On Monday 6 June 2016, Committee Members visited Google DeepMind—one of the largest machine learning labs in the world—based in King’s Cross, London. The then Chair, Nicola Blackwood MP, Victoria Borwick MP and Matt Warman MP were present. Following a tour of the building, Members met with Mustafa Suleyman, one of the co-founders of the company, and were presented with an overview of Google DeepMind’s work to date. Members were then joined by Demis Hassabis, another co-founder of the company, and took part in a question and answer session. Topics covered during the discussion included:
(1) Background to DeepMind’s work: This included an overview of how artificial intelligence has developed over the years, from IBM’s Deep Blue —which evaluated 200 million positions per second—beating Garry Kasparov at chess in 1997 to the ‘Pocket Fritz’ computer—which needed to search less than 20,000 chess moves per second—winning the Mercosur Cup in Argentina in 2009, and achieving a higher performance level than DeepBlue. A video was also shown highlighting DeepMind’s success at building an artificial intelligence agent that could learn directly, via experience, to play classic Atari games better than humans.
(2) AlphaGo: the game of Go, and why it is considered the most difficult game devised by humans, was discussed together with how AlphaGo trained for the match against Lee Sedol in March 2016, and the reinforcement learning framework it used. It was noted that AlphaGo made unprecedented moves that commentators initially thought were mistakes and that, in doing so, AlphaGo has provided human Go players with new knowledge and new insights on the game. Lee Sedol, for example, has won all his games since he played AlphaGo and has stated that it taught him to be more creative in his play.
(3) DeepMind Health: it was explained that DeepMind Health was focused on improving patient safety and that they were adopting a user-centred approach to research so that they could learn more about what tools doctors and nurses need to improve patient safety. DeepMind Health’s work on acute kidney injury (AKI) was highlighted, alongside its development of an app to help a) detect which patients are deteriorating and b) manage the subsequent intervention. Though it was anticipated that AI could become a powerful modelling tool that could be applied to solve ‘wicked problems’ in the future, it was noted that the AKI app does not use AI.
(4) Ethics: it was acknowledged that AI will have significant implications and that ethics and safety must be paramount. Google DeepMind noted that it was talking with other companies developing AI about standards and best practices.
(5) Safety: it was noted that safety is incredibly important to the development of AI and that responsible measures must be taken—such as by enabling ‘safe exploration’—to ensure its operation within defined constraints.
212 For further information about ‘safe exploration’ see: Dario Amodei et al, , 21 June 2016
5 October 2016