in business, but are not in a position to advance the state of the art”.215 Furthermore, Research Councils UK stated, “a wider range [of skills] will be needed in the AI workforce as it increasingly overlaps with ethics and social sciences. For example economists are needed for the development of fintech systems, [and] linguists for the development of language processing systems”.216

166.In the Hall-Pesenti Review, it was recommended that the Government and universities should create, at a minimum, an additional 200 PhD places dedicated to AI at leading universities, and develop new master’s level courses in AI, in collaboration with industry, with an initial cohort of 300 students. These recommendations have subsequently been adopted as Government policy, with an announcement that an additional 200 PhD places in AI-related subjects would be funded per year by 2020–22, and plans to work with universities and businesses to develop an industry-funded master’s programme in AI.217

167.When we asked Dr Pesenti, co-chair of the Hall-Pesenti Review, for the rationale behind these numbers, he explained that there had been considerable discussion on this point with civil servants, and noted that they had ultimately been revised down to ensure the recommendation’s sustainability:

“If you look at the demand right now, it needs to be counted in the thousands, quickly, in the next decade for sure. You cannot get there tomorrow because people are not able to be trained. You need to have faculty and fellows, which we also recommended in the review. There was this question: should you put the big number first or should you start with 300? There, we got a lot of back-and-forth”.218

After further questioning, he suggested it was important to think in terms of “tens of thousands” of PhD places within the next decade, but re-emphasised the importance of building up to this number in a sustainable way.219

168.We welcome the expanded public funding for PhD places in AI and machine learning, as well as the announcement that an industry-funded master’s degree programme is to be developed. We do believe that more needs to be done to ensure that the UK has the pipeline of skills it requires to maintain its position as one of the best countries in the world for AI research.

169.We recommend that the funding for PhD places in AI and machine learning be further expanded, with the financial burden shared equally between the public and private sector through a PhD matching scheme. We believe that the Doctoral Training Partnership scheme and other schemes where costs are shared between the private sector, universities and research councils should be examined, and the number of industry-sponsored PhDs increased.

170.We further recommend that short (3–6 months) post-graduate conversion courses be developed by the Alan Turing Institute, in conjunction with the AI Council, to reflect the needs of the AI development sector. Such courses should be suitable for individuals in other academic disciplines looking to transfer into AI development and design or to have a grounding in the application of AI in their discipline. These should be designed so as to enable anyone to retrain at any stage of their working lives.

Diversity of talent

171.Our witnesses raised questions over the diversity of those working in AI development. In the early decades of the computer industry there was once a significant proportion of female workers. Unfortunately this is no longer the case, and some of our witnesses spoke of the need to “democratise AI”, and address what Bill Gates has described as the “sea of dudes” problem, with mainly male attendees at AI conferences.220 PwC told us of research they had conducted, which found that only 27% of female students they surveyed said they would consider a career in technology, compared to 61% of males, and that only 3% of females said it would be their first choice. 78% of students surveyed could not name a famous woman working in technology, compared to two thirds that could name a famous man working in technology.221 As a consequence, according to Liberty, only 7% of students taking the computer science A-Level, and 17% of those working in technology in the UK, are female.222

172.Ensuring that those from low-income households and disadvantaged socio-economic backgrounds can still participate in the development and adoption of AI was also raised as part of wider efforts to facilitate social mobility.223 During our meeting with representatives of UK AI start-ups, we heard some scepticism regarding the inflexibility of the apprenticeship levy with regards to the AI development sector. Paul Clarke, Chief Technology Officer at Ocado, told us that the apprenticeship levy had “carved a hole in the available budget that companies, including ours, have to spend on that continual learning”, and urged that it be converted into a less ring-fenced “training levy”.224 However, at least one company in attendance at the techUK roundtable told us of their success using apprentices in their company. PwC also informed us of their recently announced technology degree apprenticeship scheme, beginning in September 2018 with 80 students splitting their time between study for a degree in Computer Science and work for PwC in Birmingham and Leeds.225

