AI in the UK: ready, willing and able? Contents
Appendix 5: Note of Committee visit to DeepMind: Wednesday 13 September 2017
The Select Committee on Artificial Intelligence visited DeepMind’s headquarters in King’s Cross, London, on 13 September 2017. The Committee met with Demis Hassabis (CEO and co-founder of DeepMind), Mustafa Suleyman (Head of Applied AI and co-founder of DeepMind) and Joe Ledsam (a clinical research scientist with DeepMind Health). Lunch was provided.
Eight members of the Committee were in attendance, as was Dr Mateja Jamnik, Specialist Adviser to the Committee.
Members were given an overview of DeepMind’s history and work, and took part in a question and answer session. Topics covered included:
- Background to DeepMind’s work: This included discussion of DeepMind’s focus (making sure they are carrying out the right research, and that such work is as effective as it can be), and of DeepMind’s objectives (to solve intelligence by understanding natural intelligence and applying this knowledge to machines; to use this work to make the world a better place; and to develop the world’s first general purpose learning machine that can adapt to any task). Several global challenges were highlighted (access to clean water, food waste, energy consumption and climate change), as was the potential for artificial intelligence to help address them.
- Technical work and research: This included discussion of DeepMind’s development of AlphaGo (a computer programme which can play Go) and its successes. It was made clear that artificial intelligence developments were in their infancy and no one could fully comprehend the capabilities of the systems that could be created. A demonstration was given of DeepMind’s application of a Deep Q-Network (DQN) algorithm to the arcade game Breakout. The algorithm was able to learn over hundreds of games to win more quickly than a human player. ‘Blackboxing’ was discussed as an important engineering challenge to address.
- Working in the UK: This included discussion of DeepMind’s commitment to the UK, its relationships with universities and its establishment of scholarships and sponsorship programmes to support machine learning in the UK (to address the shortage of skilled workers they needed). DeepMind was engaging with the public sector to inspire other AI companies to do the same. DeepMind anticipated that artificial intelligence systems would assist the work of humans, rather than replace it. It was a concern that the technology would be prematurely regulated, which could hinder development and further research. There was a discussion around whether a kite mark scheme could be a potential first step towards an industry-led regulatory framework, with artificial intelligence products marked to show they had met an agreed standard.
- DeepMind Energy: Data centres around the world use up to 3% of global energy. Given this was a key resource in the development of artificial intelligence, DeepMind had sought to address this. DeepMind’s first application of AI in the energy sector was to try and reduce the amount of energy used for cooling a Google data centre. The AI project led to a reduction of up to 40% of the amount of energy used for cooling, and a 15% improvement in the building’s overall energy efficiency. They were considering further applications of this in the context of other services.
- DeepMind Health: DeepMind had identified that the data used by the NHS was mostly paper-based and the process of caring for someone was an incredibly complex system. DeepMind had focused on two health problems: acute kidney injury (AKI) and eye conditions which lead to blindness. They had created an application which allowed doctors to react quickly to evidence of acute kidney injury, which is currently deployed at the Royal Free Hospital. They had worked with Moorfields Eye Hospital to improve the triage and time taken to see patients at risk of serious eye conditions. It took a year to clean and label the data required to inform the system being developed, and eye experts were used to help train the algorithm. There are promising early signs and the peer reviewed research should be published at some point in 2018.
- Ethics: This included discussion of DeepMind’s work in creating the Partnership on AI, which is intended to be an industry forum to improve understanding of AI and establish best practice for its development. DeepMind were also establishing an internal ethics unit within the next year. DeepMind had created an Independent Review Panel for their work in healthcare as a demonstration of their commitment to transparency and because of the ethical issues raised in accessing patient data.