Is it a pear or an apple?
An introduction to Machine Learning
There are times where the same words are taking over all Business IT discussions, articles and arguments, and we tend to wonder if we really ‘get it’ or if we should ‘get out of it’.
So what is Machine Learning? A first definition
Etymologically speaking, it’s about some learning done by some machine. Generally speaking, the goal of learning is to get knowledge and from the use of knowledge, we get a sort of Intelligence, .i.e. one of the different forms of Intelligence Machine Learning is a sub-part, a step towards Artificial Intelligence.
As us humans are learning a certain way, machines use the same process - or should I say programmers develop the learning process so that a machine can “mimic” the way we learn.
How do we/it learn?
Humans need 3 things to learn:
- Information (provided by our senses, experiences and the repetition of this information)
- Synapses and neurons that will be rerouted by the ingestion of that information
- Brains that allow to process this information
Interesting enough, machines need 3 things to machine learn:
- Data sets (ideally a great number of tagged different information that describes the knowledge expected)
- Machine Learning Algorithms
Pear or apple?
Let’s take an example. If we can recognize an apple from a pear, it is because we’ve seen in our lives a great number of apples and pears of different shapes, colors, sizes, etc. Likewise, for a machine to tell a picture of an apple from a pear, it will take a great number of different images of both, tagged “this is an apple” or “this is a pear”. Flowing these images through the Machine Learning algorithm running on computers will create an ability, a “knowledge” for the machine to separate apples from pears.
Yet in both cases, if shown a banana, both will deduct it is neither an apple nor a pear, but that’s it.
To compare it to regular IT, for a computer to tell apart apples and pears, it would require a program, in other words a human that will try and find a systematic approach to separate both; like apples are rounder than pears, but both have possibly still a stem. And some oddly shaped pears can be round, while oddly shaped apples can be a bit oblong. It would be very difficult to program a machine to detect apples from pears, while machine learning will create a “model”, which is the outcome of the learning, that represents the ‘thinking to apply to a picture to tell if it is an apple or a pear’. At the difference of computer program that is “readable” by programmer, a model is not “intelligible” to humans but just to machines with this algorithm.
And even just as if you will probably learn faster the accordion if you already can play the piano, transfer learning techniques in Artificial intelligence allow to jump-start learning phase from a close ability either when little interesting data is available or when one wants to go a bit faster. So yes, Machine Learning is making machines learn from their own experience rather than program them out of human rules created from human experience.
Machine Learning is the focus of this year’s Atos IT Challenge, the first international student competition dedicated to technological innovation launched in 2011. The students will come up with an innovative and sustainable concept and develop an application based on Machine Learning technologies – visit www.atositchallenge.net.