Machine Learning

image showing an apple and several oranges in a line

When most people think of artificial intelligence, it conjures up sci-fi movie images of hyper-intelligent robots taking over the world. Scary, no?

Fortunately, those kinds of stories belong firmly in Hollywood. The reality is much less terrifying and much more useful for research.

Machine learning is a particular type of artificial intelligence. It involves designing computer algorithms which can get better at doing a particular task over time. For example, let’s say you want a computer to help you separate pictures of apples and bananas. You can make an algorithm designed to do it, give the algorithm a whole bunch of fruit pictures, and leave it’s circuits buzzing away to label them as apples or bananas. At first, the algorithm will almost definitely be terrible at this task. After all, the computer has no idea even what fruit is!

That’s where the ‘learning’ comes in. When the algorithm has finished, you can tell it which images it correctly recognised as apples or bananas, and which ones it got wrong. The algorithm will then tweak itself, try again, and see if it does any better. If it correctly recognises more fruits, then it keeps these new changes. It then tweaks itself again to try to improve further. It’s basically an enormous series of trial and error. The computer tries over and over again to identify apples and bananas, until it does it very well.

And congratulations – you’ve trained a computer algorithm to identify apples and oranges! You can now unleash it on all of the images of fruit that you can possibly gather.

So, why is this so useful in epilepsy research?

One example is detecting seizures from electrical recordings of brain activity (like EEG recordings, for example). Just like the apples and bananas in our example, seizures look different to normal brain activity. This means computers can learn to sort between the two. A big hope is that, as computer algorithms get better at this, they will be able to actually predict seizures. They could do this by being able to recognise what the EEG signal looks like just before a seizure.

A recent example used machine learning to predict whether particular changes to our genetic code would be likely to cause epilepsy. This has the potential to change how quickly we can detect, diagnose and also treat genetic epilepsies.

Until recently, machine learning has really only been used seriously by computer scientists. Excitingly, more and more epilepsy scientists and other biologists are seeing the power of this approach, and it has huge potential to drive future discoveries in science and medicine.