1.1 Google DeepMind makes this submission to the committee as part of the inquiry on artificial intelligence and robotics. We write with reference to all four parts covered in the committee’s terms of reference.
1.2 We welcome the Science and Technology Committee’s inquiry into the potential of artificial intelligence to benefit the United Kingdom, and appreciate the opportunity to provide input based on our experience.
1.3 Google DeepMind is a British artificial intelligence company founded by Demis Hassabis, Shane Legg and Mustafa Suleyman in 2010. The algorithms we build are capable of learning for themselves directly from raw experience or data, and are designed to be ‘general’ in that they can perform well across a wide variety of tasks straight out of the box. Our world-class team consists of many renowned machine learning experts in their respective fields including, but not limited to, deep neural networks, reinforcement learning and systems neuroscience. While the Committee’s inquiry is looking at both AI and robotics, we will focus purely on the former and not on robotics as that is not our area of expertise.
1.4 In the announcement of this inquiry, mention was made of the recent historic Go match between the World Champion Lee Sedol and our program AlphaGo. The game of Go is the most complex game mankind has devised, and was widely viewed as an unsolved “grand challenge” for artificial intelligence. Despite decades of work, the strongest computer Go programs still only played at the level of human amateurs. On 28th January 2016, we published a Nature paper that describes the inner workings of AlphaGo. This program was based on general-purpose AI methods, using deep neural networks to mimic expert players, and further improving the program through learning from games played against itself.
1.5 The most important thing about AlphaGo is not so much what it does, but the way it does it. Although the AlphaGo system can’t for the moment do anything besides play Go, our plan is to extend the techniques developed in the process to one day be applied to important real-world problems that are similarly complex and long range (e.g. climate modelling or complex disease analysis). Artificial intelligence, with the right approach, will be able to make significant leaps in what we as a society are able to achieve, especially as we grapple with increasing volumes and complexity of data sets. It is the opportunity to complement and enhance our human decision-making that offers the most potential for benefit in the long term.
2. The implications of robotics and artificial intelligence on the future UK workforce and job market, and the Government’s preparation for the shift in the UK skills base and training that this may require.
2.1 The advent of new technologies has always helped shape the employment landscape, and we should expect that increased use of AI and machine learning will be no different. In many sectors, machine intelligence will augment and enhance the work that people do, enabling them to be more effective in the same roles. As with all technological innovation, we should expect that new areas of economic activity and employment will be made possible, and some types of work and some skills will decrease in relevance. It is important that government focus on investment in the digital and creative skills that will support a strong UK economy as these technologies develop and mature.
2.2 At this point one of the most important steps we must take is to ensure that current and future workforces are sufficiently skilled and well-versed in digital skills and technologies, particularly STEM subjects. The UK government has been proactive and vocal in support for digital education, such as introducing computer science into the curriculum from 2014 but it is important not to be complacent about the leaps that are needed.
2.3 A digital skills gap in the UK has been widely recognised by organisations including the British Chamber of Commerce, Tech UK, and the Tech Partnership. Go ON UK produced a heatmap of geographical digital exclusion showing the highest levels of Basic Digital Skills are in Greater London (84%), Scotland and East Anglia (both 81%) with the lowest levels in Wales, where only 62% of adults have the five Basic Digital Skills they need.
2.4 We can address the digital skills gap by focussing on education, teacher supply, adult skills and digital inclusion. For example, while a renewed focus on coding in the curriculum is strongly welcomed, it is important also that teachers are fully trained in how to deliver it. Likewise, recruitment of teachers for non-core subjects such as computing is critical. Google has announced a partnership with Teach First to help support and train the next generation of Teach First computing teachers specifically to address the acute teacher shortage in this area, but the full scale of the problem will clearly need larger scale investment to tackle completely.
2.5 It is also important that the UK is able to harness the talents of the widest pool available, which means putting real effort into encouraging more women into technology, focussing on adult digital literacy as well as youth education, and enabling the next generation of entrepreneurs no matter their socio-economic background.
2.6 For example, it is clear that the technology industry faces a problem of gender disparity that can be traced back to the relatively small numbers of girls who take up STEM subjects at school and university. It is for this reason that Google DeepMind is working on a programme to encourage more women into machine learning, but initiatives like this alone are not enough. We welcome the work of Martha Lane Fox and Doteveryone in enabling technologies that advantage all British citizens.
3. The extent to which social and economic opportunities provided by emerging autonomous systems and artificial intelligence technologies are being exploited to deliver benefits to the UK.
