Contrary to much commentary, machine learning is not a brand new subject. Indeed, as long ago as 1957, Frank Rosenblatt invented the Perceptron – an algorithm that could be trained to classify images through supervised learning.
The New York Times breathlessly described the hope that the machine might, one day, “be able to walk, talk, see, write, reproduce itself and be conscious of its existence”.
The machine learning conversation has recently hit new volumes, powered by three separate revolutions: first, computing power; second, data generation, storage and retrieval (an estimated 90 per cent of the data in existence today were created in the last two years, while in 1981 a GB of storage cost $300,000 whereas today the price is just $0.10); and third, the combination and maturing of methodology from statistics, computer science, mathematics and engineering, amongst others.
Putting a precise definition on machine learning is difficult, combining as it does methods from a wide range of disciplines. The fundamental idea behind machine learning, however, is that it can help us to identify repeatable patterns and relationships in data, without requiring a predetermined idea of what to look for.
This ability to understand relationships in data, without first forming a hypothesis about what those relationships might be, is what distinguishes machine learning from more traditional data science techniques or statistical models.
The excitement around the subject of machine learning seems to have reached new levels in the last 12 months. The public’s imagination has been captured through events like Google’s AlphaGo beating the world’s Go champion Lee Sedol.
Mr Sedol is rightly feted as one of the game’s greats and some had predicted he would win by a landslide.
It is almost 20 years since IBM’s Deep Blue beat Garry Kasparov at chess, but many had thought it would be a while longer before the same feat was achieved in Go, due to the game’s greatly increased complexity (by one estimate there are more board permutations than there are atoms in the universe). AlphaGo’s 4-1 victory thus surprised many.
Man AHL has been actively researching machine learning techniques for several years (including through the Oxford-Man Institute, a collaboration between Man AHL and the University of Oxford) and applies these techniques in client trading programs.
As a systematic investment manager, Man AHL seeks to understand relationships in financial data, so that we can take things we can measure and observe today and use them to forecast what financial markets might do in future.
The deep learning approach used in AlphaGo, to such success, is a modern day extension of artificial neural networks, which can in turn be traced back to Rosenblatt’s seminal and pioneering work.
That said, there remains much to do before machine learning reaches the sophistication anticipated in that New York Times article of nearly 60 years ago.
Dr Anthony Ledford is chief scientist at Man AHL.
