I was excited to attend Re·Work’s Deep Learning Summit at the end of September 2016 in London. I was looking forward to spending time immersed in the wonderful world of all things Artificial Intelligence (AI) and data-related and it really did not disappoint. I came away with so many new thoughts and perspectives, as well as having met some great minds.
I’d love to share some of this with you in a round-up of the event and Deep Learning in general.
What is Deep Learning?
Many people associate Deep Learning with the concept of AI and Machine Learning or just robots, like in the movie Ex Machina, the TV series Humans or WestWorld. Yes, Deep Learning is all these things, but it is also a lot more than that. As the name suggests, it is actually about learning. It’s how machines learn.
In Deep Learning, algorithms are used to give the computational programmes hundreds and thousands of examples that it can learn from. Out of these examples, the programme learns how to master something as we would do, it tries different ways of doing things and refines its approach over and over.
In the past the challenge for this type of development was access to a large enough amount of data for these artificial neural networks to process. Today, however, in the age of the internet, data is available like never before. DeepMind recently fed one system every Daily Mail article available online. It subsequently trained the programme to recognise certain language patterns from this and predict missing words in headlines.
Where Are We Right Now?
DeepMind is a British AI company founded in 2010 and bought by Google in 2014. Its aim is simple: ‘Solve Intelligence. Use it to make the World a better place.’
So how are they doing this? DeepMind Health was launched in February this year. This part of the company is building clinician-led technology in the NHS. Their first project is to use Machine Learning to recognise life-threatening conditions detected in the eye, and degenerative eye conditions.
At the core of all their projects though, is the research and knowledge-building needed to perfect true machine intelligence.
DeepMind created AlphaGo, a programme that can play the ancient Chinese game Go to world championship standards. In 2016 it beat the top Go player in the world from the last decade, Lee Sedol, 4-1 in a series of 5 games.
What is remarkable about this is that AlphaGo wasn’t programmed directly to play the game Go. Instead it uses deep neural networks, alongside supervised learning and reinforcement from self play. In effect, AlphaGo taught itself to play Go to this standard.
It is important to note why Go, as a game, was chosen. This incredibly simple and yet infinitely complex game has few rules but so many possible strategies and moves. Think Chess on steroids. It has been held up for years as the ‘grand challenge’ for any artificial intelligence being built. As DeepMind notes ‘there are more possible positions in Go than there are atoms in the universe.’
Re•Work Deep Learning Summit
At the Deep Learning Summit I listened to many speakers. It was interesting to hear from such a range of voices on the subject of Deep Learning.
Miriam Redi, from Bell Labs -Nokia’s research and innovation division – spoke about how they are working on Deep Learning that will allow programmes to see the ‘invisible’ in images and videos. Programmes are now adept at identifying objects and their technology has moved to focus on how programmes would identify more abstract concepts such as the beauty of the image (not the person in it) or the creativity of a video.
In order to do this they are building frameworks that give the programme a way to look at an image and apply the rules and aesthetics that photographers would use to appraise the image themselves. Again, access to large datasets is key and they have used videos from Vine and images from Instagram and Flickr. Their programme can now identify creative videos with an accuracy of 80%, which is high for this kind of technology.
What I liked about their approach is the way in which they are going right back to the start of this process – what makes an image beautiful? – and defining that in detail across cultural backgrounds to ensure that any outcomes from this technology can have applications across the globe.
Raia Hadsell, from DeepMind, talked about continual deep learning through games. In February 2015 DeepMind announced that they had built an artificial intelligence agent that can learn to play 49 Atari games. In contrast to other AI systems – IBM’s Watson or Deep Blue – DeepMind’s agent was given little information. With just basic notions of pixels and scoring the agent played each game over and over to learn the best strategies and skills to apply to each one.
What is striking and new about the agent playing these games is not just that they have learnt how to play the games themselves. Through experience the agent tried over and over to play the games. When it scored in any game this showed it how to improve and it continued to learn from there. What is remarkable is that once it had mastered these games it went on to play them to an expert level. In some cases it found strategies and ways to win that we, as game players, had not thought of.
Does AI have a role in THE STRATEGY JOURNEY for businesses and people?
There is no doubt that in the business world, data has become king. For individuals, the famous quote from Francis Bacon still holds true today.
“Knowledge is power”
Along with the right processes in which to act on the data, a data-rich and data-driven company or individual could really take over the world. All businesses and individuals actually go through a journey (in fact many journeys) and they need to apply different strategies along the way to succeed or fail in whatever they do. I have called this THE STRATEGY JOURNEY in my book (coming in 2017) where I discuss the importance of learning. My own company Stratability is built around data and machine learning. We are building rich data about people’s mindsets that are outside normal profiling methods using assessments and games. My belief is that businesses and people will succeed or fail in the future on data, and what they do with it, along their strategy journeys.
What’s Next? Looking to the AI Future
When I got back from the Re•Work summit I was chatting to one of my colleagues about the AI breakthrough at DeepMind with the agent that is playing Atari games. Whilst she was impressed by what it can do I pointed out that we are still very much at a basic stage.
When we think of an AI future we are surrounded by examples of what that could be from novels and films. With AI Professors as expert advisors to films such as Ex Machina, our fictional worlds of AI are beginning to seem a close future indeed. Whilst the agents mentioned in this article are able to learn, they are nowhere near the level of the fictional AI that we see around us.
As Miriam Redi noted in her presentation, it is currently on us to give any agent not just the data to go through and learn from, but parameters with which to analyse the data. AI that can see the invisible is still being shown where to look.
The future could be something very different. AI that can know where to look, and what it wants to find, without even being told.