The spotlight aimed at artificial intelligence has been impossible to miss. It’s been a consistent feature of the recent media landscape, with every new innovation in the field being hailed as the revolution in how humans can interact with technology. Companies are also quick to market themselves at the forefront of AI in an attempt to appear cutting edge. It is in this atmosphere of hype that people have started throwing around terms such as “machine learning”, “deep learning” and “neural networks” as interchangeable buzzwords to describe anything that has the slightest fragment of AI technologies.
The truth is that these terms refer to different subsets of the surprisingly large field that is artificial intelligence. So, what exactly are these different concepts that often dominate any conversation regarding technology? First of all, we must understand that AI is a wide area of study. It encompasses algorithms from those that drive self-driving cars to those made of simple conditional statements (known as Good-Old-Fashioned AI). Any definition of AI immediately highlights this problem, as shown as the by Encyclopedia Britannica: “The ability of a computer to perform tasks commonly associated with intelligent beings.” It is in this vague field that we find Machine Learning and Deep Learning.
If we take a dive down, Machine Learning is the first common subset of AI that we encounter. Although there has been a lot of focus on it, it does not represent AI as a whole. It is responsible for many of the technologies that are commonly referred to as “narrow AI” such as recognising faces on Facebook or image classification on Google. They are, at their core, algorithms that use known values to make estimates on new data.
Machine Learning is normally done with a neural network – a series of interconnected equations, each with different internal parameters. The algorithm uses a given training dataset, which consists of values for which the answers are known, to learn patterns in the data. It constantly re-adjusts its internal parameters to match these patterns, and uses these for new scenarios. Such “weighting” of its parameters ensures that the algorithm will move towards outputs with the highest potential rate of success.
In short, when we say an algorithm is capable of machine learning, we are saying that it is “performing a function on the data provided, and it is getting progressively better at it”. However, such algorithms still need readjustment from engineers when they return an inaccurate prediction. This can happen due to a variety of reasons, for example when its training data is biased in some way. Another layer down and we arrive at deep learning, which is the cutting edge of machine learning. It is considered a subset of it and improves upon its capabilities by utilising “deep” neural networks. The specific details are too long to go into here, but it allows using larger data sets than regular neural networks. This in turn leads to better results from the algorithm as a whole, at the expense of a longer training time and computational cost.
While all three terms are very closely linked, they shall not be confused or used interchangeably. Although they may be currently the media's focus, they only constitute the tip of the AI iceberg.