For most of my undergrad I treated machine learning as something other people did. People with stronger math backgrounds, people who had taken the right courses, people doing research. Not something a software engineer just picked up.
This year changed that for me.
What shifted
The tooling got accessible. scikit-learn makes it possible to train a real classifier in an afternoon without implementing anything from scratch. The Andrew Ng Coursera course exists and it's genuinely good. IPython notebooks make it easy to explore data interactively without setting up a whole project structure.
None of this is new in 2015 but it all reached a kind of critical mass where the barrier between "software engineer" and "person who can do machine learning" became much smaller than I had assumed.
I spent the first half of this year working through my thesis, building classifiers, breaking them, understanding why they broke. The second half I started reading more broadly. Word2Vec, convolutional nets, what Theano actually is and why people use it. Google just open-sourced TensorFlow last week and I haven't had time to look at it properly yet but the timing feels significant.
What I got wrong along the way
I assumed that if the accuracy number was good, the model was good. It took me a while to learn that a number going up is not the same as something working. Your metric can be going up while the model is learning the wrong things entirely.
I also spent too long reading about techniques before trying them. Reading about gradient descent is much less useful than implementing it badly, watching it fail, and figuring out why.
Where things seem to be going
Deep learning feels like it's moving from research curiosity to something engineers actually use. The results on image recognition in the last few years have been striking. I don't know how much of that will translate into practical products in the near term, but it's hard to watch what's happening with convolutional nets and not think something is changing.
I'm finishing my MSc soon and thinking about what comes next. Whatever it is, I want machine learning to be part of it. Not as a specialty I defer to other people, but as a tool I can actually use.
2015 was the year I decided to take that seriously.
With gusto, Fatih.