No that wasn’t a typo, we didn’t mean to refer to Florence + The Machine (OK, maybe a little bit) but actually we wanted to start off this topic with a bold statement: Computers are stupid. They don’t know what they want or what they need.
Anyway, with that out of the way we can say the big difference between people and computers is that people learn new things over time. Computers, well, they have no way of realizing that there is more to life than just crunching numbers and singing an algorithm here and there in their ‘heads’.
But hey! I thought this was a post about the magic they call machine learning – and you’re right my friends, I am getting there… ahem…
Basically machine learning is a way for the computer to learn stuff. A more familiar term could be artificial intelligence, they’re not actually the same, but sort of related.
Here at Stylight we use machine learning to make sure you get the fashion you want the way you want it. Every time you type in the search box ´thing´ in the right-hand corner, we suggest you terms that are similar to what you are typing.
Let’s say you are looking for T-shirts and so you start typing:
Each one of the suggestions is what we call a ‘category’, and to add to that, we have many category levels too. So, T-Shirts for example, is a top level category, while Casual, Printed, Long Sleeve and V-Neck are its subcategories.
We use machine learning in order to keep your fashion neatly classified into all these categories (which are over 800 different ones!!!). Long story short, our computers read the descriptions on each article, and based on that they decide whether a T-shirt is a T-shirt and not a Shoe, and then whether it is Long or Short Sleeved.
It is called machine learning because we make the machine “learn” from previous knowledge. We feed it fashion articles of which we are sure about their category, so it can later differentiate between T-shirts and Shoes on its own. I’m not going to cover our secret sauce in detail in here, but you can read more about it on our engineering blog:
An important point is that the computer can’t make up knowledge on its own. If it has only learned about Shirts and Shoes in the past, and you give it Pants, it will try to tell you that they are either Shirts or Shoes. For this computer Pants don’t even exist!
As a result, we can show you the right articles for your search most of the time. Some wrongly tagged ones still slip through the cracks. But after all, computers aren’t perfect.
We use the description on each article to label it, so we need to teach our machines about all the different languages we work with. The computer doesn’t know when English is English and not German, we just make sure to let it choose only fashion in the English language, and from there it looks for similar stuff in English only too. Right now we work with German, English, French, Spanish, Portuguese, Swedish, Norwegian and Dutch and we are proud of such an assortment!
Before we built our current machine learning system, which we affectionately call HAL, all that tagging and categorization was done by our great quality assurance team. A whole unit of people looking at fashion items and tagging them right. While people do excel at this task, they also get tired and need to rest. Thanks to HAL we have been able to tag around 70,000 products a day, freeing a great workload from our QA team and allowing us to show products from all of our stores much faster.
Here at Stylight we work hard everyday to improve our machine learning process. Our awesome engineers are always up to date with the latest technologies and implement them within reason. While we love to experiment and play, we take careful steps when choosing new paths. We’ve recently experimented with Deep Learning, the new hot stuff that everybody is talking about – we even hosted the Deep Learning meetup in Munich!
Someday we will be able to make suggestions to you on what to wear every morning based on the weather, your mood, how hungry you are or even your schedule for the day… As machine learning improves and engineers have more and more access to data, life will become easier as most of your daily menial activities will be automated, leaving you with more free time to enjoy being under the sun.
|By Rafael Perez |