3.1 As we detail below, the UK is a world-leader in artificial intelligence and machine learning, both academically and in industry, and the need to maintain and extend that lead is clearly in the interests of national economic prosperity at a macro level.
3.2 In everyday terms, the benefits of machine learning and AI are already being felt across many aspects of Google’s products that UK citizens find useful in their everyday lives, from translation tools to getting rid of spam from their email inbox and suggesting smart replies.
3.3 DeepMind’s mission is to solve intelligence and in doing so develop technologies that help society tackle some of its toughest problems, like science and healthcare. One of the key reasons it is hard to make progress on these big challenges is that even the smartest humans struggle to fully understand the relationships between cause and effect in these systems. Scientists are overwhelmed by the complexity of interacting factors and volume of information. Machine intelligence may help to model and better understand this complexity, and in turn allow us to design more effective interventions.
3.4 However, this data is also narrower in scope than the rich diversity of human experience. It’s still going be many, many decades before AI can begin to factor in the kind of nuanced social and cultural context to its perceptions that humans rely upon to make reasoned judgements. This is why it’s important that we use AI as a tool to augment and enable human expertise and insight, rather than seeing AI as a replacement for human decision-making.
3.5 We envisage machine learning systems being designed as tools that complement and empower the smart and highly motivated experts working in such fields, by enabling efficient analysis of large volumes of data, extracting insights and providing humans with recommendations to take action. This could be in areas ranging from early diagnosis of disease, discovery of new medicines, advances in materials science or optimising use of energy and resources.
3.6 We strongly believe that technology interventions should be developed in conjunction with existing experts in the field, which is why DeepMind Health is working with clinicians to develop technologies that present timely information to clinicians and facilitate provision of care. Over time we envisage exploring healthcare technologies that make direct use of machine learning, but we wanted to start with relatively simple tools that clinicians felt could make a massive impact to patient care and in doing so prepare the ground for more sophisticated technologies where clinicians see the most benefit.
4. The extent to which the funding, research and innovation landscape facilitates the UK maintaining a position at the forefront of these technologies, and what measures the Government should take to assist further in these areas.
4.1 DeepMind was founded in the UK, which is now a world-leader in artificial intelligence research and stands to benefit significantly from continued progress and investment in this area. With further steps, the UK is poised to secure its place at the forefront of AI research and innovation. We make our recommendations below for continued investment in the research but in order for the AI that powers the apps we rely on everyday to continue to flourish there are also two core commodities needed: secure access to data for research and a secure serving infrastructure for that data.
4.2 We are a part of the UK’s artificial intelligence research community, which includes many leading centres of research at universities such as UCL, Oxford, Cambridge, Imperial, and Edinburgh, as well as new and innovative research bodies such as The Alan Turing Institute. We support and engage with the research community through PhD sponsorship, lectures, conferences and papers, to date we have listed over 60 publications on our website.
4.3 To support the UK research community, we recommend that government engage with cutting-edge researchers by convening an advisory panel of both academic and industry experts. This panel would determine research funding priorities and directions with an emphasis on transparency and accountability, and feed these through to research councils and other funding bodies.
4.4 As with funding of direct machine learning research, an advisory panel of machine learning experts should provide direction to educational institutions and funding bodies on broad investment in machine learning skills development.
4.5 The government should also consider funding for machine learning masters and PhD programmes at British universities, to encourage more research in the field and nurture the next generation of scientists who will help preserve the UK’s preeminent position. This funding could also include direct support for modules within programmes that train machine learning researchers in the ethics of data science and increasingly autonomous decision-making, to ensure that the pursuit of beneficial outcomes is embedded in the science of machine learning at every level.
4.6 Machine learning technologies benefit not only from large volumes of data, but also the right types of data, for innovation and research. At DeepMind we have made extensive use of simulated environments allowing significant research without access to public datasets, and, where possible, funding research to produce more sophisticated and versatile simulated environments would support research progress.
4.7 In some research areas, simulation is difficult or intractable, and so open access to data is needed to enable successful research. The open data panel that Minister for the Cabinet Office, Rt Hon Matthew Hancock MP, has recently convened will be tackling open data questions. We support the work of this panel and sit on the steering group. We will continue to recommend measures that facilitate access to datasets, whilst protecting the rights of individuals to privacy and control over their data, and respecting the integrity and security of institutional data. It is vital to maintaining British leadership in machine learning research that the government does not lose momentum and continues to make firm commitments and progress towards a strong and innovative data policy.
