24 Mar 2016
What you need to know about machine learning – part 3
Part 3: how does machine learning work?
A supervised learning machine is a lot like a toddler.
You can’t do this to a toddler, though:
Well, we suppose you could… But you probably shouldn’t.
*Note from the Phrasee legal team: Do not do those things to a toddler*
Also, toddlers sometimes do this:
If a learning machine behaves in this way, somebody messed up real bad.
But we digress.
A toddler, like a learning machine, explores and learns about its world through an exhaustive process of trial and error.
They have a limited set of data, in this case, pertaining to balance and remaining upright. When they are presented with a new scenario, like a slippery surface, each must adjust their behaviour to account for the new information.
The next time they encounter this situation, they apply the new data they received during the previous encounter, and, presumably, navigate the situation more effectively than they did the first time around.
This process is called “learning”.
Remember our definition from part 1: what is machine learning? If not, here’s a reminder:
A computer program is said to learn from experience “E” with respect to some class of tasks “T” and performance measure “P” if its performance in tasks “T”, as measured by “P” improves with experience “E”
Granted, the world of the learning machine is vastly more limited, but the same principles apply.
In this analogy, the programmer is the learning machine’s parent.
The issue for the parent and the issue for the programmer is essentially the same: How exactly does one explain all the fine points of how to walk on ice/snow?
You just have to let your kid/robot fall down a couple of times and figure it out on their own.
Image credit: Walt Disney Pictures
For the child, the adjustments will be a bit more intuitive.
For the learning machine, which lacks intuition altogether, the adjustment process will be more clunky, and probably involve a lot more slipping and falling.
In the end, in theory, both will end up in the same place, strutting around on slippery surfaces like the cock of the walk.
Of course, Atlas the snow-walking robot is an extreme case. Most machine learning is currently dedicated to the completion of much more mundane tasks, but the process is still essentially the same.
Machine learning is the process of a machine attempting to accomplish a task, independent of human intervention, more efficiently and more effectively with every passing attempt.
Since we are talking about Supervised Learning here (which you will no doubt remember from part 2 of this series the three types of machine learning), we can assume that the machine has been given clear parameters for what is considered a positive result and what is considered a negative result.
All that is left for the machine to do is try thousands, or even millions, of different ways to achieve the positive result its human handlers have requested, inching ever closer to a perfectly efficient solution and adjusting for changing external factors as it goes.
And this, ladies and gentlemen, is where our analogy falls down.
The machine will (we hope) remain 100% dedicated to doing exactly what we’ve asked it to do forever.
Good luck with that.