173.Gender, ethnic and socio-economic diversity are important for a variety of reasons. Careers in AI are well remunerated and an area of rapid growth, and the dominance of these positions by already privileged groups in society is likely to exacerbate existing inequalities further. But this lack of diversity also has a significant impact on the way that AI systems are designed and developed. If we are to ensure that these systems, which are exerting growing influence over our lives and societies, serve us fairly rather than perpetuate and exacerbate prejudice and inequality, it is important to ensure that all groups in society are participating in their development. As CognitionX put it, “one of the reliable ways we know we can mitigate [the problem of bias and discrimination] is to have more diverse development teams in terms of specialisms, identities and experience”.226 Companies are making efforts to address this issue, and we are aware that many issues need to be tackled within primary and secondary education, which we address later in this report.227 Nevertheless, we believe there are still measures the Government can take to address this problem in the short term.

174.We recommend that the Government ensures that publically-funded PhDs in AI and machine learning are made available to a diverse population, more representative of wider society. To achieve this, we call for the Alan Turing Institute and Government Office for AI to devise mechanisms to attract more female and ethnic minority students from academic disciplines which require similar skillsets, but have more representative student populations, to participate in the Government-backed PhD programme.

175.We acknowledge the considerable scepticism of at least some technology companies who believe that the apprenticeship levy is of little use to them, despite the success that others in the sector have had with apprenticeships. The Government should produce clear guidance on how the apprenticeship levy can be best deployed for use in the technology sector, in particular in SMEs and start-ups.

Immigration and overseas skilled workers

176.While a majority of witnesses believed measures, such as the funding of additional PhD places, should be taken to develop the UK’s home-grown AI talent, this was generally seen as a long-term solution, which would not address shortages for some years to come. Many witnesses highlighted the importance of overseas workers to the UK AI development sector, and voiced concerns that this supply could be jeopardised by illiberal immigration policies, especially in the wake of Brexit. The think tank Future Advocacy observed that:

“In the current climate of uncertainty, there has already been a sharp decline in EU applications to UK tech jobs. There are 180,000 EU workers in the tech sector but the UK Government is yet to confirm new visa rules for EU workers after Brexit. If these workers left the UK it would tear open the already vast ‘skills gap’”.228

177.Those with machine learning expertise are globally sought after, and therefore constitute a highly mobile population. Eileen Burbidge emphasised that the “uncertainty (of Brexit) alone raises issues and gives people pause before they consider coming to the UK, or challenges for companies trying to recruit outside the UK”. Professor Wooldridge warned that Brexit “could quite genuinely be the death knell for UK tech start-ups, which are heavily reliant on overseas talent”.229

178.A 2017 assessment by Tech City found that non-UK workers made up 13% of the digital technology workforce, compared to 10% in the wider economy in 2015. Interestingly, non-EU workers accounted for a larger share of employment (7%) in the digital technology industries than EU workers (6%)—however, employment for EU nationals had grown faster than for non-EU nationals, growing by two percentage points over the five years from 2011 to 2015.230

179.Many witnesses called on the Government to confirm the position of EU technology workers after Brexit, and to liberalise the visa regimes for overseas technology workers in the UK more generally. Balderton Capital suggested that special measures should be taken: “special visas attached to students studying in this field could be considered to make sure they have the ability to remain in the UK while studying and afterwards when starting companies”.231 Eileen Burbidge also asked that the Home Office Migration Advisory Committee consider adding artificial intelligence-related roles to the Tier 2 Shortage Occupation List, and that the quota on Tier 1 (Exceptional Talent) visas be increased.232

180.On 15 November, shortly after we spoke with Eileen Burbidge, the Government did indeed announce that it would be doubling the number of Tier 1 (Exceptional Talent) visas from 1,000 to 2,000 a year. The 2,000 visas will be made available to individuals who are “recognised as existing global leaders or promising future leaders in the digital technology, science, arts and creative sectors” by one of five UK endorsing organisations:

181.The Government’s announcement that it will increase the annual number of Tier 1 (exceptional talent) visas from 1,000 to 2,000 per year is welcome. While top-tier PhD researchers and designers are required, a thriving AI development sector is also dependent on access to those able to implement artificial intelligence research, whose occupations may fall short of the exceptional talent requirements.

