Can we beat state of the art from 2013 with 0.046% of training examples? Yes, we can

I just published an article along with a repository where I use what we learn in Lesson 1 to beat the state of the art result from 2013.

Maybe I made a mistake somewhere, but ran it a couple of times, seem to be good.

I was blown away by the results. And listen you guys, this is just what we learn in lesson 1 of part 1 :smiley: