What is machine learning?

What does it mean for a machine to learn? And how can a machine learn to solve a problem given training data? Machine learning is a completely different way of creating computer programs. In conventional computer programming, an expert software engineer writes some code that converts input into the desired outputs. This relies on the expertise and ingenuity of the programmer to come up with steps and rules that go into the program. But for many problems, it’s very difficult to write a program to solve them by hand. For example, consider the problem of recognising a face from a photograph. How do we go about deciding what features to extract from the image or how to compare features between faces?

Since we don’t know how this problem is solved, we can’t write a program to do it. Or maybe we do have some idea of how to solve a particular problem but the program to do it might be horrendously complicated. This is where machine learning comes in. Instead of writing a program by hand,

we collect training data: example inputs along with the corresponding desired output. A machine learning algorithm takes these examples and produces a program - we call this training. This program can then be given new inputs that were not in the training examples and, if we’ve done things right, it should still produce sensible output. Our training data can come in many forms. The input might be one dimensional, like the price of a stock over time. It could be two dimensional like an image with a grayscale value at each pixel. It could be three dimensional, like an image that also has colour information at each pixel. Or it could be a sequence of colour images over time, i.e.

a video, which is four dimensional. There are two types of output that we might want our machine learning system to predict. The first type is where we want the system to choose one from a set of discrete options. For example, we might want to say whether a photograph contains a cat, a dog or a person. This is called classification. The second type is where we want the system to predict a continuous value. This is called regression. For example, we might want to predict the age of a person from their photograph. Or we could even output an image or video in which case we want to predict the colour values for each pixel.

So, we can think of machine learning as the task of finding a function that maps inputs onto outputs in a way that approximates the training examples we were given.