by Tash Clarke
You’ve probably heard lots about artificial intelligence (AI), or machine learning. Whether in the latest blockbuster or in the news, AI is often painted as a villain, with robots taking over the world - or at least putting people out of jobs. But let’s take a look at what it really is. You will more than likely see examples of machine learning every day, and not even notice it - even in the fashion industry! So what is it and how does it work?
Humans are great at spotting patterns. As a child develops, they will start to notice similarities in things, and will instinctively learn to categorise things based on features they possess. For example, cats tend to be small, have long tails and meow, whereas dogs are bigger, can have shorter tails, and bark. The older we get, as we come across more and more cats and dogs, the better we become at knowing that a Chihuahua is a type of dog, even if it can fit in a handbag. Well, it turns out that machines are also pretty good at this. They can learn in a similar way, which can be really useful for things you may not have even considered, like online shopping.
Say you wanted a website that will scour the internet and find all different shoes for you, but you want them in groups, e.g. boots and trainers. You can imagine if a human had to go through ASOS, Topshop and New Look, finding all the boots and trainers, it would take absolutely ages. They’d then have to start again every couple of weeks when a new collection is released! Not possible. But, with machine learning we can teach an algorithm (a set of instructions used to perform a task) to do this for us.
We start by showing the algorithm some pictures we already have labelled as either boots or trainers (label in this case means the group it belongs in, not the one that tells you which wash cycle to use!). This is the training phase, where the algorithm sees lots of training data and learns what features are important for each group. Imagine seeing boots and trainers for the first time - what features are different about the two groups? Our algorithm also has these pattern recognition skills, so might learn that boots tend to be black, are taller, and have zips. On the other hand, trainers are colourful and have laces.
Next comes the testing phase. The algorithm is shown boots and trainers that it’s never seen before, and guesses which group they belong in. If it’s shown 100 pictures (say, 50 boots and 50 trainers) and guesses the correct label for 85 pictures, we can say that it’s 85% accurate. This isn’t really good enough – we don’t want any boots appearing when we’re only interested in trainers, and vice versa. So, to improve our algorithm, we go back to the training phase, showing it more examples of each. This time when we test it on 100 brand new pictures, it’s 99% accurate. Just like a child growing up, the better we train our algorithm, the better it is, and can then be used in the real world.
There are some great apps that use this technology, such as SSENSE, where you browse outfits and accessories and select ones that you like. The app then learns your style and will show you items you’re more likely to lust after. My SSENSE has learnt that I’m a leopard print kinda gal, who rarely wears skirts. The more you use it, the better it will be at knowing your taste, because it has more training data to learn from!
So what about that taking over the world thing? Well, an algorithm cannot learn about something that it hasn’t seen during its training. For example, if we showed our boot vs. trainer algorithm a picture of a belt, it would just blow its mind! It will have to guess if it’s a boot or a trainer because it hasn’t learnt the features of a belt. This means, the chance of machines becoming ‘conscious’ and being able to think by themselves is very slim (in the near future at least), because they can only learn from data that they’ve seen before. So, you’re probably safe from a robot apocalypse for now, but they will help you find an awesome new pair of kicks!