What is unsupervised machine learning?

- [Presenter] Machine learning is such a rapidly changing area that we really want to pause for a moment and define both machine learning and our topic, unsupervised learning. Let's start with machine learning. Wikipedia defines it in the following way. Machine learning is a field of computer science that gives computers the ability to learn without being explicitly programmed. This is the phrase that you'll commonly hear that I think is causing a bit of confusion. Let's continue just a bit, but I want to contrast this with my own definition of machine learning. Wikipedia's continues, machine learning explores the study and construction of algorithms that can learn from and make predictions on data. I certainly agree with this. So here's the problem. The not being explicitly programmed, what are we talking about exactly? So take a look at my definition. A broad term that generally refers to presenting carefully curated data to computer algorithms that find patterns and systematically generate models, and those models are in the form of formulas and rule sets. What I'm trying to clarify here is that humans write the algorithms. So this whole notion that we're finding things without programming, we're not writing the models, but we are writing the algorithms that write the models. Back in the '80s expert systems were a hot topic. But back then a human would interview another human expert, let's say a lawyer, and then actually write the rules, and actually write it that way. We haven't done it that way for decades. But again, we write the algorithms. We curate the data. The algorithms take care of that next piece generating the models. So what form does this take in supervised learning? In supervised learning we give the computer a data set with what is called a target variable. That's whether or not the person had heart disease, or did not have heart disease, let's say. And we give a series of input variables. The modeling algorithms automatically generates the model, which establishes a relationship between the target and the input variables. This is usually described as giving a series of pairs. You give the inputs and along with the input you give the outcome. This is not the subject of our course today. Decision trees is perhaps the most common supervised learning technique, and I actually have a course on decision trees right here in the library. Unsupervised learning is a search for homogeneous subsets that produces one or more possible segmentations of the data that differ by how similar the cases are within the segment and how many subsets there are, and how common or rare they are. So the whole idea of homogeneous is just put similar cases together. Now, the computer will help you do the search. But you have to decide how similar is similar enough for your application. You have to decide how many subsets you want there to be. And you have to decide how rare or common the subsets are going to be. So it's a true partnership between the human and the algorithm. I've seen definitions of unsupervised that imply the computer does all of these things. Unsupervised is actually more difficult for the human analyst than supervised in many cases. For instance, take a look. Let's just say we were asking our computer to perform unsupervised segmentation on these blocks. Should be separate them so that they're blue and non-blue? Or maybe red and non-red? Should we be focusing on shape or is that less important? Should we be focusing on size? It's probably impossible to answer outside some context. Specifically for us, some business context. So the computer will help group, but only in partnership with a human analyst that knows what the goal of the segmentation will be. So sometimes I see definitions of unsupervised that make it seem like it's pretty much supervised, but when you can't find a target variable or you don't have ready access to a target variable it's actually a very different set of goals when we do unsupervised. Now we come to a third area of machine learning called reinforcement learning. Only recently has this been a source of confusion, mostly because of two application areas that have gotten a lot of excitement. Technologies like driverless cars and technologies like the one that was able to beat the world champion at go, these operate in a very different way. At first, they sound like supervised. But they're quite different. What's happening here is you're not presenting the algorithm with a very large number of input and output pairs. Rather, it's working in an environment and it's correct and incorrect not for each row of data, but it's rather rewarded for correct actions it takes in the environment. This is the kind of technology that arises for very cool things like watching a robot stumble and then as it practices it gets better. A lot of folks think this is the way supervised learning works, but it's not true. You're actually presenting that carefully crafted data set in one shot to the computer, and the algorithm takes that data and immediately creates the model. The model does not learn over time. Reinforcement learning is the style that learns over time and is quite different from the other two.