‘Deep learning [is] the rocket fuel of the current AI boom.’
Despite often being used interchangeably, ‘machine learning’ and ‘deep learning’ are, in fact, distinct concepts. The confusion is understandable – the world of tech is filled with buzzwords (think ‘agile’ and ‘disruptive’) that quickly shed any meaning they may once have had with over- and incorrect use. It’s easy to assume that ‘deep learning’ is just a more exciting, marketable way to talk about machine learning.
Why then, do respected industry commentators like Tom Simonite distinguish between the two, and why are they so excited about deep learning in particular?
In this case it’s even more confusing than usual because the two are so closely linked. There’s a quick answer, though:
Put simply, deep learning is a subset of machine learning that requires less human input.
Both fall under – and aim to achieve – ‘artificial intelligence’.
The basic differences
‘When we say something is capable of “machine learning”, it means it’s something that performs a function with the data given to it and gets progressively better over time’
In order to explain how basic machine learning works, we’ll make use of the ubiquitous image recognition example:
- Input data (pictures of apples and bananas) is gathered.
- A programmer ‘teaches’ the machine learning software the features it will use to perform its function (identifying each by their shape).
- The software works through the data using the features it has been provided (classifying an image ‘apple’ or ‘banana’).
- The software outputs its answers.
- If the software gets it wrong, the programmer has to correct the software in order for it to improve.
The elements that fundamentally set machine learning and deep learning apart are in bold. Let’s now take a look at the workflow for deep learning software:
- Input data is gathered on a much larger scale.
- The software determines the differences between apples and oranges itself, then categorises them.
- The software outputs its answers.
- If the software gets it wrong, it is able to correct itself and improve over time.
The biggest difference? There’s a step missing. That’s because deep learning algorithms don’t need to be told what features to look for – they’re able to figure it out themselves. How does that work, though? Here’s a (very quick) rundown.
Artificial neural networks
Deep learning owes its capabilities to artificial neural networks. Where machine learning relies on the input of a programmer to ‘teach’ it the language of its task – i.e. the features that set an apple apart from a banana – deep learning is able to look at millions of pictures of apples and bananas and figure it out on its own by creating its own ‘language’.
Artificial neural networks are inspired by the structure of the human brain but work slightly different in practice. Instead of a cluster of billions of neurons firing at will, the artificial alternatives are organised in layers:
- Each layer looks for something different in the input data, and weights its importance based on past experience.
- The weighted data is then passed on to the next layer for the same process. This step is repeated for as many layers as the programme has, and data can be passed back and forth again if necessary.
- An output is delivered.
- The process is repeated until the programme understands what it is looking at, and has developed its own language (in this case, visual) to understand the input data.
This is why deep learning requires so much more data than machine learning; it has to learn from its mistakes on its own. Without millions of photographs, for example, it’s not much use. The increased scale is one of the reasons why deep learning also needs a lot more processing power behind it.
The future of machine learning and deep learning
As it’s hopefully become clear, machine learning and deep learning are distinct concepts within AI. While deep learning is more technically complex, machine learning is still and will continue to be very widely applied. Whenever there’s a smaller amount of data and less processing power, machine learning will win out, and vice versa. They’re suited to different tasks.
Deep learning’s surge to ‘rocket fuel’ status has been fairly recent – breakthroughs only really took place in 2012 with the AlexNet neural network. Before 2012, neural networks were seen as an impracticality because of the immense processing power they required.
Since then, it’s been put to incredible use, (for some jaw-dropping examples, take a look at Generative Adversarial Networks) but it’s been held back from more sensitive applications in industries like healthcare and law, because the neural networks have had trouble explaining themselves. In other words, we’ve struggled to keep up with our own creations.
The even more recent advent of ‘Explainable AI’ (XAI) is poised to remove this new barrier, and see deep learning (and AI as a whole) applied more widely than ever before.