182.We are concerned that the number of workers provided for under the Tier 1 (exception talent) visa scheme will be insufficient and the requirements too high level for the needs of UK companies and start-ups. We recommend that the number of visas available for AI researchers and developers be increased by, for example, adding machine learning and associated skills to the Tier 2 Shortage Occupations List.

Maintaining innovation

183.While deep learning has played a large part in the impressive progress made by AI over the past decade, it is not without its issues. Some of our witnesses believed it would not continue to deliver advances at the current rate, and that other avenues of research needed more support. Deep learning requires large datasets, which can be difficult and expensive to obtain, and can require a great deal of processing power. As a number of witnesses emphasised, recent advances in deep learning have been made possible with the growth of cheap processing power.234 Some indications suggest, however, that Moore’s law—the observation that the number of transistors on a circuit board tends to double every two years—is starting to break down, with the growth in cheap processing power slowing.235 Innovations such as quantum computing may yet restore, or even accelerate, the historic growth in cheap processing power, but it is currently too early to say this with any certainty.236

184.Several of our witnesses also pointed to the difficulties of transfer learning, or “the ability of computers to infer what might work in a given scenario based on knowledge gained in an apparently unrelated scenario”.237 This can also be thought of as common sense in human beings, which cannot currently be replicated in any AI system.238

185.The Foundation for Responsible Robotics believed it could be the case that “once all the low hanging fruit has been picked, severe limitations will be found and then the technology will plateau”.239 Geoff Hinton, a pioneer in deep learning who is still revered in the field today, has warned that the deep learning revolution might be drawing to a close.240 Others were more optimistic. Witnesses told us of advances in custom-designed ‘AI chips’,241 such as Google’s Tensor Processing Unit, which trades general-purpose processing power for extra power in AI applications and, more speculatively, advances in quantum computing which might provide further boosts in future.242 In terms of transfer learning, DeepMind has had success in applying its AlphaGo AI, originally trained on Go, to chess, but for the most part this is still challenging.

186.Some of the most innovative AI research we observed went beyond deep learning, and combined it with other aspects of AI, creating hybrid systems which sought to compensate for the drawbacks of any one style of AI. For example, Prowler.IO, an AI start-up in Cambridge, told us of their own issues with deep learning, in particular the amount of data it often requires, and the lack of transparency behind the decisions it comes to. They outlined their own approach to us, which combines a range of approaches, including probabilistic modelling, multi-agent systems and reinforcement learning to create more robust AI which can cope with less data and more uncertainty. A number of other AI experts have also suggested that combining different approaches, some of which were once prominent within the AI field but have since fallen out of fashion relative to deep learning, may well be a productive way forward.243

187.Google told us that the success of AI in the UK was in considerable part due to the Government’s traditional role in “supporting long term fundamental research”, and that this role should continue.244 Professor Wolfgang Wahlster also emphasised the UK’s role in pioneering AI research.245 It will also be important for the UK to continue to participate in European research and innovation programmes such as Horizon 2020, and its successor Framework 9. We welcome the Government’s commitment to underwrite any bids for Horizon 2020 projects while the UK is still a member of the EU, and we hope that this continues, where possible, after we leave the European Union.246

188.We believe that the Government must commit to underwriting, and where necessary replacing, funding for European research and innovation programmes, after we have left the European Union.

189.We should also consider this approach to research in the light of the evidence we received on deep learning in relation to other aspects of AI. As Jonathan Penn, a historian of AI at the University of Cambridge, told us, AI has long been a varied and even incoherent field at times, with different sub-disciplines constantly vying for attention and funding. Neural networks were, for example deemed a “sterile” area of research by Marvin Minsky, a view shared by many in the discipline for many decades.247 Geoff Hinton renewed interest in neural networks just as most AI researchers were fully investing in the now largely defunct area of expert systems.

190.Furthermore, in order to track and assess the likely impact of AI on the economy, work, politics, health care and medicine, education and other fields, it will be crucial for experts from different disciplines to work closely together. In particular, researchers specialising in AI will need to collaborate with those studying other academic areas. Institutes such as the Leverhulme Centre for the Future of Intelligence at the University of Cambridge, and the Oxford Internet Institute at the University of Oxford are excellent, existing, examples of this collaboration, and universities across the UK should encourage the development of their own such centres.

