A few weeks ago, Google DeepMind’s AlphaGo beat the world’s highest-ranked Go champion, Lee Sedol, in a feat of machine intelligence that until recently many thought was still a decade away. This achievement is, at the very least, thought provoking.
Unlike chess, the possibilities on a board of Go are virtually endless. It cannot be “brute-forced”, the strategy that the Deep Blue chess computer took in 1997 against Grandmaster Garry Kasparov, by processing every permutation. In this case the artificial intelligence does some extensive decision tree searching, but after that relies on something much more akin to — for lack of a better word —intuition.
So how does this depart from traditional machine learning? And, more importantly, what does this approach mean for those of us looking to improve digital services by solving “Big Data” challenges?
The idea of deep learning, basing machine intelligence on the way the human mind learns, is not new. It has its roots in the 1940s, when the hypothesis of learning based on neural adaptively was put forward and then first applied to computational models with Turing’s B-type machines.
The early neutral network models were created for pattern recognition; however, it was not until the late 2000s — when computing power had achieved a level sufficient to train complicated, multi-level neural networks — that they would become powerful enough to start addressing image or speech recognition, for example, and finally play a pivotal role in beating a human in a game of Go.
While deep learning can be trained with example games played in the past by humans, the strength of AlphaGo is related to the idea of reinforcement learning, which itself was inspired by behavioural psychology.
During this process, AlphaGo played a large number of games against itself, continually improving by exploring and identifying winning strategies. AlphaGo combines different techniques rather than relying on a single solution, and will continue getting stronger the more games it plays.
Going forward, there are two key points that make Deep Learning so interesting in how it approaches “solving” Big Data problems.
Take a challenge: decreasing the response time of medical emergency services. In London a typical ‘life-threatening emergency’ 999 call aims to bring an ambulance to you within eight minutes, 75 per cent of the time. If we could raise that percentage to 90, we would save 3,000 more heart attack victims per year.
Deep learning algorithms can take in many different types of Data.
‘Availability’ of ambulances is surely a factor of geography, and how strong the coverage of such services is in your area.
‘Demand’ is at least in part a function of population density and demographics; the more people of a certain age in a given area, the higher the likely demand. On the surface these are fairly static data points, and surely some straightforward data mining would reveal areas with better or worse coverage and response times, and permit more effective distribution accordingly.
But what if we could go one step further and include different sources of real-time data? Traffic conditions, weather conditions, medication purchase data from local pharmacies, even readouts — via the relevant apps — from your Fitbit. All of these could potentially contribute to a much more relevant “real-time” map of risk areas, and allow for a ‘smarter’ day-to-day distribution of emergency services.
Deep learning algorithms become smarter the more they ‘play’.
One of the key differences about ‘reinforcement’ learning as opposed to rules-based is that the AI teaches itself the best way to ‘win’ the game.
Taking the example above, a neural network could be deployed in a virtual environment (designed on real historical data), with the task of analysing evolving ‘risk’ patterns in a geography and positioning ambulance units in advance of likely emergencies. Simulated response times could then be measured against actual RTs, and positive/negative reinforcement given in relation to the outcome, e.g. 90 per cent under eight minutes. Once the AI has been ‘trained’ to consistently outperform existing practices in simulations, it would then be deployed in the real world and continually monitored as it adapts and improves.
Technology has reached a landmark point, though the issues remain around the Law and ethics around using such tools in the public space. Once a consensus has been reached however, we can anticipate seeing a host of new possibilities opened up by Deep Learning, with public service improvements being just a few of the many changes in peoples’ lives.
Erik Moltgen and Sami Niemi are consultants at Valtech. Learn more about Valtech at their website.