A Primer on Machine Learning
Single and wondering how everyone from Grandma to Google knows? Read through to the end, and you might find the answer to one of these questions.
It is a quiet evening, and so you find yourself watching videos of cats playing the piano on YouTube when an ad draws your attention to a purchase you have recently been considering. Welcome to Machine Learning (ML), a branch of Artificial Intelligence (AI) dedicated to eradicating humanity. Well, maybe not quite—yet. For now, ML is confined to well-defined tasks. That is because ML uses algorithms (set of instructions) to tackle discrete problems and draw valid inferences (make educated guesses).
ML can broadly be categorised into three, namely supervised learning, unsupervised learning, and deep learning. Central to each is learning from experience to make informed decisions. Where they differ is in the precise approach adopted. This piece is intended to be an overview of those similarities and differences.
Supervised learning can be understood as directed learning. That is, when someone with the right answer, called a supervisor, points out mistakes during the enlightenment process. Supervised learning is often applied to asset price prediction (referred to as regression) and fraud detection (referred to as classification). What makes these types of problems well-suited to supervised learning is the availability of accurate historical data. Said otherwise, supervised learning requires labelled data (or answers) to differentiate right from wrong.
Knowledge, however, is not limited to the classroom. We also learn through real-world experience, in what is known as experiential learning. Unsupervised learning, therefore, excels in contexts where there is no single correct answer. Customer segmentation (or clustering) is one such example, where the result of grouping customers based on similar attributes is likely to be a function of the criteria used. In many ways, unsupervised learning resembles the way we humans make progress. It is, for instance, how infants learn to walk. Without explicit instruction, they practice, stumble, and slowly get better at the delicate act of keeping one foot in front of the other. Evaluating the performance of unsupervised learning models can thus be a vexing proposition. Imagine grading a toddler on the proficiency of their walk.
That brings us to the third type of ML, deep learning. Deep learning is learning from past mistakes by iteratively adjusting behaviour to minimise error. Think back to the first time you accidentally touched a hot stove. Chances are the tingling sensation you felt on your fingertips, dissuaded you from repeating it. Similarly, a deep learning model adjusts its forecast by minimising a cost function (a mathematical penalty for making poor decisions). For this reason, deep learning has proven amenable to time-series analysis, i.e. predicting the movement of stock prices and other traded assets.
Deep learning, or neural networks, can also be applied to classification type problems with a higher degree of accuracy. Say, for example, you buy your partner a £40 bouquet every Friday at 1 p.m. in London. After the first few instances, a deep learning model will begin to interpret this as normal behaviour for the card. If you (or somebody else ostensibly on your behalf) then tries to withdraw £3000 from an ATM in Tasmania, the model will automatically generate an alert. While the example may seem contrived, imagine scaling the same phenomenon across hundreds of millions of transactions across a global network in real-time, and it is easy to see why financial institutions have become some of the most enthusiastic adopters of neural networks.
So there you have it. Google infers your relationship status by using your browsing history to assign you to a group of users with similar browsing habits, where a binary classifier is typically employed to label you as either ‘single’ or ‘in a relationship’. Just try explaining it's complicated’ to a machine.