17 Feb 2016
What You Need to Know About Machine Learning – Part 2
The three types of machine learning
When considering machine learning as a concept, it is important to remember that it is a complex field. One that’s rife with categories and subcategories, with yet more subcategories being added by the day.
To delve too deeply into all of these would be to curse you, dear reader, to several torturous hours of maths and more maths until you would simply give up and decide to watch YouTube videos about X-rays of objects found in people’s butts.
The fact is that there are only three types of machine learning that you really need to know about for the moment.
We say “for the moment”, because advancements in this field, and the creative endeavours it is delving into are moving and changing so rapidly that it is impossible to predict where things are headed.
The three types, in no particular order are:
1) Supervised learning:
Unlike other types of machine learning, supervised learning always has a specific preset outcome that is determined by a human before the machine begins to learn.
By specific preset outcome, we mean that the data is funnelled into categories for which the parameters have already been decided. The machine simply tries to get the results we ask it for.
For example, when an email is sent to your inbox it is categorised, before it arrives, as either a “junk” email or a “regular” email.
Ever wonder who decides which is which, and how the distinction is made?
Ever find an email from a friend or family member in the junk box despite clearly not being junk?
That’s our new friend, the supervised learning machine, hard at work.
Before it started its new job, it was given millions of emails to look at. Each email it was given had been prelabelled as either “junk” or “regular”.
The machine uses these emails to “learn” what different patterns and attributes are found in junk emails and regular emails. It then sifts through billions of emails every day, looking for those patterns and attributes and labelling your emails accordingly.
Each time you click the “not junk” or “junk” button on a miscategorised email, the machine learns a little bit more and becomes more efficient at discerning the two from each other. Every time this happens the machine learns from its mistake and adjusts the attributes it looks for in each email accordingly.
The reason supervised learning is different from the other types of machine learning is contained in its limitations.
The email-sorting machine can only do one thing with all the data at its disposal; sort emails.
It can complete that one task very efficiently, and becomes more efficient as it goes. However, it cannot interpret the data it analyses in any other way, unless it is shown how to do so.
If a user wishes to delve deeper into the data and glean something else from it, he/she is much better served using:
2) Unsupervised learning
For an unsupervised learning computer, there is no right or wrong answer. There is only data and more data.
Its mission is to extract structure from the data it is given or learn how to represent the data in a way that it can be more useful to us.
It is, in a way, the supervised learning process in reverse. The machine is given a mass of data and told to look for patterns or “clusters”. It then draws its own conclusions, rather than the conclusions we want.
A good example of this type of machine learning is discount cards at big supermarkets.
You know the ones. All the supermarkets have them now. The transaction is simple; you give the supermarket your shopping habit data in exchange for discounts on some of the products they sell.
Then the data from your card, along with everyone else’s cards, is fed into an unsupervised learning machine.
The machine sorts through the purchases and buying behaviours of all cardholders and separates them into different categories for marketing purposes. It does all this without any instructions on how to separate groups, how many groups to have, which people belong in which group, or even how to describe each group.
The machine is left to sort through the data and look for patterns with no human input and then represent it visually in a way that can be analysed.
The benefit of this type of machine learning is that it can give insights that a business didn’t even know it was looking for.
The limitation of unsupervised learning is that it generally requires human input after the fact to decide how to use the information gleaned from the process before it can make anyone money.
If you want a machine that not only analyses data, but can also learn how to utilise the resulting information to complete tasks more efficiently, you need:
3) Reinforcement learning
In reinforcement learning, there is no supervisor, only a reward signal. The only data relevant to such a machine is how much reward it receives each time it completes a task.
The machine’s actions and the rewards they produce effect the subsequent data it receives. Therefore, it continues to get better at completing the task it has been asked to perform.
A great example of such learning is the machine that defeated the world champion at backgammon.
This machine was taught the rules of backgammon, but nothing else. It had to try out millions of different strategies for each situation it came across and then develop its own overarching strategy. All it knew at the end of each game was whether it had won or lost.
The machine was left to its own devices to learn where it had gone wrong with each loss and how it could do better the next time.
For the reinforcement learning machine, there is always a right or wrong answer, but the machine has to figure it out for itself.
The potential for this type of machine learning to change the business landscape entirely is very real.
We here at Phrasee have developed highly effective practical applications for it that are already helping businesses across the globe reach markets more effectively than ever before.
If you’d like to know more about how machine learning can benefit your business, check out our website.