Machine learning — sexy and it knows it. Post 1 of 3.

Many of us remeber the scramble that occurred when the hype surrounding big data made its way up to the boardroom. There we were, quietly doing analytics, when at some point in 2011 the big data hype made its way up to the C suite. I’m not sure how it first got there, but I can hazard a guess as to how it became such a hot topic. Senior execs tend to be, by nature of their roles, highly competitive. As big data was (and still is) seen as an area of competitive differentiation in many industries, it stands to reason that when this sexy new area of technology made its way into the board level lexicon it stuck. One can only imagine the cries of “my data is bigger than yours” ringing out across the C-suite golf courses…

We seem to be seeing similar hype at the moment around machine learning with interest really ramping up over the last year, presumably at least in part due to the press coverage of Google’s AlphaGo algorithm. As with big data, the hype is making its way up the corporate ladder to the exec suite. I have noted execs starting to litter their rhetoric with phrases such as ‘robotics’ and ‘artificial intelligence’, and when McKinsey start recommending that C-level executives incorporate machine learning into their strategic vision, you know it’s here to stay.

So, what is machine learning? Is the hype justified and how can organisations use it to benefit their business?

Old news
The first thing that machine learning isn’t is new. It was first defined in 1959 by Arthur Samuel as a “Field of study that gives computers the ability to learn without being explicitly programmed”. However, more than a decade earlier Alan Turing was thought to be considering the possibility of ‘machine intelligence’ in his work at Bletchley park; and in 1950 he published his seminal paper ‘Computer Machinery and Intelligence’ in the journal Mind where he explicitly discussed the idea of ‘learning machines’.

Arthur Samuel, however, is credited as the first to construct a machine learning algorithm with his Computer Checkers program which sought to win a game of draughts. It did this by identifying the optimum next move using a directed graph (an algorithm where nodes called vertices are connected by links called edges and travelling from one vertices to the next is in a predefined direction). Crucially, Computer Checkers also learnt from itself by remembering every position it had already seen and the final outcome of the game. Samuel also had his machine play thousands of games against itself as another way of learning.

Following Samuel’s early work, the science of machine learning entered somewhat of a lull until the late 1970’s when computing power began to catch up with the scientific theory and rendered the theoretical ideas of early pioneers a plausible reality. Machine learning didn’t really take off in a big way, however, until the 90’s when the focus shifted from a knowledge-driven approach to a data-driven approach. This data-driven approach, combined with the continuing increase in computational power, gave rise to many of the examples of machine learning with which we are familiar today, such as IBM’s Big Blue and Watson; and Google’s AlphaGo and driverless cars.

What is it?
So, it’s old news and can play draughts. The same is true of Robbie Williams (according to the daily mail), but I wouldn’t want to incorporate him into my corporate data and analytics strategy. What exactly is machine learning? How do machines learn? What useful purposes does this serve beyond winning games? What is the difference between artificial intelligence, robotics, data mining and machine learning ?

In a nutshell, machine learning is analysis that iteratively learns from data without having to be explicitly programmed with rules. As analytical models are exposed to new data they independently lean and adapt to this new information. Consider how this applies to a simple old school computer game like space invaders. The ‘fresh’ machine learning algorithm is given the inputs of the pixels, and the moves (up/ down/ left/ right etc.); and the output it must produce (to get the highest score possible within the allotted time). It then starts playing the game. Initially it uses random moves (and probably sucks as much as my little sister did in the early 90’s when we played it. I was, of course, much better). Despite playing like a 5 year old, it will accidentally score some points. It recognizes the connection between what it did (the inputs) to score these points (the output) and then uses this data to determine the inputs it should make in subsequent games instead of selecting them randomly. As it iteratively plays the game (tens of thousands of times) the quality of its inputs continues to improve until it is almost as good at space invaders as I was aged 8.

Consider how complex it would be to program the same outcome with rules. Even in a game as simple (and awesome) as space invaders, the number of potential combinations of inputs is vast. Explicitly programming each one and feeding this to your machine (instead of having it figure it out for itself) would be a mammoth task.

Now consider how the same concept applies to driverless cars. Identifying and coding every possible situation that a vehicle could encounter and the optimum response is impossible. However, feeding a machine learning algorithm tens of thousands of hours of driving footage and having it simulate the response to learn which is the optimum output, that’s do-able (if you’re Google anyway). Then consider what happens when you let the driverless cars loose in the real world. They continue to learn from their real world experiences. As they are all connected up to each other (via the cloud, natch) they can learn not only from their own experience but from the collective experiences of all the driverless cars…. It’s at this point that you realise why they might actually be better drivers than us mere mortals.

Science imitating nature?
It strikes me that there is an interesting parallel with nature here. Natural selection (bear with me) starts off with a lot of variations and randomness. Different phenotypic traits (inputs) confer different rates of survival and reproduction (outputs). The inputs that result in the optimum outputs are selected for and plugged back into the next iteration of the model (the next generation). Thus evolution is in effect a machine learning algorithm. The nature of the machine, however, is beyond the scope of this blog (and quite probably my intellect).

Post two in this series of three will consider the relationship between machine learning and some other current buzz words like artificial intelligence and robotics, as well as ask why machine learning is suddenly all the rage. Thanks for reading!