191.The state has an important role in supporting AI research through the research councils and other mechanisms, and should be mindful to ensure that the UK’s advantages in AI R&D are maintained. There is a risk that the current focus on deep learning is distracting attention away from other aspects of AI research, which could contribute to the next big advances in the field. The Government and universities have an important role to play in supporting diverse sub-fields of AI research, beyond the now well-funded area of deep learning, in order to ensure that the UK remains at the cutting edge of AI developments.

169 Goldman Sachs, China’s rise in artificial intelligence (31 August 2017), p 6: [accessed 23 February 2018]

170 Govini, Department of Defense: artificial intelligence, big data and cloud taxonomy (December 2017): [accessed 7 March 2018]

171 Indeed, a number of commentators have pointed out that China’s AI strategy appears to have been influenced by the Obama administration’s AI strategy, published in late 2016, which has since become dormant under the Trump administration. Cade Metz, ‘As China marches forward on AI, the White House is silent’, New York Times (12 February 2018): [accessed 13 February 2018]

172 Q 164 (Professor Wolfgang Wahlster)

173 Ibid.

174 CIFAR, ‘Pan-Canadian artificial intelligence strategy overview’ (30 March 2017): [accessed 21 February 2018]

176 Q 191 (Lord Henley)

177 Q 48 (David Kelnar)

179 Q 42 (Dr Marko Balabanovic)

180 Q 48 (Eileen Burbidge)

181 Q 49 (Libby Kinsey)

182 Q 42 (Dr David Barber)

183 Written evidence from Royal Society (AIC0168)

184 Ibid.

185 Q 3 (Professor Michael Wooldridge)

186 Ibid.

187 Written evidence from Balderton Capital (UK) LLP (AIC0232)

188 HC Deb, 24 May 2011, cols 789–790

189 National Audit Office, The Green Investment Bank (12 December 2017): [accessed 11 January 2018]

190 HC Deb, 22 November 2017, col 1049

191 British Business Bank, Budget 2017 (23 November 2017): [accessed 12 January 2018]

192 Q 194 (Matt Hancock MP)

193 HM Treasury, Autumn Budget 2017 (November 2017), p 49: [accessed 17 January 2018]

194 HM Revenue & Customs, ‘Guidance: Research and development (R&D) tax reliefs’ (14 August 2017): [accessed 14 February 2018]

195 Ibid.

196 Written evidence from Bikal (AIC0052), BSA The Software Alliance (AIC0153) and techUK (AIC0203)

197 In addition, 1,770 SMEs also claimed for research under the RDEC scheme and large company R&D scheme, which is permissible if a larger company sub-contracts research to an SME. HM Revenue & Customs, Research and Development Tax Credits Statistics (September 2017): [accessed 22 February 2018]

198 It should be noted that, due to the far-reaching nature of AI technology, in many cases it could also fall into other sub-categories of R&D.

199 Additionally, a total of £20 million was claimed by SMEs under larger company tax relief schemes.

200 SMEs receive more in tax relief, even though much more qualifying R&D is conducted by large companies, because incentives for each individual small company are substantially more generous than under the large company schemes. HM Revenue & Customs, Research and Development Tax Credits Statistics (September 2017): [accessed 22 February 2018]

201 Zen Terrenlonge, ‘Autumn budget 2017: R&D tax credits and investment to propel UK into the future’, Real Business (22 November 2017): [accessed 12 February 2018]; IPPR, Industrial strategy: Steering structural change in the UK economy (November 2017):–11/1511445722_industrial-strategy-cej-november17.pdf [accessed 7 March 2018]

202 HM Revenue & Customs, ‘Guidance: Research and development (R&D) tax reliefs’ (14 August 2017): [accessed 14 February 2018]