4.8 Perhaps even more important are ensuring the highest standards of data security. Managing data securely is critical to being able to continue to improve the apps and services we all rely on with AI and machine learning. With UK citizens beginning to see the benefits of big data, data protection questions remain key to building and maintaining public trust, especially with a number of public services and organisations using different security protocols to share data. As secure and protected ways of providing data continue to evolve, government should play a significant role in supporting academic research into world-leading data security practices that would be widely adopted in the UK. Secure data will be one of the key foundations upon which success in AI research and innovation is built.
4.9 Further, the UK government should continue to maintain its public commitment to ensuring that encryption standards are never weakened, something that the current Investigatory Powers Bill is not clear about, as this Committee has already noted. As the Committee has also noted, further clarification is needed from Government to ensure that the UK remains a world-leader in the data security that is so pivotal to evolving technological innovation, especially AI and machine learning.
4.10 The UK Digital Strategy document produced by the government recently rightly recognised that reliable and high quality broadband connections are vital for the ‘dynamic economy’, supporting growth and labour market participation in both rural and urban areas. Good broadband infrastructure is particularly important for the delivery of improved public services through technological innovations and so we welcome the government’s commitment to 100% broadband coverage but encourage them to go further and faster in delivering this to the whole of the UK.
4.11 For example, DeepMind Health, in partnership with the Royal Free Hospital, piloted a mobile app called ‘Streams’ which presents timely information that helps nurses and doctors detect cases of acute kidney injury (AKI). A scheme like this, and potential others in future, will not be possible without world-class broadband facilities available in UK hospitals, and so we welcome the government’s commitment to 100% broadband coverage but encourage them to go further and faster in delivering this to the whole of the UK.
4.12 Our offices are based in The Knowledge Quarter, a world-class knowledge cluster in the heart of London that contains, amongst others, the British Library and Central St Martins. It is also home to some of the world’s leading scientific institutions: Google DeepMind, The Francis Crick Institute and The Alan Turing Institute are all based in King’s Cross, allowing unrivalled opportunities for collaboration and learning. The government should consider how it can build on this success and increasing the number of science-led organisations in the King’s Cross area so that a scientific cluster is allowed to flourish.
5. The social, legal and ethical issues raised by developments in robotics and artificial intelligence technologies, and how they should be addressed.
5.1 As with all scientific research, ethical oversight is important. Developing innovative and beneficial real-world applications requires access to real-world data. This raises privacy, security and ethics issues which require attention both by the practitioner community and by government. The Data Steering Group convened by the Cabinet Office is doing valuable work in exploring the ethical landscape here, and DeepMind are participating in and supporting this effort. DeepMind also has our own internal ethics board of philosophers, lawyers and businesspeople.
5.2 We believe that graduate degrees within computer science should incorporate mandatory ethics courses along the same lines as the ethics training required for medical and legal qualifications, including training in the ethics of data science and algorithmic fairness.
5.3 There are also some real-world applications of these technologies that deserve early attention, in advance of their widespread development and use. For instance, we are concerned about the possible future role of AI in lethal autonomous weapons systems, and the implications for global stability and conflict reduction. We support a ban by international treaty on lethal autonomous weapons systems that select and locate targets and deploy lethal force against them without meaningful human control. We believe this is the best approach to averting the harmful consequences that would arise from the development and use of such weapons. We recommend the government support all efforts towards such a ban.
5.4 As indicated in the above section, there are also key ethical and safety concerns around the security of data. Secure access to data, protected by strong encryption, is critical to both current and future innovation.
5.5 Ultimately, as with any advanced technology, the impact of AI will reflect the values of those who build it. AI is a tool that we humans will design, control and direct. It is up to us all to direct that tool towards the common good. We at DeepMind are incredibly excited about the potential of this technology to bring benefits and opportunity to people’s lives.
(2) BCC Workforce Survey 2014
(4) Building the Talent Pipeline, Tech Partnership and Nesta, October 2015
(5) For example, in 2014 we announced a partnership with Oxford University including a donation to sponsor PhDs and a collaboration that enables DeepMind employees to lecture at the University, and students to intern at DeepMind https://www.cs.ox.ac.uk/news/847-full.html
(6) In response to a parliamentary petition, the Government stated in February 2016 that “The Government is not seeking to ban or limit encryption. The Government recognises the important role that encryption plays in keeping people’s personal data and intellectual property safe online.”
(7) Science and Technology Committee, Investigatory Powers Bill: technology issues
(8) AKI is a contributing factor in up to 20% of emergency hospital admissions as well as 40,000 deaths in the UK every year. Yet NHS England estimate that around 25%of cases are preventable. Using Streams has enabled doctors to review blood tests for patients at risk of AKI within seconds of them becoming available, often meaning earlier intervention and improved care.