Machine Learning: The Rise of the Machines
What is Machine Learning?
When first hearing the words “Machine Learning”, most folks will immediately conjure up ideas, popularized by Sci-Fi movies, of robots taking over the world and crushing every, last existence of humanity. However, quite often actually, people use some form of artificial intelligence in their everyday lives. For instance, if you’ve ever used Siri voice-assistant, or an autocomplete feature in a text messaging app or maybe you’ve applied for a line of credit online, you are using some form of Machine Learning.
So, what exactly is Machine Learning? Well simply put, Machine Learning is applied, computational statistics. Or in other words, it’s the science of building intelligent, computational artifacts that learn over time based on past experiences. In fact, Machine Learning can be broken down into 3 core disciplines:
- Supervised Learning
- Unsupervised Learning
- Re-enforcement Learning
Supervised Learning is the process (or challenge) of taking labeled data sets and extracting information to create new data sets. Wait…what? What exactly does that mean? Perhaps this will become clearer with a simple example.
Example: Think of the table below as a system with an input and an output. The first row in the table represents a series of inputs, and the row below represents the corresponding outcome of that input.
Challenge: What is the output if 5 is the input? Name the function that best represents this system?
Many would take a leap of faith and guess that the answer would be 10 and that the function would be: f(x) = 2x – whatever number is put into the system the output would that same number doubled. While this is a pretty trivial example, it does illustrate some of the important points of Supervised Learning, one of them being that induction is crucial. In other words, through mathematical induction we can take a data set, like the example above, and make fundamental assumptions to predict an outcome in the future. This is the basis of Supervised Learning and which is why it is sometimes described as functional approximation.
Pitfall: The example above assumes that the function is well-behaved. Imagine if, however, only numbers less than 10 were doubled. This is, in fact, a fundamental problem that plagues Machine Learning, generalization.
Unsupervised Learning is a lot like wearing a blindfold in a dark room. In this category of Machine Learning, there are only inputs and no outputs; there is no feedback. We can, however, derive some sort of structure by looking at the relationships of the inputs themselves. We can look at the given data and determine a way to label the data into different categories.
Unlike Supervised Learning, in which we were given clearly defined examples of inputs and their corresponding outputs (feedback) and concluding with a general approximation of the data, Unsupervised Learning is not about concluding an approximation at all. Rather, it is the process of creating concise, compact descriptions of the data. For instance, in Image Processing, an example of Unsupervised Learning would be given a group photo with many people of varying characteristics, divide a group of pixels (people’s faces) into categories of male or female faces.
Simply put, Re-enforcement Learning is the process of learning from delayed reward. In this discipline, feedback may be received several steps after a decision has already been made. This is similar to Supervised Learning, however the feedback in Re-enforcement Learning is not present immediately. Let’s look at a simple example of Re-enforcement Learning.
Example: A Tic-Tac-Toe Game
Challenge: Determine a winning strategy
In this example, Player X doesn’t receive feedback that it was using a losing strategy until Player O has placed 3 O’s in a row and wins the game. It isn’t until then, that Player X can revisit the choices made to determine where things went wrong and even which moves were critical.
No, Machine Learning is not some sort of evil invention that will eventually lead to the overthrow of mankind. We are at least a few light-years away from that point. In fact, Machine Learning can improve us human’s environments quite a bit. From predicting regular intervals throughout the day to heat and cool your home, to predicting smarter drug regimens, Machine Learning can even be used to predict which stocks to choose to optimize returns on a portfolio. By taking a quick glance at the 3 core disciplines of Machine Learning, perhaps you have a newfound appreciation for Machine Learning engineers the next time you ask Siri what the weather’s like today or spin the dial on your Nest home thermostat.