205 Written evidence from Royal Society (AIC0168)

206 Q 50 (David Kelnar)

207 Ibid.

208 Ibid.

209 Q 50 (David Kelnar) and written evidence from Balderton Capital (UK) LLP (AIC0232)

210 Written evidence by Balderton Capital (UK) LLP (AIC0232)

211 Written evidence from the Royal Society (AIC0168)

212 EPSRC, ‘Doctoral Training Partnership’: [accessed 10 January 2018]

213 Written evidence from Future of Humanity Institute (AIC0103)

214 Ibid.

215 Written evidence from Royal Academy of Engineering (AIC0140)

216 Written evidence from Research Councils UK (AIC0142)

218 Q 203 (Dr Jérôme Pesenti)

219 Q 212 (Dr Jérôme Pesenti)

220 Written evidence from Dr Huma Shah and Professor Kevin Warwick (AIC0066)

221 Written evidence from PricewaterhouseCoopers LLP (AIC0162)

222 Written evidence from Liberty (AIC0181)

223 Written evidence from Vishal Wilde (AIC0004)

224 Q 107 (Paul Clarke)

225 Written evidence from PricewaterhouseCoopers LLP (AIC0162)

226 Written evidence from CognitionX (AIC0170)

227 Google (AIC0225) and PricewaterhouseCoopers LLP (AIC0162)

228 Written evidence from Future Advocacy (AIC0121)

229 Q 47 (Eileen Burbidge) and written evidence from Professor Michael Wooldridge (AIC0174)

230 Tech City, ‘The nationality of workers in the UK tech industry – Tech Nation Talent: Part 1’: [accessed 31 January 2018]

231 Written evidence from Balderton Capital (UK) LLP (AIC0232)

232 Q 54 (Eileen Burbidge). In the UK immigration system, Tier 1 (Exception Talent) visas are for people recognised by the Home Office as a leader (or emerging leader) in their field. Tier 2 visas allow skilled workers to enter the UK on a long term basis to fill a skilled job vacancy (defined by an occupation list). Tier 2 jobs must usually be advertised to workers from within the European Economic Area (EEA) before they can be offered to those from outside the EEA. The Tier 2 Shortage Occupation List is a list of occupations which UK employers have not been able to recruit, and the jobs on this list can be offered to people from beyond the EEA, without being advertised within the EEA first.

233 Home Office, ‘Government doubles exceptional talent visa offer’ (15 November 2017): [accessed 7 February 2018]

234 Written evidence from Capco (AIC0071); Dr Toby Walsh (AIC0078); Economic Singularity Supper Club (AIC0058) and BioCentre (AIC0169)

235 Tony Simonite, ‘How AI Can Keep Accelerating After Moore’s Law’, MIT Technology Review (30 May 2017): [accessed 8 February 2018]; Mark Pesce, ‘Death notice: Moore’s Law. 19 April 1965–2 January 2018’, The Register (24 January 2018): [accessed 8 February 2018]

236 George Musser, ‘Job one for quantum computers: boost artificial intelligence’, Wired (10 February 2018): [accessed 14 February 2018]

237 Written evidence from Royal Academy of Engineering (AIC0140) See also Touch Surgery (AIC0070)

238 Written evidence IBM (AIC0160); Q 2 (Professor Nick Bostrom) and The Reverend Dr Lyndon Drake (AIC0108)

239 Written evidence from Foundation for Responsible Robotics (AIC0188)

240 James Somers, ‘Is AI riding a one-trick pony?’, MIT Technology Review (29 September 2017): [accessed 8 February 2018]

241 Written evidence from Research Council UK (AIC0142) and Deep Learning Partnership (AIC0027)

242 Written evidence from BioCentre (AIC0169) and Economic Singularity Supper Club (AIC0058)

243 Richard Waters, ‘Why we are in danger of overestimating AI’, Financial Times (5 February 2018):–59efdb70e12f [accessed 6 February 2018]

244 Written evidence from Google (AIC0225)

245 Q 168 (Professor Wolfgang Wahlster)

246 Department for Exiting the European Union, Collaboration on science and innovation: a future partnership paper (6 September 2017): [accessed 1 March 2018]

247 Written evidence from Jonathan Penn (AIC0198)

© Parliamentary copyright 2018