Is this a bird, a plane ...?!


Fast.ai

I’ve been wanting to specialize in Data Science and Machine Learning for a year or two now. I completed a few MOOC (Introduction to Computational Thinking and Data Science, Udacity’s Intro to Machine Learning), and I have read Deep Learning with Python by F. Chollet (which is really great !), but everything didn’t “click” with me. Ask me to do a project and I wouldn’t even know where to start. I managed to do the Titanic Kaggle competition. Using a DecisionTree Classifier in Sklearn, I got a score of 0.746. Given that it’s a binary classification problem (“Did the people managed to escape the Titanic or not ?”), the baseline is 0.5, so my score isn’t awful but it could be better.

I have been hearing about Fast.ai for a long time, so I decided to try it this week. I’m not used to the “top-down” approach to teaching they use (hey, I’m an undergrad, of course I’m not!), but I’ll wait until I finish part one to judge.
Anyway the second lesson of the course leaves us with homeworks: we have to implement a Deep Learning algorithm on a web page. That seemed unrealistic, but it turned out to be quite a fun challenge.

Big S or Flannel Shirt ?

Using the code depicted in the second lesson notebook (Available here), I made a dataset of two classes : Superman/Clark Kent (Mostly the Cavill incarnation). Got impressive result super fast. It makes almost no mistakes on pictures of Superman or Clark Kent. On unrelated pictures, it acts as a superhero detector. If the people in the picture wear superhero costumes they’re labeled as Superman, otherwise most pictures are labeled as Mr. Kent.

I quickly made a view in Django to put it on my website and it is now testable. You can check it out here : The Lois Lane AI.

Overall It was a great project as I could immediately put in practice what I saw in all those courses. And even if it’s a small project, it’s always a pride booster to make something that works!