Arguably the world leading Artificial Intelligence company has some incredible tricks up its sleeve
Founded in 2010 and acquired by Google in 2014, DeepMind is one of the world’s foremost artificial intelligence companies. By combining two areas of AI research – deep neural networks and reinforcement learning algorithms – DeepMind has created programmes which can teach themselves, and also apply their knowledge to new situations. This has truly revolutionary consequences for the application of AI to the real world. Here D/SRUPTION takes a look at some of the most impressive achievements of Google DeepMind to date.
1) Lip reading AI
In conjunction with the University of Oxford, DeepMind created Watch, Attend and Spell (WAS), an AI system which can lip read speakers on television. In order to achieve this feat, researchers applied computer vision and machine learning techniques to a dataset of more than 5,000 hours of silent TV footage, focusing on speakers on programmes such as Newsnight, BBC Breakfast and Question Time. Significantly, the WAS software system was better able to identify what was being said when compared to an expert human lip reader, with the software accurately recognising 50 per cent of words to the human’s meagre 12 per cent. What’s more, any errors made by the computer were small – such as single letter misspellings, or failing to recognise an ‘s’ at the end of a word – meaning that it was still possible to understand most of the transcribed speech. This technology has huge potential to improve accessibility for people with hearing loss, by providing fast, accurate speech to text transcription.
2) AI navigation
Navigation has traditionally been a huge challenge to artificial intelligence, with programmes trained by machine learning failing to recreate the spatial behaviour of humans and other mammals. However, when scientists at DeepMind trained their artificial intelligence to move through a landscape, it spontaneously developed electrical activity like that found in the human brain cells which control navigation. After discovering this activity, the researchers developed another version of the AI programme. This went on to beat experienced human players at a game which involved racing through a virtual environment to find a prize. These breakthroughs show the potential for human like activity to evolve naturally in AI systems. They also highlight how AI models can be useful to neuroscientists for understanding the human brain, with important links between artificial intelligence and our own, natural version.
3) Self taught AI
AlphaGo, one of the earliest iterations of AI developed by DeepMind, gained fame in 2015 when it beat Go’s European champion – becoming the first ever computer programme to beat a human professional at the ancient Chinese game. The following year AlphaGo went on to defeat the 18 times world champion of Go, earning itself a nine dan professional ranking – the highest certification in the game. Not content to stop there, DeepMind researchers then created AlphaGo Zero. In contrast to the previous versions of AlphaGo which learned how to play by studying thousands of human games, AlphaGo Zero taught itself. This involved the AI playing against itself, beginning with completely random moves, until it worked out the tactics of the game. It took only three days of training for AlphaGo Zero to beat its predecessor. After 40 days, AlphaGo Zero beat all of the previous versions of itself, arguably becoming the best Go player in the world. This process of AI self teaching, known as reinforcement learning, demonstrates the exponential power of artificial intelligence which will revolutionise the way we solve problems in the future.
4) Identifying disease
DeepMind has a large division devoted to health. In a partnership with The Royal Free hospital in London which began in 2015, they developed a data analysis and diagnosis app, Streams, to help doctors identify cases of Acute Kidney Injury (AKI). Whilst the app doesn’t yet use AI, DeepMind see it as an important step towards an AI-enabled medical future. In fact, in 2016 the company began a project with Moorfields Hospital NHS Foundation Trust, exploring the use of AI to diagnose sight threatening eye conditions. In results published in July, DeepMind revealed that they have created AI technology which can detect eye disease in seconds, drastically reducing the time it takes for patients to receive the treatment they need. By making the screening process for eye disease more efficient, DeepMind hopes to tackle one of the major causes of sight loss, which currently affects more than 285 million people worldwide – a figure which is expected to triple by 2050.
5) Cooperative AI
As humans, we may exist as discrete individuals but we still possess the ability to join forces with others and work as part of a team. In the computing world the acquisition of this skill is known as multi agent learning – and it’s an extremely difficult problem to solve. Programming such behaviour involves getting a large amount of individual agents to work independently, whilst also making sure they can collaborate with others whose actions are constantly changing. To investigate this problem, DeepMind applied AI to agents in the first-person multiplayer video game Quake III Arena, where two teams of individual players compete to capture the opposing team’s flag. Results showed that the AI agents spontaneously developed tactics used by human players, such as defending their own base, camping outside their opponents base, and following their teammates. What’s more, the AI agents were not only more successful than human players, but also more collaborative – showing that AI has the potential to work together with other computer programmes and humans in the future.
These breakthroughs demonstrate that DeepMind is putting its mission of developing AI to benefit the world into practice. In the future, as the state of affairs in domains such as climate change, economics and disease becomes more complex, we will need AI to help us solve some very complicated problems. Whilst DeepMind’s AI developments are certainly impressive, they are therefore small stepping stones to the highly powerful, flexible AI systems that we can expect in years to come.
For more AI insights, subscribe to our free weekly newsletter here.