Believe it or not, there’s more than one way to teach a machine how to do things. Two of the ways we’ll examine are supervised and unsupervised learning.
These two terms are subsets of the broader topic of machine learning and concern how machines encounter data and deals with it. These are the two base methods and they’re often used in conjunction with one another. Rarely is one method employed successfully over the other. In practice, effective solutions often leverage elements from both approaches.
Supervised learning is the most common form of machine learning and is how most systems are set up. In broad terms, a supervised learning model seeks to teach a machine to follow a set of parameters based on information that is already known.
For example, let’s say we wanted to organize a group of geometric shapes (triangle, square, circle, rectangle, etc). Training supervised learning model would entail providing the machine with data to learn from. Triangle, rectangles, etc would be shown to the machine. It would learn what a triangle is and be able to group triangles.
This might be hard to believe, but in basic supervised learning, the system may actually end up with the same basic result as humans. That is, it may end up classifying triangles, squares, and circles as we’d expect them to. However, that doesn’t mean it’s drawing the same conclusion. There are some awesome and even humorous examples where machines get it very wrong.
Depending upon how you’d like to organize the shapes, we would teach the machine this information. More accurately, we would teach the machine to discern these differences by providing examples. Let’s say the triangle and square are both blue and the circle is red. If we wanted to group shapes by color, we first teach the machine that ‘this is blue’ and ‘this is red.’
How do we accomplish this? Provide the machine with samples. Ultimately, when enough data has been provided, the machine is able to make the distinction between blue and red on its own.
To an extent, unsupervised learning seeks to accomplish the same thing but in a different fashion. We present the same group of shapes to the machine, however, we do not teach it to recognize color or other attributes.
Rather, the machine analyzes the data presented to it and groups items according to identifiable characteristics. It may notice that the square and rectangle have a similar shape or that the circle has no angles, etc.
The machine is still being taught. However, it is not being taught the desired outcome. We are not going into the experiment trying to get the machine to recognize the color red. It will arrive there on its own.
Given that the machine is not taught the desired outcome, unsupervised learning can yield many different results. If you wanted to group your shapes by color, the machine may not go in that direction. The machine may group the rectangle and square together based on the fact that they have a similar shape. It may put the triangle in there as well based on the fact that all three have angles and so on.
In a real world scenario, picture a traditional classroom environment. A teacher presents their students with these same shapes and instructs them to organize them by color. Most students would have no problem accomplishing this since they’ve already been taught colors.
Now take that classroom scenario and change it so that the only instructions given are to ‘organize these items.’ As a result, we’ll have several different groupings based on any number of characteristics.
In this scenario, technically there is no wrong answer. However, the end result may not be what was originally intended or even all that useful.
Unsupervised learning is a more complex process. However, in theory, the system will improve itself over time as more information becomes available. This will become more important for future inceptions of artificial intelligence. As machines learn to teach themselves based on experience (as humans do), further applications in multiple fields will become apparent.