fastai: a library that makes deep learning accessible

Recently our company, n-waves, has won Poleval, national NLP competition. I’d like to share with you some resource without which it wouldn’t have been possible: the fastai library, fast.ai’s courses, and the community around them. If you’re struggling with finding an acceptable service to address your automation needs or if you are interested in applying recent breakthroughs in artificial intelligence, then these resources can also work for you.

Until recently, entrepreneurs that wanted to run quick experiments in AI were doomed to use black box online solutions that were hard to adapt to their needs. For instance, I recall how I wanted to run a quick proof of concept and analyze the sentiment of my slack channels to quantify the mood in the team. I couldn’t find anything that would work with the Polish language! Let alone two languages, and I knew that English sentiment analysis at that time was somehow working!

That experience led me to believe that you need to understand deep learning to be able to do anything useful with it.  But if you are on your own, without a team of researchers, how you can keep up? You may feel like a small fish that swims next to the freighter, you get scraps here and there, but if the scrap required $10k of computing power, then you are out of luck.

Fortunately, the situation has changed thanks to the open work that the fast.ai team is doing:

  • They have the most approachable course for Deep Learning out there if you are on to get your hand dirty and get results quickly. 

  • Their new library fastai v1.0 is specially tuned for quick iteration and applicability that helps you do your research quickly.

  • It is built for entrepreneurs that want to apply deep learning to their business but do not have an entire data center at their disposal. (A small team of student AI coders beats Google’s machine-learning code)

  • It differs from other frameworks in that it implements the state of the art algorithms by default - often, no configuration is needed. This gives you a way to keep up.

  • Out of the box you get best data augmentation techniques, learning schedules that allow for super covergence (18 mins to train ImageNet),  state of the art in text classification, some of the best image classification models, all of this just 6 lines of code away.

  • Also, you know what? It is going to have the best sentiment analyzer for Polish. :)

With the help of this library and the support and teaching of Jeremy Howard we managed to win the Polish Natural Language Competition. fast.ai is all about applicability, so we are already in talks with some of the smart companies in Poland and Germany to apply fast.ai to build things like:

  • Better brand monitoring tools that support not only English, but also other European languages,

  • Tools to extract text from smartphone quality pictures of region-specific documents  like retyping of recipes, invoices or flight tickets,

  • Deep learning supported stocktaking,

  • And finally, some models to automate law offices, with smart text generation and outcome prediction.

As you see there is a middle way between using out of the box inflexible services and funding your own research team. I hope you feel encouraged to give it a try with your own ideas if so let’s meet on their forum I’m happy to help, and if you need a quick proof of concept you know where to find us.


Photo by chuttersnap on Unsplash